Upload 8 files

EPFL_Enterprise_OSAI_Adoption_Interview_Data.csv

@@ -1,157 +1,7 @@
-ID,Interview Type,Date,Confidentiality,Role,Industry,Stage,Context Summary,Theme 1: Decision Factors for AI Solution Selection,Theme 2: Integration Challenges and Organisational Readiness,"Theme 3: 
-Motivations for Adopting Open-Source AI",Theme 4: Use Cases and Demonstrated Value,Theme 5: Adoption Gates for Governance and Compliance,Theme 6: Execution Levers for Adoption Success,Link to Framework,Main Takeaway
-I1,Semi-structured (online video),20250605,Anonymous,AI Lead,Technology/ Software,Adopted,"The AI Lead shared insights on enterprise evaluation of open-source AI models, focusing on cost-performance trade-offs, integration barriers, and governance readiness. Customers increasingly mention DeepSeek for its popularity, but most underestimate infrastructure demands.","Bigger models mean higher inference cost, latency, and infra demands. The participant stressed cost�capability trade-offs and said clients compare models like Qwen, Mistral, and DeepSeek based on fit and affordability.  �Bigger models equal higher inference cost.� Customers often ask for DeepSeek due to its popularity, but few grasp its compute requirements.",Most clients lack internal evaluation capacity. Fine-tuning remains complex and rare: �70% only go for fine-tuning as the last solution.� Teams typically rely on APIs or pre-integrated tools instead of custom retraining.  Integration remains a major friction due to limited expertise and documentation.,"Open-source adoption is motivated by flexibility, cost, and customization. Proprietary APIs are easier to use but limit control. Open models are chosen once organizations trust the governance layer and want adaptability.","Value arises from multi-agent use and task-specific models, enabling latency reduction and cost savings. Token costs dropped by ~80%, encouraging broader experimentation. Demonstrates tangible cost and performance benefits when matched to context.","Governance prerequisites include licensing clarity and documentation completeness. �Transparency and licensing: yes, they matter.� Deployment guidance is insufficient: �How can I deploy straight from the model card?� Without structured documentation and licensing, enterprises hesitate to scale.","After governance gates are cleared, key enablers include cost efficiency, performance, and latency optimization.  �Use case first, then cost, then latency.� Confirms the Gate�Lever model: governance first, then optimization through cost and performance levers.","Reinforces RQ2 and RQ3: � Gates: Licensing, documentation, and governance must be satisfied before scale. � Levers: Cost, latency, and performance drive adoption once trust is established. � Integration capability is a critical maturity barrier.","�Enterprises are drawn to open-source AI for flexibility and cost, but trust gates (licensing, documentation, and evaluation capability) determine whether adoption moves beyond pilot stage.�"
-I2,Semi-structured (online video),20250605,Anonymous,AI Lead,Energy,Adopting,"The AI Lead discussed challenges in adopting open-source AI within a highly regulated, risk-averse energy enterprise. The firm primarily uses proprietary AI for governance and reliability reasons.",Cost is secondary to reliability and compliance. The participant explained that the energy sector�s tolerance for uncertainty is low. Proprietary solutions are preferred due to clear SLAs and security posture.,"Integration complexity is high because of legacy systems, strict internal IT policies, and dependence on external vendors for compliance checks. The firm avoids open-source tools lacking enterprise-grade documentation.","Motivation to explore open models exists, but the risk profile is seen as incompatible. The participant said, �We monitor the open space, but for now we stay on managed APIs.�","The organization focuses on analytics and forecasting tools built into proprietary platforms. No open-source use cases yet, only sandbox experiments under supervision.","Compliance and security audits are mandatory. The firm requires formal certification, licensing clarity, and full documentation before any production use. These factors are non-negotiable gates.","If gate conditions were met, the main lever would be cost savings and customization for forecasting models. For now, vendor dependence is accepted as a trade-off for safety.","Reinforces RQ2 and RQ3. Compliance, documentation, and licensing act as strict Gate conditions. Levers such as cost efficiency remain theoretical until trust is achieved.","Highly regulated industries remain risk-averse; adoption depends on trust, certification, and documentation rather than cost or performance."
-I3,Semi-structured (online video),20250611,Anonymous,Head of AI Strategy,Finance,Pre-adoption,"The participant described a cautious approach to AI adoption. The firm uses primarily proprietary solutions within a tightly regulated environment, with minimal exposure to open-source AI.","Decision factors center on compliance, security, and auditability. The participant said cost is �not the first concern; risk is.� Proprietary systems are preferred due to full traceability and vendor accountability.","Integration challenges stem from strict data governance policies and internal compliance frameworks. Open-source tools lack legal vetting and risk scoring, which halts experimentation.","Motivation to adopt open-source AI is very low. The organization prioritizes stability and regulatory trust. The participant noted, �We cannot risk using code without full control over provenance.�","No open-source use cases yet. Internal pilots focus on data anonymization and model validation tools, all within managed cloud environments.","Governance is gate-dominant. The firm requires verified licensing, source traceability, and complete documentation before considering open models. Compliance officers must approve every vendor or repository.","Execution levers are not active at this stage. If adoption occurs, it will depend on proven compliance, auditable data flows, and certified model risk frameworks.","Reinforces RQ2 by showing that pre-adoption sectors remain blocked by Gate conditions: compliance, documentation, and provenance. RQ3 logic holds as no Levers are activated.","In highly regulated finance, adoption is ruled by governance gates. Cost and innovation only matter once compliance assurance is guaranteed."
-I4,Semi-structured (online video),20250618,Anonymous ,Data Science Manager,Telecom,Adopting,The participant described a telecom enterprise that uses both proprietary and open-source components. The firm experiments with open models but faces integration and compliance hurdles.,"Decision factors include performance consistency, cost efficiency, and interoperability with legacy systems. Open models are evaluated but often fail due to unclear SLAs.","Integration challenges are significant. Telecom workloads demand low latency and high reliability. The participant said, �We need models we can monitor continuously.� Internal ops teams find deployment documentation lacking.","Motivation for adopting open-source AI is driven by cost reduction and vendor independence. However, production teams still depend on managed APIs for scaling.","Successful experiments exist in customer churn prediction and text summarization. However, scaling these pilots remains slow due to monitoring complexity.",Governance is a constant concern. The firm requires clarity in licensing and provenance before integrating open models into live customer systems.,"Execution levers include cost optimization, customization, and faster iteration cycles once internal validation is done. Performance and latency drive adoption success.",Reinforces RQ2 and RQ3. Integration barriers (Q10) and documentation gaps (Gate) slow adoption. Levers emerge only when trust and monitoring readiness are proven.,"Telecom adoption is paced by integration maturity. Cost and performance attract interest, but documentation and monitoring readiness determine scale."
-I5,Semi-structured (online video),20250623,Anonymous ,AI & Analytics Lead,FMCG,Adopting,"The participant leads analytics initiatives in a global FMCG firm. The company integrates AI mainly through Azure services, focusing on automation, forecasting, and marketing optimization. It uses proprietary tools due to internal skill constraints and reliance on enterprise-grade SLAs.","Decision factors center on reliability, integration ease, and vendor support. The participant said, �We use Azure because it works with everything we already have.� Cost is monitored but secondary to scalability and security.",Integration challenges arise from legacy systems and distributed data silos. The company lacks deep ML engineering capacity and depends on managed APIs. Internal teams face difficulty maintaining custom pipelines beyond Azure workflows.,"Motivation to adopt open-source AI is limited. The firm monitors OSS models but avoids them due to maintenance risk and lack of expertise. The participant said, �We prefer solutions that come supported � we do not have time to retrain or host models ourselves.�","Use cases include demand forecasting, campaign optimization, and customer insights using Azure ML and Cognitive Services. These applications deliver measurable ROI and are integrated with BI dashboards.","Governance is focused on data security, privacy, and brand risk. All AI deployments must pass internal IT compliance review. The participant emphasized data residency and model auditability as key gates before production.","Execution levers include vendor support, fast deployment, and cost predictability. Managed environments reduce operational risk and allow rapid scaling. �We measure success by how fast a model moves to production,� the participant said.","Supports RQ2 (decision factors and readiness) and RQ3 (lever activation without open-source gates). Demonstrates adoption driven by ROI and integration simplicity, not openness.","FMCG enterprises adopt AI through proprietary cloud ecosystems such as Azure, prioritizing speed, support, and ROI over transparency or control."
-I6,Semi-structured (online video),20250708,Anonymous ,Research Director,Education / Research,Adopted,"The Research Director oversees AI adoption in a higher education institution. The organization integrates open-source models through academic partnerships and Hugging Face resources. Its focus is on transparency, reproducibility, and academic collaboration.","Decision factors emphasize transparency, accessibility, and peer validation. Cost is important but secondary to control and reproducibility. The participant said, �If we cannot inspect it, we cannot trust it.�","Integration is supported by strong internal technical expertise but limited enterprise infrastructure. Challenges arise around maintenance, version control, and alignment with institutional IT policies.","The main motivation for adopting open-source AI is academic integrity and collaborative innovation. The participant stated, �We teach and research with open models because they reflect how science should work.�","Use cases include NLP research, student AI projects, and open-access teaching tools using Hugging Face and local deployments. These initiatives increase learning efficiency and research output.","Governance is centered on ethical AI use, licensing awareness, and data transparency. The institution follows strict research ethics protocols and uses model cards to document provenance and bias.","Execution levers include community support, documentation availability, and model reproducibility. Performance is valued when it serves academic goals, not just commercial outcomes.","Supports RQ2 (decision and motivation factors) and RQ3 (Gate - Lever interaction). Shows that governance gates are met through transparency, enabling levers such as innovation and collaboration.",Education and research adopters favor open-source AI for transparency and reproducibility. Governance alignment and internal expertise allow them to scale experimentation responsibly.
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-I1,,,,AI Lead,Tech,Adopted,"The AI Lead shared insights on enterprise evaluation of open-source AI models, focusing on cost-performance trade-offs, integration barriers, and governance readiness. Customers increasingly mention DeepSeek for its popularity, but most underestimate infrastructure demands.",,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
-,,,,,,,,,,,,,,,
+ID,Interview Type,Date,Confidentiality,Role,Industry,Stage,Context Summary,Theme 1: Decision Factors for AI Solution Selection,Theme 2: Integration Challenges and Organisational Readiness,"Theme 3: Motivations for Adopting Open-Source AI",Theme 4: Use Cases and Demonstrated Value,Theme 5: Adoption Gates for Governance and Compliance,Theme 6: Execution Levers for Adoption Success,Link to Framework,Main Takeaway
+I1,Semi-structured (online video),20250605,Anonymous,AI Lead,Technology/ Software,Adopted,"The AI Lead shared insights on enterprise evaluation of open-source AI models, focusing on cost-performance trade-offs, integration barriers, and governance readiness. Customers increasingly mention DeepSeek for its popularity, but most underestimate infrastructure demands.","Bigger models mean higher inference cost, latency, and infra demands. The participant stressed cost-capability trade-offs and said clients compare models like Qwen, Mistral, and DeepSeek based on fit and affordability. 'Bigger models equal higher inference cost.' Customers often ask for DeepSeek due to its popularity, but few grasp its compute requirements.",Most clients lack internal evaluation capacity. Fine-tuning remains complex and rare: '70% only go for fine-tuning as the last solution.' Teams typically rely on APIs or pre-integrated tools instead of custom retraining. Integration remains a major friction due to limited expertise and documentation.,"Open-source adoption is motivated by flexibility, cost, and customization. Proprietary APIs are easier to use but limit control. Open models are chosen once organizations trust the governance layer and want adaptability.","Value arises from multi-agent use and task-specific models, enabling latency reduction and cost savings. Token costs dropped by ~80%, encouraging broader experimentation. Demonstrates tangible cost and performance benefits when matched to context.","Governance prerequisites include licensing clarity and documentation completeness. 'Transparency and licensing: yes, they matter.' Deployment guidance is insufficient: 'How can I deploy straight from the model card?' Without structured documentation and licensing, enterprises hesitate to scale.","After governance gates are cleared, key enablers include cost efficiency, performance, and latency optimization. 'Use case first, then cost, then latency.' Confirms the Gate-Lever model: governance first, then optimization through cost and performance levers.","Reinforces RQ2 and RQ3: Gates: Licensing, documentation, and governance must be satisfied before scale. Levers: Cost, latency, and performance drive adoption once trust is established. Integration capability is a critical maturity barrier.","'Enterprises are drawn to open-source AI for flexibility and cost, but trust gates (licensing, documentation, and evaluation capability) determine whether adoption moves beyond pilot stage.'"
