Corporate Data Confidentiality When Using AI: What a Manager Needs to Know
Practical analysis of confidentiality risks when using cloud AI in enterprises, provider data policies comparison, and architectural decision framework.
Serhiy Andrejev
Lead Machine Learning Developer
Every prompt submitted to a cloud-based AI system is, by definition, a data transfer to a third party. This fundamental fact remains poorly understood in most boardrooms. This article provides a practical framework for managers to assess confidentiality risks, compare provider policies, navigate regulatory requirements, and select the appropriate architectural approach for their enterprise AI deployments.
Introduction
The adoption of AI across industrial enterprises has accelerated dramatically. Managers approve deployments of cloud-based language models for document analysis, report generation, quality control optimization, and operational decision support. Yet in our practice of implementing AI solutions across manufacturing and process industries, we consistently observe a critical blind spot: the majority of decision-makers treat AI tools as if they were installed software running on company servers.
They are not. When an employee pastes a production report into a cloud AI interface, that data leaves the corporate perimeter. When an engineer submits equipment specifications for analysis, those specifications are transmitted to servers operated by a third-party provider — often in a jurisdiction different from the enterprise's own. When a quality manager uploads batch records for summarization, the contents of those records become, at minimum temporarily, available to the AI provider's infrastructure.
The consequences of this blind spot are already visible. In March 2023, Samsung's semiconductor division experienced three separate data leaks within weeks after lifting its internal ban on ChatGPT. Engineers submitted proprietary source code for semiconductor measurement equipment, yield optimization algorithms, and transcripts of internal meetings discussing unreleased fabrication technology [1]. The data was entered voluntarily, by employees who understood neither the data flow nor the retention policies of the tool they were using.
Samsung is not an isolated case. According to Cisco's 2025 Data Privacy Benchmark Study, 48% of organizations acknowledge that employees have entered non-public company information into generative AI tools, while 27% of organizations have at some point banned generative AI use entirely due to privacy and data security concerns [2]. The gap between the rate of adoption and the level of organizational awareness regarding data handling represents one of the most significant governance challenges in enterprise technology today.
For managers, the question is not whether to use AI — competitive pressure makes that increasingly untenable — but how to use it without creating unacceptable confidentiality exposure. Answering this question requires understanding what data is at risk, how providers handle that data, what regulatory obligations apply, and which architectural patterns match which risk profiles.
The Risk Map: What Data Leaves Your Organization
The first step toward effective data governance in AI deployments is understanding exactly what information is transmitted when employees interact with cloud-based AI systems. In our experience, most organizations significantly underestimate the scope and sensitivity of the data flows involved.
Direct prompt content. The most obvious category: the text, data, or documents that users explicitly submit to the AI system. This includes production reports, financial summaries, technical specifications, customer correspondence, employee records, and any other content pasted or uploaded into the interface.
Contextual metadata. Many enterprise AI deployments use retrieval-augmented generation (RAG) or similar techniques that automatically attach context to user queries. When a manager asks "What were last quarter's defect rates?", the system may retrieve and transmit entire sections of quality management databases, including data the user did not explicitly intend to share.
Conversation history. Multi-turn interactions accumulate context. A series of individually innocuous questions about production scheduling, equipment capacity, and supplier delivery timelines may, in aggregate, reveal proprietary operational strategy.
Embedded sensitive data. Documents submitted for summarization or analysis often contain information that the submitting user does not consciously register as sensitive: IP addresses in log files, employee names and identifiers in production records, customer-specific pricing in contract documents, geographic coordinates of facilities in maintenance reports.
Based on our observations across multiple industrial deployments, we categorize enterprise data into four sensitivity tiers, each demanding different handling protocols:
| Data Tier | Definition | Examples | Risk Level |
|---|---|---|---|
| Public | Information intended for external audiences | Marketing materials, published specifications, public financial reports | Low |
| Internal | Non-sensitive operational information | General process documentation, non-proprietary meeting notes, industry benchmarks | Moderate |
| Confidential | Business-critical information with competitive value | Production formulas, pricing models, supplier agreements, unreleased product roadmaps | High |
| Restricted | Regulated or high-impact data requiring statutory protection | Personal employee data, patient health records, classified defense information, critical infrastructure parameters | Critical |
Source: Classification framework based on ISO 27001 information security standard and adapted for AI deployment contexts.
