The AI contract: the clauses that protect you when the vendor's model fails
The short answer
The contract is where your Module 3 judgments become enforceable, or fail to
The build-buy-wrap decision and the vendor interrogation produce knowledge; the contract is the only place that knowledge binds anyone. An answer given in a sales call and absent from the paper is a story. Evaluate the paper as the final form of every decision that preceded it.
What you will be able to do
- Evaluate an AI vendor contract clause by clause against your organization's real exposure, judging which terms protect you when the model fails, degrades, changes, or trains on your data, and which only appear to.
- Judge a data-use and training-rights clause: what the vendor may do with your inputs, outputs, and fine-tuning data, whether "service improvement" language quietly licenses training, and what a clean clause states plainly.
- Assess an IP and output indemnity for what it actually covers: the named products and tiers, the conditions (guardrails on, no intentional infringement), the exclusions, and the difference between an indemnity and insurance (see Topic 8.4).
- Weigh audit rights and evidence-access clauses against what you will genuinely need as a deployer: incident evidence, logs, model documentation, and the information the EU AI Act's value-chain and deployer provisions expect to flow to you (Regulation (EU) 2024/1689).
- Trace the subprocessor chain behind your vendor: whose model actually answers your users, what flows down by contract and what does not, and what notice you get when the chain changes.
- Critique model-change and deprecation terms using real deprecations as the measuring stick, from a three-day shutdown to a three-month migration window, and specify the notice, version pinning, and migration help a deployer can reasonably demand.
- Distinguish an uptime SLA from an accuracy commitment, explain why vendors resist warranting output quality, and evaluate the workable middle ground: documented evaluations, acceptance criteria, and your own eval suite as the enforcement mechanism (see Topic 4.2).
- Appraise a liability cap and its carve-outs against the real size of the harm an AI failure can cause, and decide when a cap makes the deal unacceptable at any price.
- Plan the exit before the entry: data return, deletion evidence, output portability, and what happens to your fine-tuned artifacts if the vendor is acquired, collapses, or deprecates the product (see Topic 3.1).
The ideas, one by one
Read the paper the way it was written: for the vendor
Standard AI terms are ordinary, not outrageous, and ordinary means the failure risks of a probabilistic system default to you: outputs disclaimed, warranties limited to uptime, liability capped at fees. The evaluation method is to start from the failure, not the clause: for each way the system can fail, find what the paper says happens next and who pays.
Training rights hide inside "improve the services."
The license grant is the commitment; the marketing page and the sales call are not, as Zoom's 2023 terms episode and Slack's 2024 default-on training showed. A clean clause forbids training on your content without express consent, sets deletion timelines, binds subprocessors, and survives termination. Defaults and opt-out mechanics matter as much as the rights themselves.
The 2023 indemnity wave is real protection with conditions attached
Microsoft, Google, OpenAI, and Anthropic all added output indemnities between September and December 2023, and every one is conditional: guardrails in use, no infringing inputs, covered tiers only, carve-outs for your modifications. Run the five-question read (claims, products, conditions, remedies, carve-outs) against your actual configuration, and verify current terms, because programs change.
An indemnity covers the lawsuit, not the failure
It pays legal costs when a third party sues over an output; it pays nothing when the output is simply wrong and costs your customer money. And it is a counterparty promise, not insurance, while insurers add AI exclusions of their own (see Topic 8.4). Know which risk each instrument actually touches.
Evidence access is what your other obligations are made of
Incident response, disclosure, and conformity work all depend on information only the vendor holds. The EU AI Act expects value-chain information to flow by written agreement (Regulation (EU) 2024/1689, Articles 13, 25(4), 26), and the Commission-published MCC-AI clauses (2025) are a free library of deployer-protective language. Land somewhere on the ladder: documentation, audit reports, scoped audit rights.
The subprocessor chain, not the logo, processes your data
Wrapped products run on upstream models and clouds you have no contract with. Demand visibility of the chain, flow-down of the key promises, notice when the chain changes, and a remedy that runs against your vendor when their chain fails. A no-training clause that stops at the wrapper is a promise about the wrapper.
