Negotiating Data Privacy Agreements

I am curious… Last week, our team spent time in a workshop with a potential partner. During the session, someone used ChatGPT to summarise our ideas in real time and make suggestions for improvement. It sparked great discussion, but it also raised an awkward question. For the AI to generate meaningful suggestions, it needed the context we had just shared, including technical details, strategic direction, and confidential roadmap items. Later, in a conversation with a legal advisor, we realised our NDAs did not explicitly cover this. They were written for a time when “sharing” meant emailing a document or handing over a printout, not pasting confidential information into a model you do not control. We ended up updating our docs to include an AI-specific clause: No Confidential Information may be uploaded to, processed by, or disclosed to any publicly available AI/ML system, model, or dataset without prior written consent. Apparently, this is starting to appear in some contracts as legal teams and AI law specialists are recommending clauses that: - Ban feeding confidential data into public models without written consent. - Require proof that approved tools will not train on the data. - Bind contractors and sub processors to the same rules. Some even provide model language allowing AI use only with “commercially reasonable assurances” the model will not train on the information and is isolated from other customers. Has anyone else encountered this or started updating their own NDAs and agreements? #AIGovernance #DataPrivacy #LegalTech #AICompliance #Contracts

🚨 AI Privacy Risks & Mitigations Large Language Models (LLMs), by Isabel Barberá, is the 107-page report about AI & Privacy you were waiting for! [Bookmark & share below]. Topics covered: - Background "This section introduces Large Language Models, how they work, and their common applications. It also discusses performance evaluation measures, helping readers understand the foundational aspects of LLM systems." - Data Flow and Associated Privacy Risks in LLM Systems "Here, we explore how privacy risks emerge across different LLM service models, emphasizing the importance of understanding data flows throughout the AI lifecycle. This section also identifies risks and mitigations and examines roles and responsibilities under the AI Act and the GDPR." - Data Protection and Privacy Risk Assessment: Risk Identification "This section outlines criteria for identifying risks and provides examples of privacy risks specific to LLM systems. Developers and users can use this section as a starting point for identifying risks in their own systems." - Data Protection and Privacy Risk Assessment: Risk Estimation & Evaluation "Guidance on how to analyse, classify and assess privacy risks is provided here, with criteria for evaluating both the probability and severity of risks. This section explains how to derive a final risk evaluation to prioritize mitigation efforts effectively." - Data Protection and Privacy Risk Control "This section details risk treatment strategies, offering practical mitigation measures for common privacy risks in LLM systems. It also discusses residual risk acceptance and the iterative nature of risk management in AI systems." - Residual Risk Evaluation "Evaluating residual risks after mitigation is essential to ensure risks fall within acceptable thresholds and do not require further action. This section outlines how residual risks are evaluated to determine whether additional mitigation is needed or if the model or LLM system is ready for deployment." - Review & Monitor "This section covers the importance of reviewing risk management activities and maintaining a risk register. It also highlights the importance of continuous monitoring to detect emerging risks, assess real-world impact, and refine mitigation strategies." - Examples of LLM Systems’ Risk Assessments "Three detailed use cases are provided to demonstrate the application of the risk management framework in real-world scenarios. These examples illustrate how risks can be identified, assessed, and mitigated across various contexts." - Reference to Tools, Methodologies, Benchmarks, and Guidance "The final section compiles tools, evaluation metrics, benchmarks, methodologies, and standards to support developers and users in managing risks and evaluating the performance of LLM systems." 👉 Download it below. 👉 NEVER MISS my AI governance updates: join my newsletter's 58,500+ subscribers (below). #AI #AIGovernance #Privacy #DataProtection #AIRegulation #EDPB

Isabel Barberá: "This document provides practical guidance and tools for developers and users of Large Language Model (LLM) based systems to manage privacy risks associated with these technologies. The risk management methodology outlined in this document is designed to help developers and users systematically identify, assess, and mitigate privacy and data protection risks, supporting the responsible development and deployment of LLM systems. This guidance also supports the requirements of the GDPR Article 25 Data protection by design and by default and Article 32 Security of processing by offering technical and organizational measures to help ensure an appropriate level of security and data protection. However, the guidance is not intended to replace a Data Protection Impact Assessment (DPIA) as required under Article 35 of the GDPR. Instead, it complements the DPIA process by addressing privacy risks specific to LLM systems, thereby enhancing the robustness of such assessments. Guidance for Readers > For Developers: Use this guidance to integrate privacy risk management into the development lifecycle and deployment of your LLM based systems, from understanding data flows to how to implement risk identification and mitigation measures. > For Users: Refer to this document to evaluate the privacy risks associated with LLM systems you plan to deploy and use, helping you adopt responsible practices and protect individuals’ privacy. " >For Decision-makers: The structured methodology and use case examples will help you assess the compliance of LLM systems and make informed risk-based decision" European Data Protection Board