The critical insight for managers is that data classification must occur before AI tools are deployed, not after an incident. An organization without a clear data classification policy is, in effect, allowing every employee to make their own judgment about what is safe to transmit — as the Samsung incident demonstrated.
Provider Data Policies: What Happens to Your Data
Understanding how major AI providers handle enterprise data is essential for informed decision-making. The landscape has matured considerably since the early days of consumer-grade ChatGPT, but significant differences remain between providers and between service tiers within the same provider.
Comparison of Major LLM Provider Data Policies
| Policy Dimension | OpenAI (API / Enterprise) | Anthropic (API / Enterprise) | Google Cloud (Gemini Enterprise) |
|---|---|---|---|
| Training on customer data | No (API and Enterprise by default) [3] | No (commercial services) [4] | No (paid Enterprise tier) [5] |
| Data retention — API | 30 days (abuse monitoring), then deleted [3] | 30 days, then deleted; zero-retention option available [4] | Not used for training; data residency controls available [5] |
| Data retention — Enterprise | Customer-controlled retention period [3] | Customer-controlled; 30-day backend deletion after user deletes [4] | Customer-controlled; data stays in chosen region [5] |
| Encryption at rest | AES-256 [3] | AES-256 [4] | AES-256 (Google-managed or customer-managed keys) [5] |
| Encryption in transit | TLS 1.2+ [3] | TLS 1.2+ [4] | TLS 1.2+ [5] |
| Zero data retention option | Available (Enterprise) [3] | Available (API and Enterprise) [4] | Available (Enterprise) [5] |
| SOC 2 Type 2 | Yes [3] | Yes [4] | Yes [5] |
| GDPR compliance | Supported [3] | Supported (DPA available) [4] | Supported [5] |
| HIPAA compliance | Supported (Enterprise) [3] | Supported (BAA available) [4] | Supported [5] |
| Data Processing Agreement | Available [3] | Available [4] | Available [5] |
| Customer-managed encryption keys | Yes (Enterprise Key Management) [3] | On request [4] | Yes (CMEK) [5] |
Sources: OpenAI Enterprise Privacy Policy [3], Anthropic Commercial Terms [4], Google Cloud Gemini Data Governance documentation [5]. Policies current as of April 2026.
Several observations merit attention. First, the distinction between consumer and enterprise tiers is critical. Consumer versions of ChatGPT, Claude, and Gemini operate under significantly different data handling terms — in some cases, user inputs may be used for model improvement unless the user explicitly opts out. Enterprise and API tiers universally commit to not training on customer data.
Second, data retention periods matter even when training exclusions are in place. During the retention window — typically 30 days for abuse monitoring — customer data exists on the provider's infrastructure. For organizations handling restricted data, even temporary third-party storage may violate regulatory or contractual obligations.
Third, the availability of zero data retention options represents a meaningful differentiator for high-sensitivity use cases. However, zero retention typically comes with trade-offs: reduced ability to investigate service issues, no conversation history, and in some cases, higher pricing.
In our practice, we observe that most organizations do not read provider data policies before deploying AI tools. This is the managerial equivalent of signing a contract without reading it. We consider it essential that any enterprise AI deployment begins with a formal review of the provider's data processing terms by both legal counsel and information security teams.
Three Architectural Approaches to Enterprise AI
The choice of deployment architecture is the most consequential decision a manager makes regarding AI data confidentiality. Each approach represents a different trade-off between capability, cost, and control. Based on our implementation experience across manufacturing enterprises, we identify three fundamental patterns.
Cloud-First Architecture
In this model, all AI processing occurs on the provider's infrastructure. The enterprise sends data to the cloud API, receives results, and relies on the provider's contractual and technical commitments for data protection.
Advantages: Lowest upfront investment, access to the most capable models, fastest time to deployment, automatic updates, no specialized infrastructure required.