Deprecation practice ranges from three days to three months, at vendor discretion, unless your contract says otherwise
Codex was shut down on roughly three days' notice in 2023; GPT-4.5 Preview got a three-month window in 2025; Anthropic published preservation commitments in 2025 that remain policy, not your clause. Measure contractual notice against your real migration time; the gap is your outage, and silent model updates need notice tied to your right to re-run evals.
Vendors will not warrant accuracy, but they can warrant evaluations
A blanket correctness guarantee on generation is undraftable; agreed test sets, metrics, thresholds, re-test rights on updates, and defined remedies are draftable and enforceable by the eval suite you build next module (see Topic 4.2). If the contract holds no measurable quality commitment, you have accepted the entire correctness risk, and that acceptance should be conscious.
The liability cap decides who absorbs the tail, and AI tails are correlated
Fee-based caps were sized for failures that hit one user at a time; AI fails at scale, and real cases dwarf any plausible cap (see Topic 3.2) (see Topic 3.5). Compare cap to blast radius, win carve-outs for the worst attributable failures, check the cap's symmetry, and when the gap is intolerable, shrink the deployment or leave.
Negotiate the divorce at the wedding
Exit terms are written when you have leverage and exercised when you have none: data return in usable formats, certified deletion reaching subprocessors, the fate of fine-tuned weights and embeddings, survival of output rights, transition assistance, and continuity triggers on acquisition or insolvency. Simulate the bad ending against the clauses as written before signing.
Spend negotiation capital in the order your failures rank, and keep the walk-away alive
Data rights first, then evidence and incident cooperation, then change notice, then liability architecture, then quality commitments. The strongest clause in any negotiation is the credible ability to leave, which is an architecture decision made back at the build-buy-wrap fork (see Topic 3.2).
The paper cannot make the model accurate, and a right without an owner is decoration
The contract decides who pays and who knows; your scoping, evals, monitoring, and oversight still decide whether customers get hurt. Every negotiated right needs a named owner and trigger inside your organization, and the checklist gets re-run at every renewal and amendment, because vendors amend.
Listen to it
The same lesson, talked through by two hosts as episode 24 of the podcast.
Real cases
The anchor is the 2023 indemnity wave, treated in depth; the other examples each sharpen one clause family, and events owned by other topics appear only as pointers.
Example 1 (the anchor): the 2023 indemnity wave, and the conditions inside it. Between September and December 2023, the four leading AI vendors added output indemnities in sequence: Microsoft's Copilot Copyright Commitment (announced September 2023, effective October 2023, extended to Azure OpenAI Service in November 2023 as the Customer Copyright Commitment), Google Cloud's two-pronged indemnity covering training data and generated output (October 2023), OpenAI's Copyright Shield for ChatGPT Enterprise and API (November 2023), and Anthropic's updated commercial terms defending API customers over authorized use and outputs (December 2023, effective early 2024). Read as a market event, the wave shows contract terms are competitive products: buyer pressure moved four legal departments in four months. Read as clauses, every commitment is conditional: Microsoft's requires the built-in content filters and guardrails in use and no attempt to generate infringing material; Google's requires adherence to its responsible AI practices; OpenAI's and Anthropic's cover paid, authorized tiers. For a deployer the wave teaches the whole method of 3D: identify the covered products, the conditions, and the carve-outs, verify your deployment actually satisfies the conditions (a team that disabled a content filter to reduce false positives may have silently exited the protection), and confirm the current program terms, since all of these are as of 2023 announcements and vendors amend terms (Microsoft, 2023; Google Cloud, 2023; OpenAI, 2023; Anthropic, 2023).
Example 2 (data-use clauses): Zoom's August 2023 terms revolt. Zoom's March 2023 terms-of-service update granted the company sweeping license rights to customer content that readers understood to permit AI training without opt-out; when the language surfaced publicly in August 2023, the backlash forced revisions within days, ending in an explicit commitment not to train AI models on customer audio, video, or chat content (TechCrunch and CBS News, August 2023). The lesson for 3C: the drafting alone was the exposure. Customers who had signed had granted the rights whether or not Zoom ever exercised them, and the protection came from public pressure, which is not a clause. A contract evaluation would have caught the grant in March; the market caught it in August.