+I2,Semi-structured (online video),20250605,Anonymous,AI Lead,Energy,Adopting,"The AI Lead discussed challenges in adopting open-source AI within a highly regulated, risk-averse energy enterprise. The firm primarily uses proprietary AI for governance and reliability reasons.",Cost is secondary to reliability and compliance. The participant explained that the energy sector's tolerance for uncertainty is low. Proprietary solutions are preferred due to clear SLAs and security posture.,"Integration complexity is high because of legacy systems, strict internal IT policies, and dependence on external vendors for compliance checks. The firm avoids open-source tools lacking enterprise-grade documentation.","Motivation to explore open models exists, but the risk profile is seen as incompatible. The participant said, 'We monitor the open space, but for now we stay on managed APIs.'","The organization focuses on analytics and forecasting tools built into proprietary platforms. No open-source use cases yet, only sandbox experiments under supervision.","Compliance and security audits are mandatory. The firm requires formal certification, licensing clarity, and full documentation before any production use. These factors are non-negotiable gates.","If gate conditions were met, the main lever would be cost savings and customization for forecasting models. For now, vendor dependence is accepted as a trade-off for safety.","Reinforces RQ2 and RQ3. Compliance, documentation, and licensing act as strict Gate conditions. Levers such as cost efficiency remain theoretical until trust is achieved.","Highly regulated industries remain risk-averse; adoption depends on trust, certification, and documentation rather than cost or performance."
+I3,Semi-structured (online video),20250611,Anonymous,Head of AI Strategy,Finance,Pre-adoption,"The participant described a cautious approach to AI adoption. The firm uses primarily proprietary solutions within a tightly regulated environment, with minimal exposure to open-source AI.","Decision factors center on compliance, security, and auditability. The participant said cost is 'not the first concern; risk is.' Proprietary systems are preferred due to full traceability and vendor accountability.","Integration challenges stem from strict data governance policies and internal compliance frameworks. Open-source tools lack legal vetting and risk scoring, which halts experimentation.","Motivation to adopt open-source AI is very low. The organization prioritizes stability and regulatory trust. The participant noted, 'We cannot risk using code without full control over provenance.'","No open-source use cases yet. Internal pilots focus on data anonymization and model validation tools, all within managed cloud environments.","Governance is gate-dominant. The firm requires verified licensing, source traceability, and complete documentation before considering open models. Compliance officers must approve every vendor or repository.","Execution levers are not active at this stage. If adoption occurs, it will depend on proven compliance, auditable data flows, and certified model risk frameworks.","Reinforces RQ2 by showing that pre-adoption sectors remain blocked by Gate conditions: compliance, documentation, and provenance. RQ3 logic holds as no Levers are activated.","In highly regulated finance, adoption is ruled by governance gates. Cost and innovation only matter once compliance assurance is guaranteed."
+I4,Semi-structured (online video),20250618,Anonymous,Data Science Manager,Telecom,Adopting,The participant described a telecom enterprise that uses both proprietary and open-source components. The firm experiments with open models but faces integration and compliance hurdles.,"Decision factors include performance consistency, cost efficiency, and interoperability with legacy systems. Open models are evaluated but often fail due to unclear SLAs.","Integration challenges are significant. Telecom workloads demand low latency and high reliability. The participant said, 'We need models we can monitor continuously.' Internal ops teams find deployment documentation lacking.","Motivation for adopting open-source AI is driven by cost reduction and vendor independence. However, production teams still depend on managed APIs for scaling.","Successful experiments exist in customer churn prediction and text summarization. However, scaling these pilots remains slow due to monitoring complexity.",Governance is a constant concern. The firm requires clarity in licensing and provenance before integrating open models into live customer systems.,"Execution levers include cost optimization, customization, and faster iteration cycles once internal validation is done. Performance and latency drive adoption success.",Reinforces RQ2 and RQ3. Integration barriers (Q10) and documentation gaps (Gate) slow adoption. Levers emerge only when trust and monitoring readiness are proven.,"Telecom adoption is paced by integration maturity. Cost and performance attract interest, but documentation and monitoring readiness determine scale."
+I5,Semi-structured (online video),20250623,Anonymous,AI & Analytics Lead,FMCG,Adopting,"The participant leads analytics initiatives in a global FMCG firm. The company integrates AI mainly through Azure services, focusing on automation, forecasting, and marketing optimization. It uses proprietary tools due to internal skill constraints and reliance on enterprise-grade SLAs.","Decision factors center on reliability, integration ease, and vendor support. The participant said, 'We use Azure because it works with everything we already have.' Cost is monitored but secondary to scalability and security.",Integration challenges arise from legacy systems and distributed data silos. The company lacks deep ML engineering capacity and depends on managed APIs. Internal teams face difficulty maintaining custom pipelines beyond Azure workflows.,"Motivation to adopt open-source AI is limited. The firm monitors OSS models but avoids them due to maintenance risk and lack of expertise. The participant said, 'We prefer solutions that come supported - we do not have time to retrain or host models ourselves.'","Use cases include demand forecasting, campaign optimization, and customer insights using Azure ML and Cognitive Services. These applications deliver measurable ROI and are integrated with BI dashboards.","Governance is focused on data security, privacy, and brand risk. All AI deployments must pass internal IT compliance review. The participant emphasized data residency and model auditability as key gates before production.","Execution levers include vendor support, fast deployment, and cost predictability. Managed environments reduce operational risk and allow rapid scaling. 'We measure success by how fast a model moves to production,' the participant said.","Supports RQ2 (decision factors and readiness) and RQ3 (lever activation without open-source gates). Demonstrates adoption driven by ROI and integration simplicity, not openness.","FMCG enterprises adopt AI through proprietary cloud ecosystems such as Azure, prioritizing speed, support, and ROI over transparency or control."
+I6,Semi-structured (online video),20250708,Anonymous,Research Director,Education / Research,Adopted,"The Research Director oversees AI adoption in a higher education institution. The organization integrates open-source models through academic partnerships and Hugging Face resources. Its focus is on transparency, reproducibility, and academic collaboration.","Decision factors emphasize transparency, accessibility, and peer validation. Cost is important but secondary to control and reproducibility. The participant said, 'If we cannot inspect it, we cannot trust it.'","Integration is supported by strong internal technical expertise but limited enterprise infrastructure. Challenges arise around maintenance, version control, and alignment with institutional IT policies.","The main motivation for adopting open-source AI is academic integrity and collaborative innovation. The participant stated, 'We teach and research with open models because they reflect how science should work.'","Use cases include NLP research, student AI projects, and open-access teaching tools using Hugging Face and local deployments. These initiatives increase learning efficiency and research output.","Governance is centered on ethical AI use, licensing awareness, and data transparency. The institution follows strict research ethics protocols and uses model cards to document provenance and bias.","Execution levers include community support, documentation availability, and model reproducibility. Performance is valued when it serves academic goals, not just commercial outcomes.","Supports RQ2 (decision and motivation factors) and RQ3 (Gate - Lever interaction). Shows that governance gates are met through transparency, enabling levers such as innovation and collaboration.",Education and research adopters favor open-source AI for transparency and reproducibility. Governance alignment and internal expertise allow them to scale experimentation responsibly.

EPFL_Enterprise_OSAI_Adoption_Survey_Data.csv

@@ -1,164 +1,162 @@
-ID,1. What is the approximate size of your organisation?,2. What industry does your organisation operate in?,3. How familiar is your organisation with Hugging Face's platform and services?,4. Which type of AI solutions does your organisation primarily use?,5. What are the most important decision factors when choosing AI solutions for your organisation?,6. How challenging has it been to integrate open-source AI solutions into your workflows?,7. Is your organisation planning to increase its use of open-source AI in the next 12–18 months?,8. What motivated your organisation to choose open-source or proprietary AI solutions?,9. Can you describe a use case where Hugging Face or another open-source AI solution delivered value to your business?,"10. What major challenges (technical, organisational, or legal) has your team faced when adopting open-source AI?","11. What additional support, features, or services would make Hugging Face more attractive for enterprise use?",12. How does your organisation evaluate the trade-offs between open-source and proprietary options when making decisions about adopting AI solutions?
+ID,1. What is the approximate size of your organisation?,2. What industry does your organisation operate in?,3. How familiar is your organisation with Hugging Face's platform and services?,4. Which type of AI solutions does your organisation primarily use?,5. What are the most important decision factors when choosing AI solutions for your organisation?,6. How challenging has it been to integrate open-source AI solutions into your workflows?,7. Is your organisation planning to increase its use of open-source AI in the next 12-18 months?,8. What motivated your organisation to choose open-source or proprietary AI solutions?,9. Can you describe a use case where Hugging Face or another open-source AI solution delivered value to your business?,"10. What major challenges (technical, organisational, or legal) has your team faced when adopting open-source AI?","11. What additional support, features, or services would make Hugging Face more attractive for enterprise use?",12. How does your organisation evaluate the trade-offs between open-source and proprietary options when making decisions about adopting AI solutions?
 1,Fewer than 100 employees,Education / Research,1,Neither,Cost,1,Not sure,"No AI in use, so no motivation yet.",None yet.,"Lack of in-house AI expertise, limited budget.",N/A,"Would weigh cost, ease of integration, and student data privacy against performance and vendor reliability."
-2,"1,000–4,999 employees",Education / Research,4,Both,Customization,3,Yes,"Open-source is chosen for flexibility, transparency, and fostering innovation in the academic environment.
-Proprietary solutions are used where specialized capabilities or vendor support are necessary.","Hugging Face models have been utilized in projects involving natural language processing courses, student research on AI ethics, and prototype development of AI-assisted tools in education.","Technical difficulties in integrating diverse open-source tools into educational platforms.
-Ensuring alignment with institutional data protection and compliance policies.",Specialized documentation and tutorials for academic use cases.,"Evaluation weighs the educational value, flexibility, and community support of open-source against the reliability and ease of use offered by proprietary solutions. "
+2,"1,000-4,999 employees",Education / Research,4,Both,Customization,3,Yes,"Open-source is chosen for flexibility, transparency, and fostering innovation in the academic environment. Proprietary solutions are used where specialized capabilities or vendor support are necessary.","Hugging Face models have been utilized in projects involving natural language processing courses, student research on AI ethics, and prototype development of AI-assisted tools in education.","Technical difficulties in integrating diverse open-source tools into educational platforms. Ensuring alignment with institutional data protection and compliance policies.",Specialized documentation and tutorials for academic use cases.,"Evaluation weighs the educational value, flexibility, and community support of open-source against the reliability and ease of use offered by proprietary solutions."