Risks: All data transits through and temporarily resides on third-party infrastructure. The enterprise's data protection posture is only as strong as the provider's operational compliance. Data residency or sovereignty requirements may be difficult to satisfy.
Appropriate for: Public and internal-tier data. Organizations in industries without stringent data residency requirements. Use cases where the competitive value of AI capabilities outweighs the residual confidentiality risk.
On-Premise Architecture
Here, the enterprise deploys and operates its own AI models on hardware within its own data center or private cloud. No data leaves the corporate perimeter.
Advantages: Complete data sovereignty, full compliance with the most restrictive data residency requirements, elimination of third-party data handling risk, ability to process restricted-tier data.
Risks: Significantly higher capital expenditure (GPU infrastructure), operational complexity (model management, updates, security patching), reduced model capability (the most powerful models are unavailable for on-premise deployment), need for specialized talent.
Appropriate for: Restricted and confidential-tier data. Defense, healthcare, financial services, critical infrastructure. Organizations with existing data center capabilities and technical teams.
Hybrid Architecture
The hybrid approach routes data through different processing paths based on its sensitivity classification. Non-sensitive queries go to cloud APIs for maximum capability; sensitive data is processed locally on smaller but adequate models.
Advantages: Optimizes the trade-off between capability and confidentiality. Allows organizations to benefit from cloud models for non-sensitive tasks while maintaining data sovereignty for sensitive operations. Most cost-effective approach for mixed data sensitivity profiles.
Risks: Requires robust data classification and routing mechanisms. Introduces architectural complexity. Demands ongoing governance to ensure sensitive data is not inadvertently routed to cloud endpoints. Requires investment in both cloud and on-premise infrastructure.
Appropriate for: Most industrial enterprises. Organizations with diverse data sensitivity profiles. Companies seeking to maximize AI value while maintaining regulatory compliance.
In our assessment, the hybrid architecture represents the most pragmatic approach for the majority of manufacturing enterprises. It acknowledges that not all data is equally sensitive and that not all AI tasks require the most powerful models. A quality documentation chatbot can operate effectively on a cloud API; a system analyzing proprietary production formulas should process data locally. The key requirement is a rigorous data classification framework and an automated routing layer that enforces classification decisions consistently.
Decision Matrix: Data Sensitivity vs. Architectural Approach
The following matrix provides a practical framework for matching data types to deployment architectures. Each cell indicates the suitability of the architectural approach for the given data sensitivity tier.
| Data Sensitivity | Cloud API | Hybrid | On-Premise |
|---|---|---|---|
| Public | ✅ Acceptable | ✅ Acceptable | ✅ Acceptable (but over-engineered) |
| Internal | ✅ Acceptable (with enterprise tier) | ✅ Acceptable | ✅ Acceptable |
| Confidential | ⚠️ With limitations (enterprise tier + DPA + legal review) | ✅ Recommended | ✅ Acceptable |
| Restricted | ❌ Unacceptable | ⚠️ With limitations (sensitive data stays on-premise) | ✅ Required |
Source: Framework developed based on ISO 27001, NIST AI RMF [7], and adapted from enterprise deployment experience across manufacturing and process industries.
Figure 1. Heatmap of architecture suitability by data sensitivity tier. Higher scores (brighter) indicate greater suitability. Source: ISO 27001, NIST AI RMF [7], adapted from deployment experience.
This matrix is not prescriptive but directional. Specific regulatory requirements, industry standards, and contractual obligations may impose additional constraints. The critical principle is that data classification drives architecture selection — not the other way around.
Regulatory Landscape: GDPR, EU AI Act, and Beyond
Managers deploying AI in enterprises operating within the European Union, Ukraine, or serving EU customers must navigate an increasingly dense regulatory environment. Three regulatory frameworks are particularly relevant.
General Data Protection Regulation (GDPR)
The GDPR, in force since May 2018, establishes fundamental principles that directly affect enterprise AI deployments [6]. When a company uses a cloud AI provider to process data that includes personal information — employee names in production reports, customer contact details in correspondence, personnel data in HR documents — the AI provider becomes a data processor under Article 28.