Example 3 (data-use defaults): Slack's opt-out by administrator email. In May 2024, attention to Slack's privacy principles revealed customer data training the company's machine-learning models by default, with opt-out only via a workspace administrator emailing the vendor; Slack clarified that its Slack AI add-on's large language models were not trained on customer data, but the default-on setting and the buried exit were the point (TechCrunch and The Register, May 2024). For your checklist: evaluate not just what rights exist but their default state and the mechanics of exercising them, and notice the tier split between the product's AI add-on and its underlying models, the kind of distinction that only a careful read surfaces.
Example 4 (deprecation, worst case): Codex and the three-day sunset. In March 2023 OpenAI told developers it was discontinuing its Codex code models, with the shutdown following roughly three days later; the models had been free in beta, and OpenAI pointed users to newer general models, but teams with Codex in production had three days to migrate, and developer confidence in the platform's stability took a documented hit (reported by Simon Willison and The Decoder, March 2023). This is the empirical floor for 3G: absent contractual notice, deprecation can be near-instant, and "the vendor would never" is not a clause.
Example 5 (deprecation, better case): GPT-4.5 Preview's three-month window. In April 2025 OpenAI announced the GPT-4.5 Preview model would be removed from its API on 14 July 2025, a three-month migration window with a named successor model, and developers who had built on its specific qualities still objected publicly (OpenAI and VentureBeat, 2025). Paired with Example 4, it brackets the real range of vendor practice, three days to three months, and proves the range is set by vendor discretion unless your contract sets it instead. A deployer whose re-evaluation and redeployment cycle takes six months should read both examples as a specification for the notice clause they need.
Example 6 (deprecation policy as commitment): Anthropic's model-preservation undertaking. In November 2025 Anthropic published commitments on model deprecation and preservation: preserving the weights of all models with significant use for the lifetime of the company and instituting a documented process around retirement, explicitly citing the costs customers bear when a depended-on model disappears (Anthropic, 2025). Epistemically this is a published policy commitment, not a contract term in your agreement, and the evaluation discipline treats those differently: a policy can change with a blog post; a clause changes with your consent. Its value in negotiation is as precedent: a frontier vendor has now stated on the record that abrupt model loss harms customers, which is a useful sentence to quote when asking for deprecation notice in writing.
Example 7 (the information floor, EU): Article 25(4) and the MCC-AI clause library. The EU AI Act requires, for high-risk systems, that providers and their component suppliers specify by written agreement the information, capabilities, technical access, and assistance needed for compliance (Regulation (EU) 2024/1689, Article 25(4)), and deployers hold their own duties under Article 26, including operating per instructions and retaining logs. In March 2025 the European Commission's Public Buyers Community published updated model contractual clauses for AI procurement (MCC-AI), a full high-risk version aligned to the Act plus a light version and commentary, voluntary but drafted precisely to put those information flows on paper (European Commission Public Buyers Community, 2025). Together they mark the direction of travel for 3E: in regulated deployments, evidence access is stopping being a negotiating preference and becoming the shape compliance takes. A deployer anywhere can borrow the MCC-AI language; a deployer in scope of the Act may need to.
Example 8 (the accuracy-commitment frontier, illustrative method). Consider the standard warranty section of nearly any AI vendor's terms: the service is provided as is, outputs may contain errors, the customer is responsible for verifying outputs before relying on them. This is not one company's clause; it is the industry default, and it means the correctness risk of the system sits entirely with the deployer unless something replaces it. The replacement that works in practice is the evaluation-based commitment of 3H: agreed test sets, agreed metrics, agreed thresholds at delivery and after material updates, agreed remedies. The method example: a procurement team that arrives with its own eval suite (see Topic 4.2) and asks the vendor to warrant performance on it has converted an undraftable demand ("be accurate") into a draftable one, and the vendor's reaction to that request is itself diagnostic, the contract-stage version of the vendor interrogation (see Topic 3.3).