🤞🏽 How To Share NDA-Protected UX Work. In B2B and Enterprise work, typically you can’t speak about a project you’ve spent years on — not the client, not the findings, not a single screenshot. But how do we show it in a portfolio or use it as a case study in a job interview? It depends on the specific NDA, but a signed NDA doesn’t necessarily mean that you aren’t allowed to share anything about the project: 🚫 Things you (often) can't share 1. Client name or client's branding 2. Client employees you worked with 3. Specific findings or research 4. Actual artifacts and deliverables 5. Screenshots of work pre/mid/after 6. Any business or management decisions 7. Customer details, sensitive data 8. Mentions of products/services used 9. Any metrics, data or analytics 10. Vendor or partner names 11. Photos from client premises 12. Unreleased feature or roadmap details ✅ Things you (often) can share 1. The industry you worked in 2. Your team's role and responsibility 3. How you planned and executed work 4. The research/design process you followed 5. Recommendation from your manager/client 6. Redacted, recreated artifacts 7. Swapped data with open source datasets 8. Publicly released materials (press release, YouTube etc.) 9. Material from a public talk deck (with cited source) 10. Data reported by third-party services (with credit) In fact, that may include text, screenshots, metrics, quotes from team members or executives — as long as it's already public. Also some NDAs are time-limited, so check how long restrictions actually apply. Jessica Ivins wrote a useful post on just that a while back (https://lnkd.in/dQDHj__6). She points out that typically you can’t disclose any findings, but you can walk through your methods, your decision-making and how it helped uncover valuable insights or make progress. As Nick Finck wrote, no need to invent numbers — disclose upfront that the case study is under strict NDA, which is why you can’t share any specifics. But you can explain your role and your process, and be upfront about what you can’t share, and why. I always struggle with disclosing patterns that emerge from usability testing. Of course, we can't generalize — these findings are often specific to the context we observed them in. But once we introduce a change and notice its impact, that's worth remembering. And because these insights are often perceived as merely technical, I generalize them and ask permission to share — no names, no numbers, no specifics. What I've found: UX findings rarely feel important to companies, as long as their brand isn't attached and nothing business-critical is revealed through them. It's really a question of how, when and who you ask — and understanding what kind of details is actually OK to make public. In the end, always check the actual wording of your NDA — don't assume. And when in doubt, ask your manager for permission directly. I hope it helps! 🤞🏽

GSA just changed the game for every federal contractor handling CUI. New mandatory cybersecurity requirements dropped January 5. Not guidance. Requirements. This is not CMMC. This is GSA building its own CUI protection framework, and it goes further than DoD in several areas. Translation: - NIST 800-171 Revision 3 required (DoD still allows Rev 2) - No self-assessment option. Third-party assessment mandatory. - One-hour incident reporting. Not 72 hours. One hour. - Nine "showstopper" controls including MFA, encryption, and vulnerability monitoring - Five-phase compliance lifecycle: Prepare, Document, Assess, Authorize, Monitor This is the first major expansion of CUI protections beyond the Department of Defense. If you have a GSA contract and you handle CUI, your compliance posture just changed overnight. The organizations that prepared for CMMC are ahead. Everyone else is scrambling. #GRCEngineering #CMMC

The most dangerous clauses in vendor contracts aren’t the ones you fight over. They’re the ones you skim past—(em dash mine 😑) the “standard” terms that seem harmless until they explode. Just ask Morgan Stanley. Overlooked contractual gaps turned a vendor’s mishandling of client-data-bearing equipment into hundreds of millions in fines, settlements, and penalties for Morgan Stanley. I have identified some top of mind examples: #1: The Subcontracting Black Hole Most vendor contracts include innocent-looking language like: "Vendor may engage subcontractors as necessary to perform services." The problem: You have zero visibility into who's actually handling your sensitive data or critical operations. What Morgan Stanley missed: Their vendor subcontracted the actual data destruction to an unqualified third party. The fix: • Require prior written approval for all subcontractors • Mandate the same security/compliance standards flow down • Include right to audit subcontractors directly • Cap subcontracting to specific, pre-approved functions #2: The Liability Cap Loophole Standard cap: "Vendor's liability limited to fees paid in preceding 12 months." The hidden trap: This covers the vendor's mistakes but not the regulatory fines, customer lawsuits, and reputational damage you'll face. What to negotiate: • Separate caps for different types of damages • Higher caps for data breaches and regulatory violations • Unlimited liability for gross negligence and willful misconduct • Minimum insurance requirements that match your actual risk exposure #3: The Termination Cost Surprise Innocent clause: "Upon termination, vendor will assist with transition for 30 days." The trap: No mention of data extraction, migration costs, or knowledge transfer requirements. Real example: A SaaS company switching CRM vendors discovered "transition assistance" meant read-only access to export screens. Manual data extraction cost $47K in consulting fees. Protection strategies: • Define data export formats and timelines • Cap termination assistance fees • Require knowledge transfer documentation • Include escrow provisions for critical operational data #4: The Change Order Cash Grab Standard language: "Any modifications require mutual written agreement." The hidden cost: No controls on pricing for change orders or scope creep. Pattern I see: Vendors lowball initial proposals then recover margins through change orders priced at 200-400% markup. The armor: • Cap change order pricing as percentage of original contract value • Require detailed justification for scope changes above set thresholds • Include right to third-party validation for major change orders • Build in quarterly spend reviews with automatic triggers The point is, most "standard" vendor contracts are written to protect vendors, not you. Don't let your "standard" vendor agreement become someone else's cautionary tale. Dig deep. #VendorManagement #ContractReview #RiskManagement