 3,Fewer than 100 employees,Education / Research,3,Proprietary ,Performance,4,Not sure,"Open-source offers flexibility, cost savings, and freedom from vendor lock-in, making it attractive for experimentation. Proprietary tools are often faster to deploy, have stronger support, and require less maintenance, which matters for our small team. The decision will depend on how well open-source can meet our needs without creating operational bottlenecks.","Experimental use of Hugging Face NLP models to summarize cohort session transcripts. While promising, we need to assess whether we can support this at scale.","Ongoing maintenance and updates to models, content accuracy and avoiding bias in learning materials.",Step-by-step guides for SaaS integration and deployment.,"We weigh the flexibility, cost-effectiveness, and innovation potential of open-source AI against the stability, vendor support, and ease-of-deployment of proprietary solutions."
-4,Fewer than 100 employees,Education / Research,3,Proprietary ,Ability to improve learner engagement and outcomes.  Ease of integration with minimal engineering overhead.  Affordability and scalability for a startup.  Compliance with educational data privacy standards.  Reliability and vendor support.,3,Not sure,"Open-source AI is valued for experimentation, and avoiding vendor lock-in.
-Proprietary AI is preferred for speed, reliability, and ease of deployment, especially given the small team size.",Piloting Hugging Face NLP models for automated summarization of cohort sessions and personalized course recommendations.,"Limited internal AI engineering resources.
-Ongoing maintenance and updates for models.
-Ensuring AI-generated content is accurate, unbiased, and appropriate.
+4,Fewer than 100 employees,Education / Research,3,Proprietary ,Ability to improve learner engagement and outcomes.  Ease of integration with minimal engineering overhead.  Affordability and scalability for a startup.  Compliance with educational data privacy standards.  Reliability and vendor support.,3,Not sure,"Open-source AI is valued for experimentation, and avoiding vendor lock-in.
+Proprietary AI is preferred for speed, reliability, and ease of deployment, especially given the small team size.",Piloting Hugging Face NLP models for automated summarization of cohort sessions and personalized course recommendations.,"Limited internal AI engineering resources.
+Ongoing maintenance and updates for models.
+Ensuring AI-generated content is accurate, unbiased, and appropriate.
 Integrating AI outputs without disrupting user experience.",More education-specific AI templates ,"Balances innovation, cost-effectiveness, and flexibility of open-source AI against the stability, support, and ease of use offered by proprietary solutions."
-5,100–499 employees,Education / Research,5,"Open-source (e.g., Hugging Face)",Cost,3,Yes,"We prioritise transparency, reproducibility, and cost control. Open source AI let us fine tune for our domains and keep research data on institutional infrastructure.",Literature triage and summarisation for researchers using HF models with a small institutional knowledge base. It reduced screening time and helped surface relevant papers.,"Budget for compute, standardising data governance across labs, and keeping evaluation practices consistent.",Improved model cards.,Cost for research workflows.
-6,"1,000–4,999 employees",Education / Research,4,"Open-source (e.g., Hugging Face)","Cost, Transparency, Collaboration",3,Yes,Accessibility and community,Course material summarisation,Infrastructure limits,Institutional SLAs,Openness > Proprietary cost
-7,500–999 employees,Education / Research,4,Both,"Transparency, Cost, Ecosystem",3,Yes,Reproducibility,Research assistant chatbot,Compute resources,Academic deployment credits,"Open first, cost later"
-8,100–499 employees,Education / Research,3,"Open-source (e.g., Hugging Face)","Cost, Learning Value",4,Yes,Freedom to experiment,Student-facing tutoring bot,Lack of skilled ops staff,Step-by-step deployment tutorials,Feasibility → Learning gain
-9,"1,000–4,999 employees",Energy / Utilities / Oil & Gas,1,Proprietary ,Data control / privacy,1,Not sure,"Managed services with SLAs and EU-region hosting simplified GDPR and NIS2 compliance and accelerated go-live, while we keep sensitive OT and customer data in a private cloud.","A pilot RAG system using HF embeddings and a local vector index over safety manuals and maintenance SOPs cut lookup time for field engineers, with all content kept inside our network.","The main issues are organisational and legal, technical friction was low in pilots.",/,"Data residency first. If a workload can stay in our VPC with clear auditability we consider open components; otherwise we prefer proprietary services with SLAs, then compare performance and TCO"
-10,"1,000–4,999 employees",Energy / Utilities / Oil & Gas,2,Both,Data control / privacy,2,Not sure,Better performance and transparency ,"Distributed optimization, optimization solvers ","Concerns about support, legal in terms of licenses or data privacy. ",It’s already good and easy to use models. Nothing to add. ,"Data privacy, quality/performance and concerns about longer term support play more important role than costs. "
-11,"1,000–4,999 employees",Energy / Utilities / Oil & Gas,2,Proprietary ,Data control / privacy,4,Yes,"Open-source: flexibility, control over models, ability to deploy on-prem for sensitive data, cost efficiency at scale.
-Proprietary: faster deployment for certain NLP and summarisation tasks, vendor support, cutting-edge capabilities.",N/A,"Technical: aligning open-source models with our internal cybersecurity and compliance frameworks.
-Organisational: upskilling engineers to work with model training pipelines.
+5,100-499 employees,Education / Research,5,"Open-source (e.g., Hugging Face)",Cost,3,Yes,"We prioritise transparency, reproducibility, and cost control. Open source AI let us fine tune for our domains and keep research data on institutional infrastructure.",Literature triage and summarisation for researchers using HF models with a small institutional knowledge base. It reduced screening time and helped surface relevant papers.,"Budget for compute, standardising data governance across labs, and keeping evaluation practices consistent.",Improved model cards.,Cost for research workflows.
+6,"1,000-4,999 employees",Education / Research,4,"Open-source (e.g., Hugging Face)","Cost, Transparency, Collaboration",3,Yes,Accessibility and community,Course material summarisation,Infrastructure limits,Institutional SLAs,Openness > Proprietary cost
+7,500-999 employees,Education / Research,4,Both,"Transparency, Cost, Ecosystem",3,Yes,Reproducibility,Research assistant chatbot,Compute resources,Academic deployment credits,"Open first, cost later"
+8,100-499 employees,Education / Research,3,"Open-source (e.g., Hugging Face)","Cost, Learning Value",4,Yes,Freedom to experiment,Student-facing tutoring bot,Lack of skilled ops staff,Step-by-step deployment tutorials,Feasibility → Learning gain
+9,"1,000-4,999 employees",Energy / Utilities / Oil & Gas,1,Proprietary ,Data control / privacy,1,Not sure,"Managed services with SLAs and EU-region hosting simplified GDPR and NIS2 compliance and accelerated go-live, while we keep sensitive OT and customer data in a private cloud.","A pilot RAG system using HF embeddings and a local vector index over safety manuals and maintenance SOPs cut lookup time for field engineers, with all content kept inside our network.","The main issues are organisational and legal, technical friction was low in pilots.",/,"Data residency first. If a workload can stay in our VPC with clear auditability we consider open components; otherwise we prefer proprietary services with SLAs, then compare performance and TCO"
+10,"1,000-4,999 employees",Energy / Utilities / Oil & Gas,2,Both,Data control / privacy,2,Not sure,Better performance and transparency ,"Distributed optimization, optimization solvers ","Concerns about support, legal in terms of licenses or data privacy. ",It’s already good and easy to use models. Nothing to add. ,"Data privacy, quality/performance and concerns about longer term support play more important role than costs. "
+11,"1,000-4,999 employees",Energy / Utilities / Oil & Gas,2,Proprietary ,Data control / privacy,4,Yes,"Open-source: flexibility, control over models, ability to deploy on-prem for sensitive data, cost efficiency at scale.
+Proprietary: faster deployment for certain NLP and summarisation tasks, vendor support, cutting-edge capabilities.",N/A,"Technical: aligning open-source models with our internal cybersecurity and compliance frameworks.
+Organisational: upskilling engineers to work with model training pipelines.
 Legal: ensuring license compliance for commercial deployment.",Pre-built domain-specific models.,"Total cost of ownership, data sensitivity, performance benchmarks, and long-term vendor lock-in risk."
-12,100–499 employees,Energy / Utilities / Oil & Gas,4,Both,Data control / privacy,2,Yes,"Open-source: Flexibility to customize solutions to specific energy utility needs, transparency of models and code, lower total cost of ownership, rapid innovation from community contributions, and faster deployment cycles.
-Proprietary: Access to enterprise-grade support, validated compliance certifications, specialized advanced AI capabilities, and service-level agreements (SLAs) for mission-critical requirements.","We used HF for time-series anomaly detection in utility billing data. This application reduced manual billing reconciliation workload by approximately 50%, improving operational efficiency and customer satisfaction. ","Technical: Integration complexity with legacy energy ERP systems; customization of general AI models to handle large-scale, high-frequency time-series billing data.
-Organisational: Building in-house expertise for managing open-source pipelines, keeping up with rapid community updates, and aligning multidisciplinary teams.
+12,100-499 employees,Energy / Utilities / Oil & Gas,4,Both,Data control / privacy,2,Yes,"Open-source: Flexibility to customize solutions to specific energy utility needs, transparency of models and code, lower total cost of ownership, rapid innovation from community contributions, and faster deployment cycles.
+Proprietary: Access to enterprise-grade support, validated compliance certifications, specialized advanced AI capabilities, and service-level agreements (SLAs) for mission-critical requirements.","We used HF for time-series anomaly detection in utility billing data. This application reduced manual billing reconciliation workload by approximately 50%, improving operational efficiency and customer satisfaction. ","Technical: Integration complexity with legacy energy ERP systems; customization of general AI models to handle large-scale, high-frequency time-series billing data.
+Organisational: Building in-house expertise for managing open-source pipelines, keeping up with rapid community updates, and aligning multidisciplinary teams.
 Legal: Compliance with diverse international energy regulations and privacy laws, ensuring explainability and auditability of AI decisions consistent with regulatory requirements.",Account management.,"The evaluation balances regulatory compliance, customization flexibility, total cost of ownership, performance, vendor support, and data control."
-13,"1,000–4,999 employees",Energy / Utilities / Oil & Gas,3,Proprietary ,Regulatory compliance,4,Not sure,Proprietary AI solutions are preferred for mission-critical applications requiring guaranteed support and compliance.,Possible pilot projects leveraging open-source NLP or predictive models for customer service automation.,"Integration complexities with specialized energy and building systems.
-Adherence to strict regulatory and security standards.
-Internal expertise limitations for maintaining and scaling AI models.
+13,"1,000-4,999 employees",Energy / Utilities / Oil & Gas,3,Proprietary ,Regulatory compliance,4,Not sure,Proprietary AI solutions are preferred for mission-critical applications requiring guaranteed support and compliance.,Possible pilot projects leveraging open-source NLP or predictive models for customer service automation.,"Integration complexities with specialized energy and building systems.
+Adherence to strict regulatory and security standards.
+Internal expertise limitations for maintaining and scaling AI models.