Example 9 (the cap-versus-harm gap, by pointer). The scale mismatch between contract caps and AI blast radius needs no invented numbers; the program has already walked two real cases whose deep treatments live elsewhere: an automated government system whose wrongful fraud accusations led to a settlement dwarfing any conceivable vendor fee cap (see Topic 3.2), and a trading firm that lost hundreds of millions in under an hour of automated malfunction (see Topic 3.5). Bring those magnitudes to 3I's evaluation: excluded consequential damages plus a fee-based cap means the difference between those numbers and your annual invoice is, by default, yours. The insurance market's response, AI-specific exclusions appearing in commercial policies, is owned by the insurance topic (see Topic 8.4) and completes the picture: the tail risk is not quietly covered somewhere else.
Example 10 (exit under the worst conditions, by pointer). When the telehealth company Babylon collapsed into administration in 2023, every organization that had built on it ran its exit clauses under the least favorable conditions possible; the case's deep treatment anchors the scoping topic (see Topic 3.1). Here it stands for 3J's simulation: exit terms are evaluated at signing precisely because, when they finally matter, the counterparty may be a bankruptcy estate, the engineers who understood the data formats may be gone, and every ambiguity in the clause resolves against the party with no leverage, which is you. The wave of AI startups and the consolidation around them make vendor discontinuity an ordinary planning assumption, not a pessimist's edge case; that projection is labeled as a projection, and the Babylon collapse grounding it is fact.
Where people go wrong
- "The vendor has an indemnity, so the copyright risk is handled." The indemnity is conditional and scoped: named products and tiers, guardrails in use, no infringing inputs, carve-outs for your modifications and combinations. A team that disabled a content filter or uses the consumer tier may be outside the protection entirely. Read the conditions against your actual deployment, and verify the current terms, because vendors amend them.
- "Standard terms are non-negotiable, so reviewing them is pointless." The review is what tells you what you are absorbing, which drives decisions you fully control: scope of deployment, human oversight, second-sourcing, insurance, or walking away. And terms move more than assumed: the entire 2023 indemnity wave was vendors changing standard terms under buyer pressure, and smaller vendors negotiate constantly.
- "The sales team confirmed they do not train on our data." If the license grant says "provide and improve the services" and nothing excludes training, the paper permits what the sales call denied. Zoom's 2023 episode began exactly there: drafting broad enough to permit training created the exposure regardless of practice. The clause is the commitment; the call is a memory.
- "We have a signed NDA, so our data is protected." A non-disclosure agreement restricts the vendor from disclosing your information to outsiders; it says nothing about the vendor using that information internally to improve or train its own models, which is not disclosure at all. An NDA and a no-training clause protect against different failures, and a vendor who points to the NDA when asked about training rights has answered a question you did not ask.
- "An indemnity is basically insurance." An indemnity is a promise to pay from one counterparty, as good as its balance sheet, its lawyers' reading of the conditions, and its survival. Insurance is a regulated third-party risk transfer, and insurers have been adding AI-specific exclusions of their own (see Topic 8.4). Confusing the two leaves you double-counting protection you may not have once.
- "Uptime SLA of 99.9 percent means the AI is reliable." The uptime SLA promises the service answers; it says nothing about whether the answers are right. Output quality is typically disclaimed entirely. If the contract contains no measurable quality commitment, the correctness risk is all yours, enforced only by your own evals (see Topic 4.2), and that is a decision to make consciously, not by omission.
- "Deprecation is a distant, theoretical risk." A major provider shut down its Codex models on roughly three days' notice in March 2023, and a 2025 API deprecation with a three-month window still disrupted developers. The range of real practice is days to months, set by vendor discretion unless your contract sets it. Measure the notice you can get against the migration time you actually need.
- "We contracted with the vendor, so the vendor is who we need to evaluate." Your data and your users' outputs run through the vendor's upstream model provider, cloud, and subprocessors. A no-training promise that does not flow down the chain protects you from the wrapper, not from the processing. Demand visibility of the chain, flow-down of the key promises, and notice when the chain changes.
- "The liability cap is standard, so it is fine." The cap being standard is exactly why it deserves evaluation: standard caps were sized for software that fails one user at a time, not for correlated AI failures across thousands of decisions. Compare the cap to your deployment's realistic blast radius, push the worst attributable failures into carve-outs, and if the gap is intolerable, shrink the deployment, not just your expectations.