DPDP Act Decoded #27: Using Data Processors Lawfully — What a “Valid Contract” Must Cover Many organisations assume DPDP compliance can be handled through a standard vendor DPA. That is not what the law says. The Act permits engagement of a Data Processor, for any activity related to the offering of goods or services to Data Principals, only under a valid contract. But it does not define that phrase or prescribe a clause-level checklist. So the real question is not “What template do we use?” It is: what must the contract enable the Data Fiduciary to do, given its non-delegable obligations? Three things matter in practice. 1. Responsibility does not shift Section 8(1) is explicit. The Data Fiduciary remains responsible for complying with the Act and Rules in respect of processing carried out by it or on its behalf by a Data Processor. A processor agreement is not just a procurement document. It is part of the Data Fiduciary’s compliance architecture. 2. The contract must enable control where the Act requires it Processing under DPDP must rest on a lawful purpose, whether through consent or certain legitimate uses. While the Act does not expressly prescribe “instruction clauses,” the strongest reading is that the contract should bind the processor to purpose-linked processing requirements and enable the Data Fiduciary to exercise control where the Act requires it. This matters because: • if consent is withdrawn, the Data Fiduciary must cause its processors to cease processing • where erasure is required, the Data Fiduciary must cause its Data Processor to erase the personal data made available to it That outcome is only possible if the contract creates enforceable control. 3. Security is now expressly a contractual issue Rule 6(1)(f) makes one point explicit: the contract should contain appropriate provision for taking reasonable security safeguards. Read with Section 8(5), this means security cannot be left as an implied expectation. This is not implied compliance. It is contractual design. In practice, the contract should support access controls, logging and monitoring, breach response support, and retention and erasure workflows. It should also be drafted with Rule 8(3) in view, because the Rules require retention of certain data and logs for at least one year before erasure in specified cases. So what is a “valid contract” under DPDP? The Act does not define it. But the safest reading is this: a processor contract should be structured so that outsourcing does not disable the Data Fiduciary from complying with its own statutory duties. The real takeaway Under DPDP, you can outsource processing. You cannot outsource accountability. Relevant provisions • Section 2(i), 2(k) • Section 4(1), 4(2) • Section 6(4)–6(6) • Section 8(1), 8(2), 8(4), 8(5), 8(6), 8(7) • Rule 6(1), especially clause (f) • Rule 7 • Rule 8(3) #DPDP #DPDPAct #DataProtection #DataPrivacy #PrivacyLaw #IndiaLaw #Compliance #DataGovernance

Forget employees experimenting with ChatGPT on their lunch breaks. That is yesterday's shadow AI problem. The more systemic risk facing HR and People teams right now is institutional shadow AI, and it is built directly into your standard software supply chain. For years, procurement teams have treated the standard Data Processing Agreement (DPA) as a definitive legal shield. If a vendor signed on the dotted line, the compliance box was checked. DataGrail’s 2026 Privacy and AI Trends Report (attached) suggests this compliance model is fractured. Out of 2,400 business software providers surveyed, 63.6% of vendors marketing AI capabilities failed to disclose their third-party AI subprocessors in their legal documentation. This means the talent acquisition platform or resume-screening tool you recently onboarded is likely routing sensitive applicant data into backend foundational models that your data protection team has never reviewed, vetted, or approved. From a Governance, Risk and Compliance perspective, this creates three distinct friction points: 1️⃣ First, it creates an auditability blind spot. Roughly 20.7% of these undisclosed systems power Automated Decision-Making (ADM). If a backend model quietly filters out a candidate pool, you cannot audit it for systemic bias. If a regulator investigates your hiring practices, pointing to an incomplete vendor DPA will not shift the liability away from your organization. 2️⃣ Second, the regulatory landscape is shifting toward personal accountability. Under updated compliance frameworks (like the CCPA’s risk assessment rule) executives are increasingly required to personally sign off on AI data safety. If your HR platform is quietly exposing PII to unvetted models, that structural risk rests on executive shoulders. 3️⃣ Finally, there is an operational logjam. The "Jobs & Professional Development" sector currently faces the highest volume of data deletion requests nationwide (US only), averaging 365 per month. Trying to manually track down and purge candidate records across an unmapped web of hidden software pipelines is driving up compliance costs, with mid-market companies spending an estimated $1.5 million annually just to manage the operational mess. The assumption that your HR tech vendors are inherently protecting your data is no longer a viable risk strategy. If you are navigating AI integration right now, the play is to shift from passive contract reviews to active technical verification. Do not stop reviewing DPAs, but ensure they require a technically verifiable map of every backend model and subprocessor your vendors actually use. The organizations that navigate the next wave of regulatory enforcement successfully won't be those with the thickest policy binders. They will be the ones that ensure their legal contracts match their engineering realities.