 Risk management and operational continuity in critical infrastructure.","Domain-specific industrial connectors, which can be addressed via customized integration.","Innovation potential and cost advantages of open-source AI against the need for reliability, compliance, and vendor support provided by proprietary solutions. "
 14,"5,000+ employees",Energy / Utilities / Oil & Gas,3,Proprietary,"Compliance, Performance, Support",3,Not sure,Reliability and regulation fit,Exploring internal anomaly detection,Long procurement cycles,Industry-specific model registry,Compliance > Cost
-15,"1,000–4,999 employees",Energy / Utilities / Oil & Gas,3,Proprietary,"Performance, SLAs",3,Not sure,Reliability and integration,Predictive maintenance POC,Access to open datasets,Deployment templates,SLA before Openness
-16,500–999 employees,Energy / Utilities / Oil & Gas,2,Proprietary,"SLAs, Data Security",4,No,"Risk mitigation, vendor accountability",Pilot for emissions monitoring,Legal review,Role-based access guides,Security → SLA
-17,"1,000–4,999 employees",Energy / Utilities / Oil & Gas,3,Proprietary,"Cost, Performance",3,Yes,Reduce vendor lock-in risk,Evaluating model-based forecasting,Internal compliance gates,HF enterprise validation tools,SLA vs Transparency
+15,"1,000-4,999 employees",Energy / Utilities / Oil & Gas,3,Proprietary,"Performance, SLAs",3,Not sure,Reliability and integration,Predictive maintenance POC,Access to open datasets,Deployment templates,SLA before Openness
+16,500-999 employees,Energy / Utilities / Oil & Gas,2,Proprietary,"SLAs, Data Security",4,No,"Risk mitigation, vendor accountability",Pilot for emissions monitoring,Legal review,Role-based access guides,Security → SLA
+17,"1,000-4,999 employees",Energy / Utilities / Oil & Gas,3,Proprietary,"Cost, Performance",3,Yes,Reduce vendor lock-in risk,Evaluating model-based forecasting,Internal compliance gates,HF enterprise validation tools,SLA vs Transparency
 18,"5,000+ employees",Energy / Utilities / Oil & Gas,3,Proprietary,"Performance, Cost",3,Not sure,Service reliability,Document processing prototype,Integration hurdles,Compliance mapping,SLA > Cost
-19,"1,000–4,999 employees",Finance / Banking / Insurance,4,Both,Regulatory compliance,4,Not sure,Price and complexity to set up.,Quick prototyping. We have not done production facing things yet with hugging face models.,"All of the above. Lack of knowledge, compliance, skepticism, resistance to change.",On premise installations.,The support of the vendor and compliance to swiss privacy needs are important. Turnkey solutions are valued.
-20,100–499 employees,Finance / Banking / Insurance,1,Proprietary ,Regulatory compliance,5,Not sure,increase productivity,we have SyzGpt to assite all administrative tasks and AI intelligence monitoring ,banking secrecy,training and workshop,prioritize proprietary options
-21,"5,000+ employees",Finance / Banking / Insurance,3,Both,Risk management,3,Yes,No vendor lock-in and a hedge against sudden pricing or policy shifts.,"A fine-tuned, open-source LLM running on our secure cluster now drafts compliance reports by summarising new regulations, cutting analyst time ≈ 45 %.","- Regulatory compliance (EU AI Act, Basel guidance). 
-- Security & patch cadence for rapidly evolving models.
-- Model governance (lineage, bias testing, audit trails).
-- Change-management friction inside legacy risk processes.",Pre-certified model cards mapped to banking risk taxonomies. Integrated guardrail & red-teaming toolkit.,"We score each option across risk, speed, cost, performance, talent fit, and lock-in.
+19,"1,000-4,999 employees",Finance / Banking / Insurance,4,Both,Regulatory compliance,4,Not sure,Price and complexity to set up.,Quick prototyping. We have not done production facing things yet with hugging face models.,"All of the above. Lack of knowledge, compliance, skepticism, resistance to change.",On premise installations.,The support of the vendor and compliance to swiss privacy needs are important. Turnkey solutions are valued.
+20,100-499 employees,Finance / Banking / Insurance,1,Proprietary ,Regulatory compliance,5,Not sure,increase productivity,we have SyzGpt to assite all administrative tasks and AI intelligence monitoring ,banking secrecy,training and workshop,prioritize proprietary options
+21,"5,000+ employees",Finance / Banking / Insurance,3,Both,Risk management,3,Yes,No vendor lock-in and a hedge against sudden pricing or policy shifts.,"A fine-tuned, open-source LLM running on our secure cluster now drafts compliance reports by summarising new regulations, cutting analyst time ≈ 45 %.","- Regulatory compliance (EU AI Act, Basel guidance). 
+- Security & patch cadence for rapidly evolving models.
+- Model governance (lineage, bias testing, audit trails).
+- Change-management friction inside legacy risk processes.",Pre-certified model cards mapped to banking risk taxonomies. Integrated guardrail & red-teaming toolkit.,"We score each option across risk, speed, cost, performance, talent fit, and lock-in.
 If a single strategic partner can meet the risk/speed bar under a tight legal framework, we stay focused to maximise velocity. Otherwise, we diversify or use an LLM gateway to abstract providers, so we keep optionality while containing contractual complexity."
 22,"5,000+ employees",Finance / Banking / Insurance,2,Proprietary ,"Productivity enhancement, regulatory compliance, security, and ability to improve hybrid client experience.",2,No,"We chose proprietary AI solutions, particularly Microsoft Azure Open AI Service, motivated by the need for regulatory compliance, enterprise-grade support, security, and productivity gains.",Currently no prominent deployments of Hugging Face solutions.,Regulatory compliance and data privacy concerns are the major challenges faced when adopting open-source AI. ,"Enterprise-grade regulatory compliance, deployment support, integration capabilities, and assured data security would be critical to making Hugging Face’s open-source AI offerings more attractive to financial institutions.","Compliance, security, vendor support, and productivity improvements."
-23,Fewer than 100 employees,Finance / Banking / Insurance,2,Proprietary ,Cost,3,Not sure,"Open-source AI is attractive for avoiding vendor lock-in, but proprietary solutions are favoured for ease of use, support, and faster implementation given our small team.",Currently in pilot or evaluation stages; potential to use Hugging Face models for automating invoice processing and financial document summarization to reduce manual work.,"- Limited in-house AI expertise to implement and maintain models. 
+23,Fewer than 100 employees,Finance / Banking / Insurance,2,Proprietary ,Cost,3,Not sure,"Open-source AI is attractive for avoiding vendor lock-in, but proprietary solutions are favoured for ease of use, support, and faster implementation given our small team.",Currently in pilot or evaluation stages; potential to use Hugging Face models for automating invoice processing and financial document summarization to reduce manual work.,"- Limited in-house AI expertise to implement and maintain models. 
 - Data privacy and security compliance with financial regulations.",Enhanced compliance-focused resources and security best practices documentation relevant to financial data processing would also be helpful for startups in regulated industries.,"Reliability, ease of deployment and support."
 24,"5,000+ employees",Finance / Banking / Insurance,2,Proprietary,"Compliance, Data Privacy",4,Not sure,"Regulatory fit, risk control",Evaluating retrieval over policy docs,Legal approval cycles,Compliance templates and SLAs,"Governance first, cost second"
-25,"1,000–4,999 employees",Finance / Banking / Insurance,3,Both,"Cost, Support, Compliance",4,Yes,Balance cost with risk management,Pilot sentiment classifier for audit reports,"Data residency, approval delays",Private deployment guidance,Feasibility → Cost → SLA
-26,100–499 employees,Finance / Banking / Insurance,2,Proprietary,"Data Privacy, Risk Management",5,No,"Minimize exposure, ensure GDPR",None yet — early evaluation,"Security clearance, policy fit",Repository trust indicators,Compliance gates before comparison
+25,"1,000-4,999 employees",Finance / Banking / Insurance,3,Both,"Cost, Support, Compliance",4,Yes,Balance cost with risk management,Pilot sentiment classifier for audit reports,"Data residency, approval delays",Private deployment guidance,Feasibility → Cost → SLA
+26,100-499 employees,Finance / Banking / Insurance,2,Proprietary,"Data Privacy, Risk Management",5,No,"Minimize exposure, ensure GDPR",None yet - early evaluation,"Security clearance, policy fit",Repository trust indicators,Compliance gates before comparison
 27,"5,000+ employees",FMCG,2,Proprietary ,"Customization, Cost, Performance, Vendor Support, Privacy",4,Not sure,We are still developing the first solutions. Exploration phase so far.,Not aware.,technical mostly,solving what is in number 5,"Once the solution is proven to provide value implementation support is most important criteria. Since the implementations can be very large scale, support both before and after deployment is extremely important. Future roadmap must exist for the technology."
-28,500–999 employees,FMCG,2,Proprietary,"Cost, Ease of Use, Support",4,Not sure,Limited internal expertise,Testing basic sentiment analysis,Lack of internal talent,End-to-end templates,Feasibility before cost
-29,"1,000–4,999 employees",FMCG,3,Proprietary,"Cost, Vendor Reliability",4,Yes,"Faster deployment, less engineering",Social media monitoring POC,"Skills gap, integration issues",Training and tutorials,SLA > Customization
-30,100–499 employees,FMCG,2,Proprietary,"Cost, Ease of Integration",5,No,"Simplicity, vendor support",Idea tagging pilot,Lack of AI literacy,Turnkey pipelines,ROI vs Risk
-31,"1,000–4,999 employees",Government / Public Sector,1,Proprietary ,Regulatory compliance,5,No,"The preference typically leans toward proprietary AI solutions due to assured regulatory compliance, vendor accountability, and risk mitigation. Open-source AI might be considered only for non-critical or experimental projects where flexibility is needed but under stringent controls.",Use cases in government are often limited but could include pilot projects involving NLP.,"Technical challenges: integrating open-source AI with legacy government IT infrastructure.
-Organisational challenges: resistance to change, limited internal expertise, and coordination across departments.
+28,500-999 employees,FMCG,2,Proprietary,"Cost, Ease of Use, Support",4,Not sure,Limited internal expertise,Testing basic sentiment analysis,Lack of internal talent,End-to-end templates,Feasibility before cost
+29,"1,000-4,999 employees",FMCG,3,Proprietary,"Cost, Vendor Reliability",4,Yes,"Faster deployment, less engineering",Social media monitoring POC,"Skills gap, integration issues",Training and tutorials,SLA > Customization
+30,100-499 employees,FMCG,2,Proprietary,"Cost, Ease of Integration",5,No,"Simplicity, vendor support",Idea tagging pilot,Lack of AI literacy,Turnkey pipelines,ROI vs Risk
+31,"1,000-4,999 employees",Government / Public Sector,1,Proprietary ,Regulatory compliance,5,No,"The preference typically leans toward proprietary AI solutions due to assured regulatory compliance, vendor accountability, and risk mitigation. Open-source AI might be considered only for non-critical or experimental projects where flexibility is needed but under stringent controls.",Use cases in government are often limited but could include pilot projects involving NLP.,"Technical challenges: integrating open-source AI with legacy government IT infrastructure.
+Organisational challenges: resistance to change, limited internal expertise, and coordination across departments.
 Legal challenges: compliance with strict data privacy, security regulations, and procurement rules.",Features for transparent auditing and explainability to meet governance requirements.,"We prioritize risk mitigation, regulatory compliance, and data governance when evaluating trade-offs. Proprietary AI solutions are preferred for critical and sensitive applications due to guaranteed vendor support, accountability, and compliance assurances."
 32,Fewer than 100 employees,Government / Public Sector,2,Proprietary ,Regulatory compliance,5,No,"Proprietary AI is favoured for assured compliance, support, and risk management. Open-source AI may be used cautiously only in pilot or experimental projects that do not impact critical operations.",Limited use cases primarily involving pilot NLP projects.,Integration with legacy systems. Ensuring strict regulatory and security compliance. Lack of internal expertise. Resistance to change within the organization. Procurement and governance hurdles.,Transparency and auditing features for governance.,"The organisation prioritizes risk mitigation, regulatory compliance, and accountability. "
-33,"5,000+ employees",Government / Public Sector,3,Both,Regulatory compliance,4,Not sure,"Open-source AI is valued for transparency, adaptability to local needs, and alignment with European digital sovereignty goals.
+33,"5,000+ employees",Government / Public Sector,3,Both,Regulatory compliance,4,Not sure,"Open-source AI is valued for transparency, adaptability to local needs, and alignment with European digital sovereignty goals.
 Proprietary AI is used for mission‑critical systems where vendor accountability, guaranteed SLAs, and compliance certification are essential.",Pilot projects focused on natural language processing for public health information services.,"Ensuring strict compliance with GDPR and specific health data privacy laws. Integrating open-source AI tools into complex, multi-agency public health IT ecosystems and legacy systems.",Regulatory-focused deployment guides.,"Adaptability, transparency, and innovation afforded by open-source AI with the operational stability, vendor accountability, and regulatory assurances offered by proprietary solutions."