This triggers specific legal obligations: a formal Data Processing Agreement (DPA) must be in place, the processor must act only on documented instructions, appropriate security measures must be implemented, and the data controller (the enterprise) retains full responsibility for the lawfulness of processing. Critically, GDPR's data minimization principle (Article 5) requires that only data necessary for the specific purpose be processed — a principle that is inherently challenged by AI systems designed to ingest broad context.
For managers, the practical implication is clear: deploying a cloud AI tool without a DPA in place is not merely a policy failure but a potential regulatory violation carrying fines of up to 4% of global annual revenue.
EU AI Act
Regulation (EU) 2024/1689, commonly known as the EU AI Act, introduces the world's first comprehensive legal framework for artificial intelligence [8]. While its primary focus is on AI system classification and risk management rather than data privacy specifically, several provisions directly affect enterprise data governance.
Article 10 establishes data governance requirements for high-risk AI systems, including obligations regarding training data quality, bias testing, and data documentation. For enterprises using AI to support decisions in areas such as employment, safety-critical systems, or access to essential services, these requirements overlay on top of GDPR obligations.
The EU AI Act's phased enforcement timeline means that different provisions take effect between 2025 and 2027. Managers should treat this period as a compliance preparation window rather than a grace period.
Ukrainian Data Protection Framework
For enterprises operating in Ukraine, the Law of Ukraine "On Personal Data Protection" (No. 2297-VI) establishes requirements aligned with but not identical to GDPR. Recent legislative initiatives aimed at harmonizing Ukrainian data protection law with EU standards are creating an increasingly convergent regulatory environment, particularly relevant for Ukrainian companies serving EU markets or operating as part of EU supply chains.
Practical Regulatory Implications
The convergence of GDPR, the EU AI Act, and national legislation creates a layered compliance obligation. For a manufacturing enterprise using AI to analyze production data that incidentally contains employee identifiers:
- GDPR requires a legal basis for processing personal data, a DPA with the AI provider, and appropriate security measures.
- The EU AI Act may impose additional requirements if the AI system is classified as high-risk.
- National legislation adds jurisdiction-specific notification or registration obligations.
In our experience, the most effective approach is to treat the strictest applicable standard as the baseline and build governance frameworks that satisfy all layers simultaneously, rather than attempting piecemeal compliance with each regulation independently.
Selection Criteria: A Decision Framework for Managers
Based on our experience implementing AI solutions across diverse industrial environments, we propose the following decision framework for managers evaluating AI deployment approaches from a data confidentiality perspective.
Step 1: Classify your data. Before evaluating any AI tool or provider, establish a data classification policy that categorizes all information assets into sensitivity tiers. This is a prerequisite, not an optional step. Without classification, every subsequent decision is made in the dark.
Step 2: Map regulatory obligations. Identify which regulations apply to your organization, your data, and your intended AI use cases. GDPR applies if you process EU residents' personal data. The EU AI Act applies if your AI system falls within its scope. Industry-specific regulations (HIPAA, PCI-DSS, defense classification standards) may impose additional requirements.
Step 3: Evaluate provider policies. For any cloud or hybrid deployment, conduct a formal review of the provider's data handling terms. Key questions: Does the provider train on your data? What is the retention period? Is zero data retention available? Where is data processed and stored? What compliance certifications does the provider hold? Is a DPA available?
Step 4: Match architecture to risk. Use the decision matrix above to identify which architectural approaches are appropriate for your data sensitivity profile. If your organization processes only public and internal data, cloud-first may be sufficient. If you handle confidential or restricted data — as most manufacturing enterprises do — hybrid or on-premise architectures are necessary.
Step 5: Implement governance. Technology alone does not solve the confidentiality problem. The Samsung incident was not a technology failure — it was a governance failure. Effective governance includes: clear policies on what data may and may not be submitted to AI systems, employee training programs, technical controls (DLP systems, API gateways with content filtering), regular audits, and incident response procedures.
Step 6: Monitor and reassess. Provider policies change. Regulations evolve. New AI capabilities introduce new data flows. The confidentiality framework must be a living governance structure, not a one-time compliance exercise.