Healthcare has made tremendous progress in exchanging data. The next challenge is ensuring that information is exchanged responsibly, consistently, and in accordance with each individual’s choices. This article and resource highlight important new guidance from The Sequoia Project’s Privacy & Consent Workgroup to help providers, payers, health information networks, and technology developers transition from fragmented, manual consent processes toward automated, computable consent. The guidance is especially valuable because it recognizes that consent is not simply a technical transaction. It requires coordinated legal interpretation, governance, organizational readiness, interoperable standards, and workflows that can reliably capture, communicate, manage, and enforce patient preferences. Through our collaboration with The Sequoia Project and Shift Collaborative, HL7 FHIR at Scale Taskforce (FAST) is helping advance the complementary technical infrastructure needed to operationalize this vision. The FAST Scalable Consent Management Implementation Guide is focused on providing reusable FHIR-based patterns for managing the consent lifecycle—including capture, decisions, delegation, revocation, discovery, status inquiries, and auditability. Together, this work can help the industry move from asking whether consent can be automated to demonstrating how it can be implemented consistently across organizations, networks, jurisdictions, and healthcare use cases. This is how we build a scalable consent trust stack—one that supports appropriate information sharing while strengthening privacy, transparency, patient choice, and trust. https://lnkd.in/gxU_dv-D #HealthcareInteroperability #ComputableConsent #PatientConsent #FHIR #HL7FAST #TheSequoiaProject #HealthDataExchange #PatientEmpowerment #HealthIT

Among other things, President Biden's May 12, 2021 Executive Order (EO) on Cybersecurity seeks to improve the security of software used by the federal government. See https://lnkd.in/e399Y2g. The EO contemplates "contract language requiring suppliers of software available for purchase by agencies to comply with, and attest to complying with" certain security requirements, including "conformity with secure software development practices" and ensuring "the integrity and provenance of open source software used within any portion of a product." On March 11, CISO and OMB released a Secure Software Development Attestation Form (the Form) for software producers to comply with this new requirement. See https://lnkd.in/ea3aSSvQ. Yesterday, CISA went "live" with its "Repository for Software Attestation and Artifacts." See https://lnkd.in/erXttezB. Among other things, the Form requires "the Chief Executive Officer (CEO) of the software producer or their designee, who must be an employee of the software producer and have the authority to bind the corporation" to attest to the existence of specific controls summarized below: (1) The software is developed and built in secure environments (including with environment segregation; regular logging, monitoring, and auditing trust relationships used for authorization and access; MFA and conditional access; continuous monitoring of operations and alerts; etc.); (2) The software producer makes a "good faith" effort to maintain trusted source code supply chains by employing automated tools or comparable processes to address the security of internal code and third-party components and manage related vulnerabilities; (3) The software producer maintains provenance for internal code and third-party components incorporated into the software to the greatest extent feasible; (4) The software producer employed automated tools or comparable processes that check for security vulnerabilities. The form also requires the software producer to attest that "it will notify any agency to which it has submitted this form if and when the producer ceases to make consistent use of the practices identified above in developing the software." Because the attestor is generally not required to provide any evidence or detail to support this attestation, it may be tempting for some software producers to "hold their breath" and just attest. Please do NOT do this. The form makes clear that "[w]illfully providing false or misleading information may constitute a violation of 18 U.S.C. § 1001, a criminal statute." The DOJ prosecutes Section 1001 violations so frequently that there are "scores of reported cases" on such prosecutions. See https://lnkd.in/ekXK_Bte. If you find yourself required to fill out this attestation, but are not yet in a position to do so truthfully, please discuss your options with competent counsel. This attestation is definitively not worth committing a federal crime over.