-34,"5,000+ employees",Government / Public Sector,3,Both,Regulatory compliance,4,Yes,"Open-source AI is valued for transparency, adaptability, multilingual capability.
-Proprietary AI is chosen for mission-critical, classified, or legally sensitive operations requiring certified vendor support.",In policy research and internal knowledge management pilots.,"- Confidentiality and security mandate alignment
-- Integration with complex, security‑hardened federal systems
-- Multilingual performance management and bias mitigation
+34,"5,000+ employees",Government / Public Sector,3,Both,Regulatory compliance,4,Yes,"Open-source AI is valued for transparency, adaptability, multilingual capability.
+Proprietary AI is chosen for mission-critical, classified, or legally sensitive operations requiring certified vendor support.",In policy research and internal knowledge management pilots.,"- Confidentiality and security mandate alignment
+- Integration with complex, security‑hardened federal systems
+- Multilingual performance management and bias mitigation
 - Compliance assurance and accuracy consistency in model updates",Model cards with more detailed information.,"Transparency, adaptability, and sovereignty benefits of open-source AI against the guaranteed SLAs, vendor accountability, and certification of proprietary AI."
-35,500–999 employees,Government / Public Sector,2,Proprietary ,Regulatory compliance,3,Not sure,"Proprietary AI is valued for reliability, turnkey deployment, and vendor accountability in the public education sector.",No direct open‑source model deployment. ,Regulatory compliance.,Clear compliance and deployment guides for education-sector AI.,"Reliability, compliance, and vendor accountability, leading to a preference for proprietary systems in national rollouts."
+35,500-999 employees,Government / Public Sector,2,Proprietary ,Regulatory compliance,3,Not sure,"Proprietary AI is valued for reliability, turnkey deployment, and vendor accountability in the public education sector.",No direct open‑source model deployment. ,Regulatory compliance.,Clear compliance and deployment guides for education-sector AI.,"Reliability, compliance, and vendor accountability, leading to a preference for proprietary systems in national rollouts."
 36,"5,000+ employees",Government / Public Sector,2,Proprietary,"Documentation, Compliance",5,Not sure,"Transparency, accountability",Policy analysis prototype,Security posture and data isolation,Public-sector audit kits,Compliance > Cost
-37,100–499 employees,Government / Public Sector,2,Proprietary,"Governance, Data Privacy",4,No,"Auditability, traceability",None yet — feasibility study,Legal review and IT policy,Public-use compliance tools,GDPR check → SLA
-38,"1,000–4,999 employees",Government / Public Sector,3,Both,"Performance, Compliance",3,Yes,Balance open vs closed ecosystems,Retrieval over public open data,Procurement friction,Licensing examples,Policy gate → Performance
+37,100-499 employees,Government / Public Sector,2,Proprietary,"Governance, Data Privacy",4,No,"Auditability, traceability",None yet - feasibility study,Legal review and IT policy,Public-use compliance tools,GDPR check → SLA
+38,"1,000-4,999 employees",Government / Public Sector,3,Both,"Performance, Compliance",3,Yes,Balance open vs closed ecosystems,Retrieval over public open data,Procurement friction,Licensing examples,Policy gate → Performance
 39,"5,000+ employees",Government / Public Sector,3,Proprietary,"Data Privacy, Vendor Dependence",5,Not sure,Reduce lock-in,Not yet piloted,Integration and internal review,Government-tailored guidance,GDPR and sovereignty first
 40,Fewer than 100 employees,Healthcare / Biotech / Life Sciences,3,Proprietary ,Data control / privacy,4,Not sure,Not there yet,None.,N/A,N/A,"Security and compliance, performance and accuracy, cost and resource allocation."
 41,Fewer than 100 employees,Healthcare / Biotech / Life Sciences,3,Both,Performance,3,Not sure,To balance customization and cost with stringent regulatory compliance and data privacy requirements inherent in healthcare.,Open-source AI solutions have helped develop clinical decision support tools that enhance diagnostic accuracy and patient care by leveraging natural language processing and machine learning models.,Challenges include ensuring compliance with healthcare regulations and addressing data security and confidentiality concerns.,Dedicated enterprise support.,"The evaluation focuses on customization and cost benefits of open-source versus reliability, vendor support, and compliance assurances of proprietary solutions."
 42,Fewer than 100 employees,Healthcare / Biotech / Life Sciences,5,"Open-source (e.g., Hugging Face)",Data control / privacy,2,Yes,"Scientific progress, reproducibility, and network effects that benefit the entire community.","Our flagship model created substantial value by demonstrating our capabilities to potential enterprise customers, generating research collaborations, and establishing our reputation in the industry.","Data aggregation, talent, complex data licensing.",Enhanced model cards.,"Open-source: reproducibility, community validation, and broader adoption. "
 43,Fewer than 100 employees,Healthcare / Biotech / Life Sciences,4,Both,Regulatory compliance,3,Yes,Innovation speed from open-source and reliability/compliance from proprietary AI.,"We use open-source NLP models as benchmarks and components in our platform for tasks like document analysis, metadata extraction, and literature monitoring. These open-source foundations enable us to rapidly prototype and validate new features before full proprietary implementation.",Technical: adapting general-purpose models for highly specialized medical and regulatory content. Legal: we must ensure all AI components meet stringent healthcare data privacy requirements and regulatory standards across multiple jurisdictions.,More robust enterprise-grade security features and audit trails.,"Regulatory compliance requirements, performance for specialised medical tasks, and total cost of ownership. "
-44,Fewer than 100 employees,Healthcare / Biotech / Life Sciences,3,Both,Performance,3,Yes,"Open-source to accelerate research into single-cell workflows and enable rapid prototyping of biomarker discovery pipelines. Proprietary for validated, production-grade deployments that meet regulatory requirements and guarantee data security for client studies.","We evaluated an open-source variational inference model for single-cell data as a foundation for our multimodal integration pipeline. It reduced dimensionality and batch-effect correction time by 60%, enabling faster identification of disease-specific cell populations for our diagnostic partners.","- Harmonizing heterogeneous single-cell datasets with varying formats and quality.
-- Ensuring reproducibility and traceability in regulated studies (audit trails, data lineage).
+44,Fewer than 100 employees,Healthcare / Biotech / Life Sciences,3,Both,Performance,3,Yes,"Open-source to accelerate research into single-cell workflows and enable rapid prototyping of biomarker discovery pipelines. Proprietary for validated, production-grade deployments that meet regulatory requirements and guarantee data security for client studies.","We evaluated an open-source variational inference model for single-cell data as a foundation for our multimodal integration pipeline. It reduced dimensionality and batch-effect correction time by 60%, enabling faster identification of disease-specific cell populations for our diagnostic partners.","- Harmonizing heterogeneous single-cell datasets with varying formats and quality.
+- Ensuring reproducibility and traceability in regulated studies (audit trails, data lineage).
 - Recruiting talent with combined expertise in deep learning and advanced biology.","Turnkey, life-sciences-specific model card templates with provenance and validation sections.","We prioritize open-source AI for research and pilot phases to benefit from community innovations and low entry-cost experimentation. For client-facing products, we layer proprietary tooling on top of open-source models or choose enterprise-grade vendors to ensure regulatory compliance and consistent performance."
-45,500–999 employees,Healthcare / Biotech / Life Sciences,3,Both,"Compliance, Data Control",4,Yes,"Data protection, patient privacy",Prototype de-identification workflow,Documentation and traceability,Health-specific compliance templates,"GDPR first, then performance"
-46,100–499 employees,Healthcare / Biotech / Life Sciences,2,Proprietary,"Compliance, Documentation",4,Not sure,"Transparency, traceability",Planning for medical text summarisation,License ambiguity,Legal templates and clearer model cards,"Governance filter, no shortcuts"
-47,"1,000–4,999 employees",Healthcare / Biotech / Life Sciences,3,Both,"Cost, Support, Time to Value",3,Yes,Fine-tuning capability,Annotating patient feedback for triage,Model reliability,Enterprise SLAs,Feasibility check → Compare cost
+45,500-999 employees,Healthcare / Biotech / Life Sciences,3,Both,"Compliance, Data Control",4,Yes,"Data protection, patient privacy",Prototype de-identification workflow,Documentation and traceability,Health-specific compliance templates,"GDPR first, then performance"
+46,100-499 employees,Healthcare / Biotech / Life Sciences,2,Proprietary,"Compliance, Documentation",4,Not sure,"Transparency, traceability",Planning for medical text summarisation,License ambiguity,Legal templates and clearer model cards,"Governance filter, no shortcuts"
+47,"1,000-4,999 employees",Healthcare / Biotech / Life Sciences,3,Both,"Cost, Support, Time to Value",3,Yes,Fine-tuning capability,Annotating patient feedback for triage,Model reliability,Enterprise SLAs,Feasibility check → Compare cost
 48,Fewer than 100 employees,Hospitality and Tourism Industry,3,Proprietary ,Performance,4,Not sure,"Reliability, brand safety, client compliance needs, and integration with content and production systems.",Currently none,"Governance/compliance, brand-safety controls for generated content, and integration with production/logistics/content systems.",Enriched model cards.,"Reliability, brand safety, compliance, and support (proprietary) against flexibility and cost (open-source)."
-49,100–499 employees,Hospitality and Tourism Industry,2,Proprietary,"Cost, Time to Value",3,Yes,Ease of entry,RAG chatbot for guest FAQs,Lack of skilled staff,Simple deployment SDKs,Feasibility before ROI
+49,100-499 employees,Hospitality and Tourism Industry,2,Proprietary,"Cost, Time to Value",3,Yes,Ease of entry,RAG chatbot for guest FAQs,Lack of skilled staff,Simple deployment SDKs,Feasibility before ROI
 50,Fewer than 100 employees,Hospitality and Tourism Industry,3,Both,"Support, Compliance",3,Yes,Data ownership and flexibility,Internal knowledge base assistant,Limited documentation,Step-by-step guides,Security and GDPR checks first
-51,500–999 employees,Hospitality and Tourism Industry,2,Proprietary,"Data Privacy, Performance",4,Not sure,Better customer experience,Pilot chatbot in sandbox,Integration complexity,Clearer hosting options,Privacy and cost
-52,100–499 employees,Manufacturing / Industrial,1,Neither,Regulatory compliance,4,Yes,We have not adopted either yet. Compliance and supplier-assurance gates come first,Currently none.,Technical and organisational.,Not sure yet.,"Compliance and safety first, then security and support"
-53,"1,000–4,999 employees",Manufacturing / Industrial,1,Neither,Data control / privacy,3,Not sure,Productivity increase ,"No production use yet. A small on premises pilot summarised maintenance SOPs and incident reports with HF models, reducing lookup time for shift leads while keeping data inside our network.",Fear about data confidentiality ,,"Data confidentiality, IP protection, and GDPR compliance first. If those pass, we assess security and support, then task performance and total cost of ownership."
+51,500-999 employees,Hospitality and Tourism Industry,2,Proprietary,"Data Privacy, Performance",4,Not sure,Better customer experience,Pilot chatbot in sandbox,Integration complexity,Clearer hosting options,Privacy and cost
+52,100-499 employees,Manufacturing / Industrial,1,Neither,Regulatory compliance,4,Yes,We have not adopted either yet. Compliance and supplier-assurance gates come first,Currently none.,Technical and organisational.,Not sure yet.,"Compliance and safety first, then security and support"
+53,"1,000-4,999 employees",Manufacturing / Industrial,1,Neither,Data control / privacy,3,Not sure,Productivity increase ,"No production use yet. A small on premises pilot summarised maintenance SOPs and incident reports with HF models, reducing lookup time for shift leads while keeping data inside our network.",Fear about data confidentiality ,,"Data confidentiality, IP protection, and GDPR compliance first. If those pass, we assess security and support, then task performance and total cost of ownership."
