The deal is at procurement. Security review passed, pricing agreed. Then the buyer’s legal team sends a 40-question “AI vendor assessment.” Your AE forwards it with one line: “can we say yes to everything?” No, and how you answer determines whether you close in two weeks or stall for a quarter. Since August 2026, every enterprise buyer with EU exposure runs some version of this questionnaire. Treat it as a product surface, not paperwork.
The twelve questions behind every questionnaire
Strip the branding and enterprise AI assessments converge on twelve asks. (1) AI Act role and risk classification of the feature you sell. (2) Which foundation models sit underneath, under what licence or API terms. (3) Whether customer data trains any model; yours or your vendor’s. (4) Where inference happens and the KVKK/GDPR transfer mechanics. (5) Retention of prompts and outputs. (6) Article 50 disclosures and content marking. (7) Human-oversight points for consequential decisions. (8) Hallucination and accuracy controls with metrics. (9) Incident response for AI failures, tied into your breach process. (10) Subprocessor list for the AI chain. (11) IP posture: who owns outputs, and indemnities for infringement claims. (12) Evidence (policies, logs, test reports) for all of the above.
Build the answer once: the AI factsheet
The scaling move is a two-page, versioned AI factsheet mirroring those twelve headings, maintained like a SOC 2 report and attached to every questionnaire response. Rules of the road: never claim “we are fully AI Act compliant” (a meaningless and falsifiable sentence, and state your classification and controls instead). Never deny model use that your architecture diagram reveals. Align the factsheet with your public marketing, because a mismatch is an AI-washing exhibit. Where the honest answer is “not yet,” write “roadmapped for Q4.” Buyers accept gaps with dates far better than discovered gaps.
Negotiating the AI clauses they will send next
Expect the buyer’s paper to demand: no training on customer data (accept; mirror it upstream to your model vendor), output warranties (resist absolute accuracy warranties; offer service levels and correction duties), AI-specific indemnities (cap them; carve out buyer-supplied data), and audit rights (channel into reports and certifications rather than raw system access). Every clause you accept must be enforceable against your vendors. The flow-down check is where most AI contracts silently break.
Question-to-evidence map
| Questionnaire theme | The evidence that answers it | Red flag if missing |
|---|---|---|
| Role & risk classification (1) | One-paragraph classification memo per feature | “We are not covered by the AI Act” with no analysis |
| Models, licences, training use (2–3) | Model inventory + vendor terms extracts | Answers from memory contradicting vendor docs |
| Data location & transfers (4–5) | Data-flow diagram, SCC pack + notification receipt | SCCs signed but never notified |
| Transparency & oversight (6–7) | UI screenshots of notices; oversight points in the flow | Disclosures living only in terms of service |
| Quality & incidents (8–9) | Eval metrics, red-team summary, incident runbook | No definition of what counts as an AI incident |
| Chain, IP & proof (10–12) | Subprocessor list, IP/indemnity positions, policy set | Indemnities accepted downstream with no upstream mirror |
A negotiation vignette
A real pattern from recent deals: the buyer’s template demands an uncapped indemnity for “any claim arising from AI output.” The vendor counters not with a flat refusal but with structure; indemnity capped at fees, covering third-party IP claims over the vendor’s own models and training; express carve-out for buyer-supplied data and buyer-configured prompts; and a service-level commitment to correct systematic output errors within a defined window. The buyer’s legal team accepts in one round: what they actually needed was a story for their own risk committee, not infinite liability. Most AI-clause standoffs dissolve the same way. Convert the absolute ask into an operational commitment you can actually perform, and price the residue into the contract.
Do we need all this if we only use AI internally?
A lighter version, yes. Buyers increasingly ask about internal AI use in vendor risk reviews, because your copilot’s access to their data is their problem too.
Who should own the factsheet?
One named owner with legal review each release, usually whoever owns security questionnaires today. It changes with every model swap, so version it like code.
This week’s homework
Draft the twelve-heading factsheet this week, even in rough form. Then answer question (3), whether customer data trains anything, in writing, from your actual vendor terms, not from memory. That single verified answer unblocks more enterprise deals than any certification.
Related: AI vendor contract clauses playbook · AI Compliance Hub.
Author
If this is on your desk
Templates and checklists are free in the Founder Academy; for a specific situation, book a 30-minute intro call.
Founder AcademyBook an intro callMore from Vircon Insights
AI in Advertising Is Now a Disclosure Question in Türkiye: The Targeted Advertising and AI Rules in Force Since 1 August
August 3, 2026A 12-Month Cap on Off-Exchange Share Sales: Rebuilding the Post-IPO Exit Plan
September 1, 2026Three Ways to Move Up the IPO Queue: Türkiye's New Priority Criteria
September 1, 2026English Is No Longer the Exception on KAP: The Process to Build Before 1 October 2026
August 31, 2026Your Users Are Younger Than Your Terms Say: Children, Age Assurance and AI Products
August 31, 2026Will Insurance Pay When the Model Is Wrong? E&O, Cyber and Product Liability for AI Companies
August 28, 2026Related Practice Areas
ICOs, Crypto & Blockchain
Crypto-asset regulation, token offerings, exchange and custody licensing.
View service →Privacy & Cybersecurity
KVKK and GDPR compliance, breach response, cybersecurity governance.
View service →Mergers & Acquisitions
End-to-end M&A: due diligence, structuring, documentation, negotiation.
View service →