Gartner's research reinforces the urgency of this approach: 25% of enterprise generative AI applications are predicted to experience at least five minor security incidents per year by 2028, up from 9% in 2025 [9]. Organizations that invest in governance infrastructure now will be materially better positioned to manage these incidents when they occur.
Author's Assessment and Conclusions
The current state of enterprise AI adoption reveals a persistent and dangerous gap between the speed of technology deployment and the maturity of data governance frameworks. Managers are approving AI tools with the same procurement mindset they apply to office software — a mindset that is fundamentally inadequate for a technology that, by its nature, requires access to organizational knowledge to deliver value.
Based on our practice, we offer the following conclusions:
First, awareness is the most urgent requirement. The majority of data confidentiality incidents in AI deployments stem not from sophisticated attacks or provider failures, but from employees who do not understand where their data goes when they type it into an AI interface. Executive-level understanding of AI data flows is a prerequisite for effective governance.
Second, the hybrid architecture is the pragmatic choice for most industrial enterprises. Pure cloud deployments expose confidential data to unacceptable risk. Pure on-premise deployments sacrifice capability and impose prohibitive costs for most organizations. The hybrid approach — cloud for non-sensitive tasks, local processing for sensitive data — optimizes the trade-off, provided it is supported by robust data classification and routing mechanisms.
Third, provider enterprise tiers are necessary but not sufficient. The commitments made by OpenAI, Anthropic, and Google in their enterprise terms are meaningful and represent genuine progress. However, they do not eliminate all risk. Data still transits through and temporarily resides on third-party infrastructure. Contractual protections are only as strong as the provider's operational compliance. Enterprises must maintain their own security controls and audit capabilities.
Fourth, regulatory compliance is a floor, not a ceiling. GDPR and the EU AI Act establish minimum standards. Organizations that treat compliance as the endpoint rather than the starting point will find themselves perpetually reactive. The goal should be a data governance framework that exceeds regulatory requirements and provides genuine organizational confidence in the security of AI deployments.
Fifth, governance is an ongoing operational function, not a project. The AI landscape changes rapidly. Provider policies are updated, new models introduce new data handling patterns, regulatory interpretations evolve. Data confidentiality governance must be embedded in organizational operations as a continuous process.
The enterprises that will derive the greatest sustainable value from AI are not those that adopt fastest, but those that adopt most thoughtfully — with clear-eyed understanding of the data flows, risks, and governance requirements that responsible AI deployment demands.
References
- The Register. "Samsung reportedly leaked its own secrets through ChatGPT." 2023. https://www.theregister.com/2023/04/06/samsung_reportedly_leaked_its_own
- Cisco. "2025 Data Privacy Benchmark Study." 2025. https://newsroom.cisco.com/content/r/newsroom/en/us/a/y2025/m04/cisco-2025-data-privacy-benchmark-study-privacy-landscape-grows-increasingly-complex-in-the-age-of-ai.html
- OpenAI. "Enterprise Privacy at OpenAI." 2026. https://openai.com/policies/api-data-usage-policies
- Anthropic. "Commercial Terms of Service." 2025. https://www.anthropic.com/legal/commercial-terms
- Google Cloud. "How Gemini for Google Cloud Uses Your Data." 2025. https://docs.cloud.google.com/gemini/docs/discover/data-governance
- European Parliament and Council. "Regulation (EU) 2016/679 — General Data Protection Regulation." 2016. https://gdpr.eu.org/art/28/
- NIST. "AI Risk Management Framework (AI RMF 1.0)." 2023. https://www.nist.gov/itl/AI-risk-management-framework
- European Parliament and Council. "Regulation (EU) 2024/1689 — Artificial Intelligence Act." 2024. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689
- Gartner. "Gartner Predicts 25% of All Enterprise GenAI Applications Will Experience At Least Five Minor Security Incidents Per Year By 2028." 2026. https://www.gartner.com/en/newsroom/press-releases/2026-04-09-gartner-predicts-25-percent-of-all-enterprise-gen-ai-applications-will-experience-at-least-five-minor-security-incidents-per-year-by-2028