- "We will negotiate the exit terms if we ever leave." You will exercise exit terms at the moment of minimum leverage: dispute, deprecation, acquisition, or the vendor's insolvency. Babylon's collapse is the standing lesson (see Topic 3.1). Data return, deletion certification, fine-tuned artifact handling, output rights survival, and transition assistance are negotiated at signing or effectively never.
- "Legal owns the contract review." Counsel owns the drafting and the law. Only you know the system's failure modes, the eval results, the incident history, and the blast radius, which is the knowledge that decides which clauses matter and what the asks are. A contract reviewed by counsel alone is legally sound and operationally blind; the checklist is built for the two of you to run together.
- "Once signed, the contract work is done." Vendors amend terms, swap upstream providers, and publish policy changes; renewals arrive with new paper. And every right you won is inert without an owner: a deprecation notice needs a watched inbox, an eval-trigger right needs someone who re-runs the evals. Assign owners to the clauses and re-run the checklist at every renewal and amendment.
- "A published vendor policy is as good as a clause." A policy commitment, like a deprecation or model-preservation policy, can change with a blog post and binds the vendor only as far as its reputation. A clause changes with your consent and binds in court. Cite policies in negotiation as precedent; rely on clauses.
- "The EU AI Act is the provider's problem, not ours." Deployers of high-risk systems hold their own obligations (Regulation (EU) 2024/1689, Article 26), including operating per the instructions for use and retaining logs, and the Act expects information to flow along the value chain by written agreement (Article 25(4)). If your deployment is in scope, evidence-access clauses are not a preference; they are what your own compliance is made of (see Topic 5.6).
- "Free and pilot tiers are just the paid product without the invoice." Free, trial, and consumer tiers commonly run on different terms: broader data-use rights, no indemnity (OpenAI's Copyright Shield, for instance, excluded free tiers from the start), no SLA, and weaker deletion commitments. A pilot run on a free tier can grant training rights over the very data the eventual enterprise contract forbids training on. Clear the pilot's terms with the same checklist before the first prompt is sent.
- "Our own customer contracts are someone else's problem." The clauses you demand upstream are the clauses your customers will demand from you, because to them, you are the AI vendor. Every evaluation skill in this topic runs in both directions: your disclaimers, your caps, your data-use language, and your incident-notice duties downstream should be written by someone who has read them the way you just read your vendor's. Asymmetry between what you demand and what you grant is a finding about your own paper.
- "We should demand an accuracy guarantee and refuse to sign without one." A blanket correctness warranty on open-ended generation is a demand no serious vendor can grant, and insisting on it wastes the negotiation. The winnable version is evaluation-based: agreed test sets, metrics, thresholds, re-test rights on material updates, and defined remedies. Arrive with your own eval suite and the undraftable becomes draftable (see Topic 4.2).
Questions people ask
- What is data-use and training-rights clause?
- The contract terms defining what the vendor may do with your inputs, outputs, and datasets, especially whether it may train or improve models on them. A clean clause requires express consent for training, sets deletion timelines, binds subprocessors, and survives termination.
- What is "Improve the services" language?
- The common license-grant phrasing in which training rights hide: a grant to use customer content "to provide, maintain, and improve the services" can be read to permit model training unless training is expressly excluded.
- What is output indemnity?
- A vendor's conditional promise to defend the customer and pay judgments or settlements on defined third-party claims (typically copyright) arising from the AI's outputs. Introduced at scale by the 2023 wave: Microsoft (September 2023), Google Cloud (October 2023), OpenAI (November 2023), Anthropic (December 2023), each with conditions and scope limits.
- What is indemnity conditions?
- The customer-side requirements an indemnity depends on, such as using the built-in guardrails and content filters, not attempting to generate infringing material, not supplying inputs you lack rights to, and using a covered product tier. Violating a condition can silently forfeit the protection.
- What is Customer Copyright Commitment?
- Microsoft's name (from November 2023) for its expanded copyright indemnification program covering commercial Copilot and Azure OpenAI Service customers, originally announced as the Copilot Copyright Commitment in September 2023; conditional on guardrail use, and Microsoft's own documentation now requires a formal approval process before a customer may turn off or modify the required content filters and keep coverage, a sign that indemnity conditions tend to tighten over time rather than loosen. Verify current name and scope in live terms.
Keep going
This lesson builds Contract terms that allocate AI risk, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →