The vendor controls the AI and you sign for the risk
AI vendors are shifting risk onto their customers by contract. Here's how to negotiate control, evidence and remedies before an agent fails.
AI contracts already decide who pays when the technology fails.
Major vendors are rewriting their terms so that customers bear the consequences of their AI. If your contracts don’t define control, evidence and remedies, your AI governance policy is just a good intention.
A third-party AI agent sends a customer a quote with a discount nobody approved. The customer accepts it. Your legal team reviews the contract with the vendor and finds that, under the terms, responsibility for generated content is entirely yours.
This isn’t a remote hypothetical. It’s the direction contracts are moving in.
In early November, Alla Valente, a principal analyst at Forrester, spotlighted a concrete example: the update to Meta’s commercial terms, which shifts broad responsibility for its AI features and their output onto business users. Continuing to use the service after November 20, 2026 amounts to accepting them, whether or not your legal team has read them.
In this article you’ll see why this isn’t a single-vendor problem, what changes when AI moves from answering to acting, and which specific clauses you can demand before signing or renewing.
| Fact | What it means for you |
|---|---|
| Meta’s new commercial terms, effective November 20, 2026 | Continued use means accepting the risk of its AI features |
| Bipartisan AI Agent Accountability Act proposal in the U.S. | Lawmakers already see autonomy as a possible escape route from liability |
| Forrester’s report “Contracts Are The AI Guardrails With Teeth” | The contract is the bridge between governance policy and enforcing control |
The contract you accept just by continuing to use it
Valente sums it up with humor: Meta gave the “it wasn’t me” defense a contractual version. The technology is the vendor’s; the responsibility is the customer’s.
The underlying problem isn’t Meta. It’s the mechanism. Many AI services are contracted under standard terms that get updated unilaterally and accepted through use. Nobody signs anything new. Nothing goes through the procurement committee.
In the region, this weighs more heavily. Much of Latin American companies’ customer service, sales and marketing runs on third-party platforms with built-in AI. Meta’s terms cover commercial use of its products: ads and marketing messages, catalogs, Pages and tools such as Meta Business Suite. Outside the United States, they send claims to courts in California or Ireland. It’s worth having your legal team review how this applies in your country.
A term accepted through use is a risk decision nobody in your organization made.
Agent autonomy opens an accountability gap
While AI only suggested text, the potential damage was contained. A person reviewed before sending.
With agents, that changes. An agent can make decisions, invoke tools, generate code or interact with other systems. When something goes wrong, pointing at the technology doesn’t answer the question that matters: who had control and who absorbs the consequences.
Washington is grappling with the same thing. The proposed AI Agent Accountability Act seeks to extend liability under the federal Computer Fraud and Abuse Act (CFAA) to operators and developers of agents that recklessly cause damage and, in some cases, to developers that fail to implement reasonable safeguards.
It’s only a bill, it focuses on hacking damage and it applies in the United States. But it signals the trend: without enforceable controls, autonomy becomes a way to dodge accountability. And if your vendor operates from there, that discussion will reach your contracts too.
Your governance policy has no teeth without a contract
Many companies already have an AI committee, an acceptable use policy and an inventory of use cases. It’s a good start.
Forrester’s core argument is that it isn’t enough. Having policies and identifying risks doesn’t guarantee anyone will answer when a third party’s AI fails, whether from a model change, a system outage or an agent exceeding its authority.
- Policy without a contract: you know what shouldn’t happen, but you can’t demand anything when it does.
- Policy with a contract: you can ask for evidence, suspend the service and recover costs.
That’s the argument to take to the board: AI governance doesn’t end at internal policy. It ends at what you can legally require of whoever controls the technology.
Four clauses that change who pays
Forrester proposes four negotiating fronts before accepting third-party AI terms.
Explicit limits on agent autonomy
Define in writing what the agent can do, what it can’t and what requires human approval.
- What data it can read and under what identity it accesses it.
- Which tools it can invoke and which transactions it can initiate.
- Which decisions require a human in the loop.
- Whether the limits are enforced by policy, by code or both.
The more the agent acts in other systems, the more precise those limits need to be.
Evidence, not promises
“Trust us, we have safeguards” won’t help you reconstruct an incident.
- Logs that show what the agent did and why.
- Advance notice of material changes to the model or its controls.
- Incident notification deadlines.
- Vendor support for forensic investigation.
Accountability proportional to control
Reject blanket risk transfers when the vendor controls the model, its safeguards or changes you can’t see.
- Using a technology shouldn’t automatically mean assuming its risk.
- If the vendor controls the conditions that create the risk, it should share it.
Real remedies
A safeguard with no consequence for breaching it is an aspiration, not a control.
- A remediation obligation with deadlines.
- The right to suspend the service and terminate the contract.
- Meaningful financial compensation, defined before the incident.
If you can’t say what you’ll require of the vendor when its agent goes out of bounds, you don’t have a control. You have a hope.
A stress test for every AI vendor
It all comes down to a question Valente poses: if this agent exceeds its bounds tomorrow, what can we contractually require the vendor to do, and who pays when the safeguards fail?
Use it to review your current portfolio:
- Inventory of embedded AI. List third-party services with active AI features, including those that arrived with an update nobody asked for.
- Review of current terms. Identify which ones are accepted through continued use and which shift responsibility for content or actions onto the customer.
- Classification by autonomy. Separate AI that suggests from AI that acts. Prioritize agents that access data, execute transactions or operate in other systems.
- Run the question. For each priority vendor, answer with the contract in hand: what evidence can I ask for, what can I suspend and who pays?
- Renegotiation plan. Take the gaps to procurement and legal before the next renewal, not after the next incident.
What the skeptic will say
“You don’t negotiate with a vendor the size of Meta or a hyperscaler.” It’s true that bargaining power is uneven. But not negotiating doesn’t eliminate the risk; it only leaves it unpriced. If you can’t change the terms, at least you’ll know what risk you’re taking on, you can limit the use cases and you can document the decision before the board. And in higher-volume enterprise contracts, there’s usually room.
“This is legal’s job, not IT’s.” Legal drafts the clauses. But only IT knows what data the agent touches, what permissions it runs with and what logs are needed to investigate an incident. Without that input, legal negotiates blind.
“Our AI use cases are still pilots.” Pilots turn into production without going back through the contract table. The best time to set the terms is before the agent becomes indispensable and you lose bargaining power.
Risk doesn’t disappear: it gets assigned
Every AI contract your company accepts answers, whether you know it or not, one question: who pays when the technology fails.
If you don’t take part in that answer, the vendor writes it. And the most recent terms show it’s writing it in its own favor.
Agent autonomy will keep growing. Contracts have to grow with it.
“It wasn’t me” is not an accountability model.
Your first step this week: pick the AI vendor with the most autonomy inside your operation. Sit down with legal and answer the stress test question with the current contract in hand. If the answer is “we don’t know,” you already have your priority for next quarter.
Sources
- Valente, A. (2026). “It Wasn’t Me” Is Not An AI Risk Management Strategy. Forrester.
- Valente, A. Contracts Are The AI Guardrails With Teeth. Forrester.
- Meta. Commercial terms preview.
- MeriTalk. Coverage of the AI Agent Accountability Act proposal.
You might also like


AI Automation Doesn't Fail Because of the Model: It Fails Because of the System
September 28, 2026
Read More