 54,"5,000+ employees",Manufacturing / Industrial,3,Both,Regulatory compliance,4,No,"Regulatory and safety constraints drive the use of proprietary AI, while open-source is adopted for non-safety-critical R&D and predictive-maintenance analytics.",We used open-source NLP models to analyze maintenance logs and service bulletins. This automated readability scoring and anomaly detection reduced manual triage time by 40%.,"Compliance with EASA certification requirements, intellectual-property and export-control restrictions, aligning cross-functional processes between R&D, engineering, and safety authorities.",Built-in audit trails.,"Innovation speed and flexibility (open-source) against the need for formal certification, traceability, and vendor accountability (proprietary)."
 55,Fewer than 100 employees,Manufacturing / Industrial,4,Both,Performance,3,Not sure,Open-source: customization of anomaly-detection models. Proprietary AI: enterprise-grade security.,We used Hugging Face on time-series sensor data to detect pump anomalies that reduced unscheduled downtime and maintenance costs.,Data quality and consistency across legacy SCADA systems. Model integration into existing OT networks with strict uptime requirements,/,"Total cost of ownership and flexibility against operational reliability, SLAs, and integration ease."
-56,500–999 employees,Manufacturing / Industrial,2,Proprietary ,Vendor support,4,Not sure,"We needed rapid deployment with clear support and certifications that fit our quality system. Managed services simplified security reviews, change control, and integration with our OT environment.","Vision based defect detection on the production line using a managed model and pipeline. Reliability, monitoring, and traceability were decisive.","Validating open components against safety and quality processes, standing up on premises GPU operations, and ensuring end to end traceability for audits.",Example change-control records to speed quality reviews.,"Compliance risk and time to production are the gates. Also, flexibility and total cost of ownership."
+56,500-999 employees,Manufacturing / Industrial,2,Proprietary ,Vendor support,4,Not sure,"We needed rapid deployment with clear support and certifications that fit our quality system. Managed services simplified security reviews, change control, and integration with our OT environment.","Vision based defect detection on the production line using a managed model and pipeline. Reliability, monitoring, and traceability were decisive.","Validating open components against safety and quality processes, standing up on premises GPU operations, and ensuring end to end traceability for audits.",Example change-control records to speed quality reviews.,"Compliance risk and time to production are the gates. Also, flexibility and total cost of ownership."
 57,"5,000+ employees",Media & Entertainment,4,Both,Performance,3,Yes,On-prem deployment for PII-heavy content pipelines; full inspectability for legal review; flexible fine-tuning on our storytelling IP.,"Fine-tuning an open-source multilingual model on 100 years of subtitles automatically tags each scene with characters, locations, and emotions, shrinking clip-search from hours to seconds and saving six-figure annual costs.","Licence clarity, Model-card completeness, Inference cost spikes, change management.","Built-in policy enforcement hooks (e.g., export-compliance checks), Curated, security-scanned model zoo flagged for commercial use","Latency goal, data control, accuracy ceiling, reg-compliance, total cost."
-58,500–999 employees,Media & Entertainment,4,Both,"Performance, Cost",3,Yes,"Creativity, Adaptability",Caption generation,IP review,Model card clarity,Performance > Cost
-59,"1,000–4,999 employees",Media & Entertainment,4,Both,"Cost, SLAs",3,Yes,Agile iteration,Metadata tagging for archives,Compliance,License examples,SLA vs openness
+58,500-999 employees,Media & Entertainment,4,Both,"Performance, Cost",3,Yes,"Creativity, Adaptability",Caption generation,IP review,Model card clarity,Performance > Cost
+59,"1,000-4,999 employees",Media & Entertainment,4,Both,"Cost, SLAs",3,Yes,Agile iteration,Metadata tagging for archives,Compliance,License examples,SLA vs openness
 60,Fewer than 100 employees,Retail / eCommerce,1,Proprietary ,Cost,4,Not sure,Proprietary solutions were the only one known so we took it,It reduce the time we are spending on creating content mainly,Understand how AI could help and How it could be helpful to our work,Nothing to add.,Vendor support
 61,"5,000+ employees",Retail / eCommerce,1,Proprietary ,Data control / privacy,5,Not sure,"We handle customer personal data at scale. Proprietary services gave us EU region hosting, SLAs, and audit-ready documentation for GDPR and security reviews, which reduced time to production.",Not in production.,"DPIA preparation, data minimisation and PII redaction, uncertainty over licence terms for commercial use, and hardening self-hosted inference with proper monitoring and patching.",Org-level policy enforcement with automatic GDPR and EU AI Act documentation export.,"Compliance, data residency, and security first"
 62,"5,000+ employees",Retail / eCommerce,5,Both,Scalability and cost.,4,Yes,"Open-source AI gives us agility, cost savings, and access to vibrant collaborative communities. Proprietary AI is preferred for mission-critical systems requiring vendor reliability and compliance guarantees.",AI-powered assistants leveraging a combination of internally built models and external large language models to enhance customer experience.,"Model adaptation to unique fashion retail data and customer preferences, brand-aligned content quality and preventing bias in AI outputs, integrating open-source tools within complex proprietary platforms, GDPR compliance in AI data handling and deployments.",Nothing here!,"Innovation speed, transparency, and community benefits of open-source AI vs. the operational stability, vendor accountability, and legal assurances of proprietary solutions."
 63,Fewer than 100 employees,Retail / eCommerce,3,Proprietary ,Performance,4,Yes,"Open-source AI may be used for experimentation, research, or components to speed up development and tap into community innovation. Proprietary solutions for specialized image generation quality, and scalability.",Research and rapid prototyping in content personalization and user interaction.,Fine-tuning models to ensure brand-consistent outputs without distortion and maintaining data privacy and compliance especially when dealing with user personalization.,"More templates focused on image generation, fashion trend analysis, and personalization workflows.","Proprietary AI is preferred for core product engineering and customer-facing features, while open-source is valuable for pilots and augmenting capabilities."
 64,Fewer than 100 employees,Retail / eCommerce,4,Both,"Performance, customization, deployment speed",2,Yes,Open-source: prototyping of our social‐media image analysis pipelines. Proprietary: managed infrastructure and vendor support for production scalability.,"We leveraged open-source models to build our trend‐detection models, reducing prototype time by 50% and improving seasonal trend prediction accuracy by 12%.",Data compliance when scraping and processing social media imagery across jurisdictions,-,"We use open-source AI for R&D and proof-of-concepts, then transition successful models to proprietary, managed services when we need enterprise-grade reliability, support SLAs, and compliance assurances for production."
-65,100–499 employees,Retail / eCommerce,5,Both,Customization,3,Yes,"Proprietary solutions offer managed services and commercial SLAs. Open-source enables customization, transparency, and cost control.",We integrated Hugging Face transformers into our product recommendations engine. The solution increased recommendation click-through rates and boosted average order value.,"Model governance and version control across multiple teams, data-privacy and compliance requirements, scaling inference infrastructure cost-effectively under variable traffic patterns.",Advanced monitoring dashboards with custom alerting on model drift.,"> Total cost of ownership (licensing vs. support)
-
-> Speed of integration (turnkey APIs vs. custom deployments)
-
-> Security and compliance needs
-
+65,100-499 employees,Retail / eCommerce,5,Both,Customization,3,Yes,"Proprietary solutions offer managed services and commercial SLAs. Open-source enables customization, transparency, and cost control.",We integrated Hugging Face transformers into our product recommendations engine. The solution increased recommendation click-through rates and boosted average order value.,"Model governance and version control across multiple teams, data-privacy and compliance requirements, scaling inference infrastructure cost-effectively under variable traffic patterns.",Advanced monitoring dashboards with custom alerting on model drift.,"> Total cost of ownership (licensing vs. support)
+
+> Speed of integration (turnkey APIs vs. custom deployments)
+
+> Security and compliance needs
+
 > Long-term roadmap alignment with vendor vs. community innovation trajectory"
-66,"1,000–4,999 employees",Retail / eCommerce,4,Both,"Cost, Performance, Support",3,Yes,Lower TCO and flexibility,Product search enhancement,Integration with legacy stack,Clearer SDKs,Compare ROI and SLA
-67,500–999 employees,Retail / eCommerce,3,Both,"Performance, Time to Value",3,Yes,Faster iteration,RAG over product manuals,Model updates,Update notifications,Test feasibility → Cost
-68,100–499 employees,Retail / eCommerce,3,Proprietary,"Support, SLAs",4,Not sure,SLA reliability,Recommendation bot POC,Limited engineering,Enterprise-tier support,SLA > Openness
+66,"1,000-4,999 employees",Retail / eCommerce,4,Both,"Cost, Performance, Support",3,Yes,Lower TCO and flexibility,Product search enhancement,Integration with legacy stack,Clearer SDKs,Compare ROI and SLA
+67,500-999 employees,Retail / eCommerce,3,Both,"Performance, Time to Value",3,Yes,Faster iteration,RAG over product manuals,Model updates,Update notifications,Test feasibility → Cost
+68,100-499 employees,Retail / eCommerce,3,Proprietary,"Support, SLAs",4,Not sure,SLA reliability,Recommendation bot POC,Limited engineering,Enterprise-tier support,SLA > Openness
 69,"5,000+ employees",Retail / eCommerce,4,Both,"Cost, Customization",3,Yes,Internalization control,Classification for returns,Model drift,Monitoring tools,Compare TCO vs reliability
-70,"1,000–4,999 employees",Retail / eCommerce,3,Both,"Performance, Customization",2,Yes,Domain adaptation,Copilot for customer queries,GPU constraints,Fine-tuning templates,Compare accuracy vs cost
+70,"1,000-4,999 employees",Retail / eCommerce,3,Both,"Performance, Customization",2,Yes,Domain adaptation,Copilot for customer queries,GPU constraints,Fine-tuning templates,Compare accuracy vs cost
 71,"5,000+ employees",Technology / Software,1,Proprietary ,Data control / privacy,1,No,"Customer zero, using in-house AI solution. Furthermore, security is always a concern. ","Unfortunately not, haven’t used it in the pass ",No open-source AI adoption ,Not able to share internal information ,No part of the decision making process 
-72,"5,000+ employees",Technology / Software,3,Proprietary ,Performance,,Not sure,"We mainly use Copilot for integrated apps use cases (e.g. Copilot for Outlook, PowerPoint, Excel, search etc...) and OpenAI for custom use cases; both are used because 1) we are producing the product or have a stake in the company developing them and 2) because of the fit for purpose","We have customers heavily using Azure OpenAI for use cases such as chatbots, pre-triage customer service, documentation processing etc.. for example we have a client (an elevator manufacturer) which trains OpenAI on technical documentation of elevators and provides a chatbot for technicians in the field to answer questions about the specific model of elevator they need to work on","We see clients mostly struggle with the following:
-1) Technical: AI models need to consume data and give outputs, most of the time it's not plug and play... they need to build interfaces between the AI models and their data (e.g. structured / unstructured databases, documents, SharePoint sites etc...). So putting all of this together can be challenging.
-2) Chargeback: a common design pattern is to share an Azure OpenAI resource across different apps, which implies building a chargeback/measurement mechanism around the consumption of the AI models. This mechanism is frequently a source of friction for our clients.
-3) Governance: many clients want to control the information their employees share with public AI services (e.g. ChatGPT) so they build their own chatbot internally, also trained on all the available ""public web"" but they run it themselves internally to ensure confidential information is not exfiltrated from the company. This is quite a common approach, and some clients even go beyond and implement some additional governance to verify their employees don't ask offensive questions etc... I personally don't understand the interest of doing this but I've seen a major watch manufacturer build this and it was clearly a source of challenge.
-
-Also on governance: many clients want to control the information that is shared with cloud providers' AI services (e.g. Azure OpenAI service); this control is generally challenging for them to implement.","I've not seen clients use HF yet, only Azure OpenAI","We don't use it ourselves but I know from clients:
-1) Governance / data protection
-2) Usability
+72,"5,000+ employees",Technology / Software,3,Proprietary ,Performance,,Not sure,"We mainly use Copilot for integrated apps use cases (e.g. Copilot for Outlook, PowerPoint, Excel, search etc...) and OpenAI for custom use cases; both are used because 1) we are producing the product or have a stake in the company developing them and 2) because of the fit for purpose","We have customers heavily using Azure OpenAI for use cases such as chatbots, pre-triage customer service, documentation processing etc.. for example we have a client (an elevator manufacturer) which trains OpenAI on technical documentation of elevators and provides a chatbot for technicians in the field to answer questions about the specific model of elevator they need to work on","We see clients mostly struggle with the following:
+1) Technical: AI models need to consume data and give outputs, most of the time it's not plug and play... they need to build interfaces between the AI models and their data (e.g. structured / unstructured databases, documents, SharePoint sites etc...). So putting all of this together can be challenging.
+2) Chargeback: a common design pattern is to share an Azure OpenAI resource across different apps, which implies building a chargeback/measurement mechanism around the consumption of the AI models. This mechanism is frequently a source of friction for our clients.
+3) Governance: many clients want to control the information their employees share with public AI services (e.g. ChatGPT) so they build their own chatbot internally, also trained on all the available ""public web"" but they run it themselves internally to ensure confidential information is not exfiltrated from the company. This is quite a common approach, and some clients even go beyond and implement some additional governance to verify their employees don't ask offensive questions etc... I personally don't understand the interest of doing this but I've seen a major watch manufacturer build this and it was clearly a source of challenge.
+
+Also on governance: many clients want to control the information that is shared with cloud providers' AI services (e.g. Azure OpenAI service); this control is generally challenging for them to implement.","I've not seen clients use HF yet, only Azure OpenAI","We don't use it ourselves but I know from clients:
+1) Governance / data protection
+2) Usability
 3) Costs"
 73,Fewer than 100 employees,Technology / Software,1,Proprietary ,Performance,1,Not sure,Ease of use and deployment speed,Chatbot for software documentation,/,/,/
-74,"1,000–4,999 employees",Technology / Software,3,Both,Regulatory compliance,4,Not sure,"Data privacy, security, customer demand, employee demand",We have an internal AI tool available based on Ollama and Open Web UI. There may be more project usecases from our data and AI business line that I'm not aware of.,My guess: providing the hardware necessary and justifying the cost.,maybe integrated cloud hosting or deployment solutions. Also model training capabilities to speed up processes and access to or pooling of specific training data,Usually we offer both based on customer demand. We have a big Microsoft solutions division that offers AI services through Azure. We also train models with customer data for specific use cases. It depends largely on cost and demand
+74,"1,000-4,999 employees",Technology / Software,3,Both,Regulatory compliance,4,Not sure,"Data privacy, security, customer demand, employee demand",We have an internal AI tool available based on Ollama and Open Web UI. There may be more project usecases from our data and AI business line that I'm not aware of.,My guess: providing the hardware necessary and justifying the cost.,maybe integrated cloud hosting or deployment solutions. Also model training capabilities to speed up processes and access to or pooling of specific training data,Usually we offer both based on customer demand. We have a big Microsoft solutions division that offers AI services through Azure. We also train models with customer data for specific use cases. It depends largely on cost and demand
 75,"5,000+ employees",Technology / Software,3,Proprietary ,Regulatory compliance,4,Yes,acceleration of development,,"main challenges are legal and compliances, then integration to existing workflows",,compliance and security
 76,"5,000+ employees",Technology / Software,4,Proprietary ,Performance,3,Yes,"We chose proprietary models for top performance, mature tooling, and latency SLAs for customer-facing features. We plan to expand open source for fine tuning on internal data and to reduce long-term TCO.",We used HF models and datasets to benchmark candidate LLMs and to power a RAG prototype over product docs. The open setup let us customise retrieval and iterate quickly without exposing customer data.,"License interpretation for commercial use, setting up reproducible eval pipelines, dependency scanning for serving images, and aligning internal OSS policies for contributions and model sharing.",," Security, privacy, and customer commitments."
 77,"5,000+ employees",Technology / Software,5,Both,Performance,3,Yes,"Hybrid for performance and speed: proprietary for SLA backed features, HF open source to fine tune our own LLMs in cloud for cost control and avoiding lock in.",Developing our own open source LLMs,Integration and Support from model Provider in AI frameworks,Easier Integration into Azure as Infra stack,Performance use case driven 
 78,"5,000+ employees",Technology / Software,5,Both,All of the above with Data control/privacy / Compliance & Risk at the forefront,5,Yes,Having the broadest possible portfolio of models available to us and our customers.,We sometimes (or our customers) use specialized OpenSource  models when specific needs arise.,"Quality, security, compliance, etc. of the models. Basically we have our Responsible AI principles and it is very difficult to assess all models at the same level that we do for our 1st party or even 3rd party managed models. ",We just got it a couple of weeks ago. Hugging Face is not integrated in Azure AI Foundry ;-),"Based on our Responsible AI principles, then it will depend on the use case: capability, cost, ease of use, ease of fine tuning, etc."
-79,500–999 employees,Technology / Software,4,Both,Customization,3,Yes,"Eliminate single-vendor risk, fine-grained control over weights & infra, lower TCO at scale, and stronger data-sovereignty guarantees.","Automated pull-request review summaries. It now drafts ~70 % of review notes, cutting engineer PR turnaround by 32 hours per sprint.",Technical: optimising GPU utilisation and memory footprint for 24/7 inference. Organisational: upskilling DevSecOps on model-license nuances. Legal: aligning Apache 2.0 & GPL-derived dependencies with customer redistribution clauses.,A push-time license-compliance gate.,"Total cost of ownership, latency & accuracy benchmarks, licence obligations, roadmap stability, and lock-in risk."
-80,"5,000+ employees",Technology / Software,4,Both,Data control / privacy,3,Yes,"Open Source for: 1. Customisability & speed-to-experiment: we can fine-tune quickly on small, domain-specific datasets. 2. Deployment flexibility: on-prem, edge or air-gapped for nuclear and defence customers. 3. Cost transparency: no usage-based surprises during large-scale inferencing.
+79,500-999 employees,Technology / Software,4,Both,Customization,3,Yes,"Eliminate single-vendor risk, fine-grained control over weights & infra, lower TCO at scale, and stronger data-sovereignty guarantees.","Automated pull-request review summaries. It now drafts ~70 % of review notes, cutting engineer PR turnaround by 32 hours per sprint.",Technical: optimising GPU utilisation and memory footprint for 24/7 inference. Organisational: upskilling DevSecOps on model-license nuances. Legal: aligning Apache 2.0 & GPL-derived dependencies with customer redistribution clauses.,A push-time license-compliance gate.,"Total cost of ownership, latency & accuracy benchmarks, licence obligations, roadmap stability, and lock-in risk."
+80,"5,000+ employees",Technology / Software,4,Both,Data control / privacy,3,Yes,"Open Source for: 1. Customisability & speed-to-experiment: we can fine-tune quickly on small, domain-specific datasets. 2. Deployment flexibility: on-prem, edge or air-gapped for nuclear and defence customers. 3. Cost transparency: no usage-based surprises during large-scale inferencing.
 Proprietary services still win when we need state-of-the-art accuracy out-of-the-box or robust vendor SLAs. ","Used on 2 M maintenance work-order logs to auto-classify failure modes and recommend remedial actions. Mean-time-to-repair dropped 8 %, saving ≈ US$3 M annually.","Technical, organisational (Upskilling engineers), legal / IP.","Industrial-Protocol Streaming Connectors to enable fast, reliable deployment of AI in industrial environments.","Data sovereignty & privacy, total cost of ownership over 3 years, performance on domain benchmarks, regulatory & safety compliance fit, vendor / community support & roadmap."
-81,Fewer than 100 employees,Technology / Software,4,Both,Regulatory compliance,2,Yes,"Open-source → rapid experimentation, transparent audits, fine-grained weight control.
-Proprietary → turnkey scalability, specialised hardware, indemnification.
-Running both provides cost leverage, bias reduction, and operational fail-over.",A module classifies >50 k contractual clauses per hour to flag regulatory risk for European banks. A fine-tuned model from Hugging Face cut review time by 73 % while boosting precision by 11 pp.,"OSS license & supply-chain vetting
-Silent upstream checkpoint changes (solved by pinning SHAs)
-GDPR/data-residency assessments
-Change-management for non-technical stakeholders",Model cards (eg. make it crystal-clear where—and where not—the model should be applied).,"Compliance, Cost, Control, Performance, Vendor-Risk. Proprietary APIs lead on latency and ease, open-source wins on transparency and strategic flexibility. We keep two functionally equivalent models (one OSS, one proprietary) in production to hedge outages and policy shifts, revisiting scores quarterly."
-82,Fewer than 100 employees,Technology / Software,5,Both,Customization,2,Yes,"We needed speed to market and long-term freedom. Our abstraction layer means we can swap in open-source models without rewriting pipelines, balancing vendor lock-in risk, cost, and performance.",We fine-tuned a Mistral-7B Instruct model on company-specific support chats via Hugging Face. The model now auto-drafts tier-1 support replies with a 67 % reduction in handling time and zero extra API cost.,"Keeping pace with weekly model releases and benchmarks.
-GPU scarcity for fine-tuning larger models.
+81,Fewer than 100 employees,Technology / Software,4,Both,Regulatory compliance,2,Yes,"Open-source → rapid experimentation, transparent audits, fine-grained weight control.
+Proprietary → turnkey scalability, specialised hardware, indemnification.
+Running both provides cost leverage, bias reduction, and operational fail-over.",A module classifies >50 k contractual clauses per hour to flag regulatory risk for European banks. A fine-tuned model from Hugging Face cut review time by 73 % while boosting precision by 11 pp.,"OSS license & supply-chain vetting
+Silent upstream checkpoint changes (solved by pinning SHAs)
+GDPR/data-residency assessments
+Change-management for non-technical stakeholders",Model cards (eg. make it crystal-clear where-and where not-the model should be applied).,"Compliance, Cost, Control, Performance, Vendor-Risk. Proprietary APIs lead on latency and ease, open-source wins on transparency and strategic flexibility. We keep two functionally equivalent models (one OSS, one proprietary) in production to hedge outages and policy shifts, revisiting scores quarterly."
+82,Fewer than 100 employees,Technology / Software,5,Both,Customization,2,Yes,"We needed speed to market and long-term freedom. Our abstraction layer means we can swap in open-source models without rewriting pipelines, balancing vendor lock-in risk, cost, and performance.",We fine-tuned a Mistral-7B Instruct model on company-specific support chats via Hugging Face. The model now auto-drafts tier-1 support replies with a 67 % reduction in handling time and zero extra API cost.,"Keeping pace with weekly model releases and benchmarks.
+GPU scarcity for fine-tuning larger models.
 Licensing clarity.","Observability & experimentation, optimised infra & cost control.","Time-to-value, total cost of ownership, privacy/compliance, and adaptability"
-83,100–499 employees,Technology / Software,4,Both,Cost,3,Yes,"Open-source models gave us speed and flexibility. We could inspect the code, adapt architectures, and fine-tune without licensing delays. They lowered initial costs and let us experiment broadly.
-
-Proprietary models gave us performance and support guarantees we could not match in-house. For certain workloads they offered higher accuracy, better latency, or compliance features that reduced operational risk.
-
-We use open-source when control, customization, or rapid iteration matters. We use proprietary when reliability, security, or competitive performance justifies the cost. The mix minimizes lock-in and maximizes delivery speed.","We used open-source AI to power semantic search in our customer support knowledge base. Fine-tuning it on our historical support tickets let users find answers by intent rather than exact keywords.
-
+83,100-499 employees,Technology / Software,4,Both,Cost,3,Yes,"Open-source models gave us speed and flexibility. We could inspect the code, adapt architectures, and fine-tune without licensing delays. They lowered initial costs and let us experiment broadly.
+
+Proprietary models gave us performance and support guarantees we could not match in-house. For certain workloads they offered higher accuracy, better latency, or compliance features that reduced operational risk.
+
+We use open-source when control, customization, or rapid iteration matters. We use proprietary when reliability, security, or competitive performance justifies the cost. The mix minimizes lock-in and maximizes delivery speed.","We used open-source AI to power semantic search in our customer support knowledge base. Fine-tuning it on our historical support tickets let users find answers by intent rather than exact keywords.
+
 Result: support ticket deflection increased 28%, average resolution time dropped by over a third, and we avoided vendor lock-in while keeping the model fully under our control for future domain adaptation.",Model size and inference costs required heavy optimization to meet latency targets. Documentation and prebuilt tooling were inconsistent across projects.,"More granular, policy-based authorization than the current role and resource-group model.","Performance, control, flexibility, cost, risk & compliance."
 84,Fewer than 100 employees,Technology / Software,4,Both,Cost,3,Yes,"For open-source models: cost efficiency, customization needs, vendor independence. For proprietary AI: specialized capabilities (multimodal processing work).","To fine-tune a customer support ticket classification system that reduced our daily triage time from 3 hours to 20 minutes while saving $180K annually compared to proprietary solutions, achieving 92% accuracy on our domain-specific fintech support data.","Our biggest challenge was the steep learning curve and resource investment required to fine-tune, deploy, and maintain open-source models in production, which initially stretched our small engineering team thin compared to the plug-and-play nature of proprietary APIs.",More granular pricing tiers between the free tier and full enterprise.,"We use a decision matrix weighing cost per inference, required accuracy thresholds, data sensitivity, time-to-deployment, and maintenance overhead."
 85,Fewer than 100 employees,Technology / Software,4,Both,Cost,3,Yes,"Cost-effectiveness, rapid prototyping and experimentation, community support and collaboration.","We built a scalable and accurate sentiment analysis solution, which delivered significant value by automating our analysis, improving customer experience, and enabling data-driven decision-making, all while reducing costs and resources.","Ensuring compliance with regulatory requirements, managing intellectual property and licensing agreements, and addressing potential compatibility and scalability issues with our existing infrastructure and proprietary systems when adopting open-source AI solutions.","Improved licensing models, enhanced security and governance controls, and more comprehensive documentation.","We evaluate the trade-offs between open-source and proprietary options by considering factors such as cost, customization, scalability, security, and community support, as well as our business goals, risk tolerance, and internal expertise, to determine the best fit for each project or initiative."
 86,Fewer than 100 employees,Technology / Software,5,Both,Customization,2,Yes,"We value open-source for flexibility, transparency, and avoiding vendor lock-in. Proprietary models are used when they offer unique capabilities or higher performance in specific contexts.","Orchestrated multiple open-source LLMs in real-time, enabling generalized multi-agent conversation outputs that improved both diversity and robustness of results.","model hosting scalability, aligning inference performance with proprietary systems, and managing model version changes in open-source projects.",More detailed enterprise-grade model cards.,"We assess trade-offs by weighing adaptability, transparency, and cost (open-source) against integration ease, reliability, and specialized capabilities (proprietary). Our abstraction layer minimizes switching costs, enabling a mixed-strategy approach."
-87,"1,000–4,999 employees",Technology / Software,5,Both,"Performance, Flexibility, Community",2,Yes,"Transparency, control, collaboration",Text generation models integrated in production,"Version control, scaling",Enterprise-grade fine-tuning API,Performance > Cost
+87,"1,000-4,999 employees",Technology / Software,5,Both,"Performance, Flexibility, Community",2,Yes,"Transparency, control, collaboration",Text generation models integrated in production,"Version control, scaling",Enterprise-grade fine-tuning API,Performance > Cost
 88,"5,000+ employees",Technology / Software,5,Both,"Performance, Governance, Security",2,Yes,Vendor independence,Custom LLM for internal knowledge base,Policy compliance,Role-based access control,Balance governance vs innovation
-89,500–999 employees,Technology / Software,4,"Open-source (e.g., Hugging Face)","Cost, Customization, Ecosystem",2,Yes,Freedom to innovate,Multi-language classification pipeline,Maintaining dependencies,Long-term maintenance support,Openness before SLA
+89,500-999 employees,Technology / Software,4,"Open-source (e.g., Hugging Face)","Cost, Customization, Ecosystem",2,Yes,Freedom to innovate,Multi-language classification pipeline,Maintaining dependencies,Long-term maintenance support,Openness before SLA
 90,"5,000+ employees",Technology / Software,5,Both,"Performance, Ecosystem, Integration",3,Yes,Open innovation,Multimodal search system,Integration with legacy apps,Simplified API monitoring,Cost-performance equilibrium
 91,"5,000+ employees",Telecommunications,2,Both,Regulatory compliance,3,Yes,"Exploring new markets, we are building AI platform for B2B customers.",Company provides internal LLM models to get help in everyday tasks. Github copilot speeds up developers in implementations and reasoning. ,"As in many big enterprise companies, the organisational and legal overhead slows down project execution. ",Hard to say,"Currently, we're mainly focused on open-source models when selling AI solutions, we use proprietary models whenever they are already integrated in a product we use, e.g. Github Copilot. "
 92,"5,000+ employees",Telecommunications,5,Both,Data control / privacy,4,Yes,"Scalability, flexibility, security, and maintaining compliance while optimizing costs.","We have used Hugging Face to develop multilingual customer support bots, resulting in faster response times and greater customer satisfaction.","Integrating open-source AI with existing legacy systems, ensuring robust data privacy and security, and navigating a complex regulatory environment.","Enhanced technical support, stronger security guarantees, better integration tools, and certification for regulatory compliance.","By weighing the flexibility and cost benefits of open-source against the managed services, reliability, and vendor support of proprietary offerings."
-93,Fewer than 100 employees,Telecommunications,2,Proprietary ,Performance,4,Not sure,Proprietary AI and in-house developed solutions are preferred for control and performance reliability. Open-source AI is considered for pilot projects but with caution regarding integration and support.,Possibly limited but may include customer support automation.,"Integration complexity with specialized telecom hardware and software.
+93,Fewer than 100 employees,Telecommunications,2,Proprietary ,Performance,4,Not sure,Proprietary AI and in-house developed solutions are preferred for control and performance reliability. Open-source AI is considered for pilot projects but with caution regarding integration and support.,Possibly limited but may include customer support automation.,"Integration complexity with specialized telecom hardware and software.
 Data security and compliance considerations.",More specialized signal processing resources. ,"The evaluation prioritizes performance, reliability, and control, favoring proprietary or in-house AI for core functions. Open-source AI is valuable for non-critical, exploratory, or supplementary use cases where rapid innovation is advantageous and resource constraints exist."
-94,"5,000+ employees",Telecommunications,5,Both,Data control / privacy,4,Not sure,Proprietary AI is favored for core operational and network management systems due to reliability and vendor support. Open-source AI is leveraged for rapid experimentation.,"Exploratory and pilot use in NLP for customer service automation, chatbots, and data analytics as part of innovation initiatives at the Innovation Lab.","Integrating with complex telecom infrastructure and legacy systems.
-Navigating regulatory and compliance requirements around data security.
-Building and maintaining AI expertise in-house.
+94,"5,000+ employees",Telecommunications,5,Both,Data control / privacy,4,Not sure,Proprietary AI is favored for core operational and network management systems due to reliability and vendor support. Open-source AI is leveraged for rapid experimentation.,"Exploratory and pilot use in NLP for customer service automation, chatbots, and data analytics as part of innovation initiatives at the Innovation Lab.","Integrating with complex telecom infrastructure and legacy systems.
+Navigating regulatory and compliance requirements around data security.
+Building and maintaining AI expertise in-house.
 Ensuring operational stability and risk mitigation in critical network services.",More specialized documentation and use cases tailored to telecom signal processing and 6G network.,"Flexibility and innovation speed of open-source AI with the reliability, compliance, and vendor support of proprietary solutions. Critical network infrastructure relies on proprietary systems, while open-source AI is embraced in labs and pilot projects to drive future digital transformation and sustainability."
-95,"1,000–4,999 employees",Telecommunications,4,Both,Regulatory compliance,3,Yes,"We handle network telemetry and personal data. We keep retrieval and fine tuning in our virtual private cloud to meet GDPR and data residency, and use a managed API for select copilots with SLAs. This balances auditability, control, and speed.","Retrieval over incident runbooks and knowledge base articles using HF in our VPC, with a vendor LLM for response drafting. We saw higher first contact resolution and shorter time to resolve incidents.","Legal review of model and dataset licences for commercial use, hardening inference endpoints.","Risk-register templates mapped to GDPR and the EU AI Act transparency duties, with links to model cards ","Compliance and privacy first, then security and support readiness. If those are satisfied, we compare task performance and lifecycle cost."
+95,"1,000-4,999 employees",Telecommunications,4,Both,Regulatory compliance,3,Yes,"We handle network telemetry and personal data. We keep retrieval and fine tuning in our virtual private cloud to meet GDPR and data residency, and use a managed API for select copilots with SLAs. This balances auditability, control, and speed.","Retrieval over incident runbooks and knowledge base articles using HF in our VPC, with a vendor LLM for response drafting. We saw higher first contact resolution and shorter time to resolve incidents.","Legal review of model and dataset licences for commercial use, hardening inference endpoints.","Risk-register templates mapped to GDPR and the EU AI Act transparency duties, with links to model cards ","Compliance and privacy first, then security and support readiness. If those are satisfied, we compare task performance and lifecycle cost."
 96,"5,000+ employees",Telecommunications,4,Both,"Cost, Performance",3,Yes,Avoid API cost,Log summarisation,Licensing clarity,Repository rating,Compare ROI
-97,500–999 employees,Telecommunications,3,Proprietary,"Performance, SLAs",3,Not sure,Reliable delivery,Chatbot for tech support,Integration,HF connector kits,SLA before openness
-98,"1,000–4,999 employees",Telecommunications,4,Both,"Customization, Support",2,Yes,Adaptability,Customer ticket triage,Tooling,Workflow templates,Balance support vs autonomy
-99,100–499 employees,Telecommunications,3,Both,"Performance, Cost",3,Yes,Cost-effective pipeline,Internal Q&A bot,Limited resources,Step-by-step MLOps guide,Feasibility → ROI
+97,500-999 employees,Telecommunications,3,Proprietary,"Performance, SLAs",3,Not sure,Reliable delivery,Chatbot for tech support,Integration,HF connector kits,SLA before openness
+98,"1,000-4,999 employees",Telecommunications,4,Both,"Customization, Support",2,Yes,Adaptability,Customer ticket triage,Tooling,Workflow templates,Balance support vs autonomy
+99,100-499 employees,Telecommunications,3,Both,"Performance, Cost",3,Yes,Cost-effective pipeline,Internal Q&A bot,Limited resources,Step-by-step MLOps guide,Feasibility → ROI
 100,"5,000+ employees",Telecommunications,4,Both,"Performance, Compliance",3,Yes,Data control,Intent detection model,Security,Compliance docs,Feasibility → SLA
 ,,,,,,,,,,,,
 ,,,,,,,,,,,,