Who Pays When the AI Is Wrong: The Indemnity Question Now Deciding Enterprise AI Deals

Your product works. The security questionnaire came back clean. The champion is bought in and the business case cleared finance. Then the redline arrives from the buyer's legal team, and somewhere around section eleven there is a paragraph that did not exist in last year's version of this contract. If your AI produces an output that is wrong, and our people act on it, and we get sued or fined, who pays?

Most vendors do not have a rehearsed answer to that. Some have never been asked it before this year. And the ones who stall on it are watching deals sit in legal review while a competitor with a written position closes.

For Revenue Leaders, Sales Executives, Legal Partners, and GTM Teams Selling AI-Powered Products

This is not a contract drafting article. It is about a commercial question that got promoted, in about eighteen months, from a footnote nobody negotiated to a gate that stops deals in the final stretch. Your liability position has become part of what you are selling, whether or not anyone in your go-to-market org has looked at it.

The clause that showed up in every enterprise redline

Start with why buyers got serious, because it was not a sudden attack of caution. It was reporting obligations and lived experience.

Seventy-two percent of S&P 500 companies now flag AI as a material risk in their public disclosures, up from 12% in 2023. When a company tells its investors that AI creates material exposure, procurement and legal inherit a mandate. The people across the table from your rep are not being difficult. They are following instructions that came down from the audit committee.

Then there is what buyers have already been through. A Cloud Security Alliance survey fielded in January 2026 with 418 IT and security professionals found that 65% of enterprises had experienced an AI agent related incident in the previous twelve months, and 35% took a financial loss from one. A DigiCert survey of 1,001 IT and security leaders across the US, UK, and Australia found that half had dealt with an incident tied to an unauthorized or misconfigured AI agent in just the previous six months.

So the buyer questioning your indemnity language has probably already cleaned up after somebody else's AI. The question is not theoretical to them, and answering it as though it were is the fastest way to sound naive.

Buyers priced this risk after watching other people eat it

Nothing changes procurement behavior like a public example with a number attached.

Deloitte Australia produced a compliance framework review for the country's Department of Employment and Workplace Relations at a cost of roughly A$440,000, about US$290,000. After publication in August 2025, a university researcher found the report contained a fabricated quote attributed to a real federal court judgment and citations to academic papers that did not exist. Deloitte repaid the final installment of the contract, about A$97,000, and the revised version carried a disclosure that a generative AI system had been used in drafting.

The refund was small relative to the contract. The story was not, because it gave every risk committee a concrete scenario to point at: a sophisticated buyer, a reputable vendor, a deliverable that looked finished and was not.

The legal profession supplied the volume data. Researchers tracking AI-fabricated content in court filings had documented 1,313 proceedings as of April 2026, with 496 of them involving licensed attorneys. Sanctions in a single matter have reached $55,597, up from around $5,000 in 2023. Stanford researchers measuring purpose-built legal research tools found hallucination rates above 17% for one major product and above 34% for another, with general-purpose models landing between 69% and 88% on legal queries.

Your buyer's counsel has read those numbers. They inform the posture that shows up in your redline. The working assumption is no longer that the AI works. It is that the AI will be wrong at some rate, that the rate is not zero, and that someone needs to own what happens on those occasions.

The contract math is lopsided, and buyers can finally see it

Standard AI vendor paper is more one-sided than standard SaaS paper, and the gap is now measurable.

Analysis of contract terms by TermScout found that 88% of AI vendors cap their own liability, roughly in line with the 81% seen across SaaS generally. But only 38% of AI vendor agreements cap the customer's liability, compared with 44% in broader SaaS. Add the two moves that have become common in AI terms, disclaiming output accuracy and carving AI-generated outputs out of the IP indemnity, and the shape of the deal is clear enough. Accuracy risk sits with the customer, and the customer's exposure is uncapped while yours is not.

That arrangement survived for a while because reading contracts closely was expensive. It is not anymore. Forty-four percent of organizations now use AI somewhere in their contracting workflow, with redlining, review, and summarization leading adoption. Benchmarking your terms against market standard used to take a specialist a week and get skipped on smaller deals. It now takes an afternoon and gets done on everything.

Assume the buyer knows how your paper compares to market before your first call with their legal team. If your position is more aggressive than the norm, that will be raised as a fact, not a suspicion, and your rep needs to know it is coming.

The insurance layer stepped back at the same moment

The other half of this story rarely reaches the sales floor, which is a problem, because it changes what your company can credibly promise.

ISO made new exclusionary endorsements available in January 2026. Two of them, CG 40 47 and CG 40 48, remove commercial general liability coverage for harms arising out of generative AI. A third, CG 35 08, does the same on the products and completed operations side. Carriers have been filing with state regulators to use them, and some are going further with absolute AI exclusions that void coverage wherever AI use or development is involved. A standalone AI liability market is forming to fill the gap and is projected to reach $4.8 billion by 2032, but that is a market under construction rather than one you can lean on this quarter.

The practical consequence for revenue teams is direct. A certificate of insurance may no longer answer the question a buyer's risk team is asking, and additional insured status on a policy with AI exclusions gives them nothing. Before anyone in your deal cycle offers coverage as a concession, somebody needs to confirm the coverage exists. A signed indemnity you cannot fund is worse than a hard conversation in week three.

Regulation moved the deadline, not the question

Anyone waiting for regulatory clarity to settle this got a lesson in July.

The EU's Digital Omnibus entered into force on 27 July 2026, deferring high-risk obligations for Annex III systems from 2 August 2026 to 2 December 2027, and pushing systems embedded in regulated products to August 2028. Article 50 transparency and AI content labeling duties stayed on their original schedule, as did the general-purpose AI provider obligations that have applied since August 2025.

Read that as a buyer and the takeaway is not relief. It is that the timeline can move again, in either direction, while the contract you sign this quarter runs for three years. When regulatory certainty is unavailable, risk teams do the only thing left to them and push the exposure into commercial terms. The deferral raised the pressure on your contract language rather than lowering it.

There is an opening in that for vendors willing to take it. Buyers are short on certainty right now, and a clear, documented, defensible position is a form of certainty you can actually supply.

This is a revenue problem wearing legal clothing

The cost of not having an answer shows up in the forecast rather than the legal budget.

Vendor risk assessment appears in 18% to 32% of deals and adds 14 to 28 days to the procurement phase. Legal review tends to be the bottleneck in the last 30 days of a cycle, which is exactly where AI liability questions land. When your answer requires escalating to a general counsel who is juggling four other priorities, add another two weeks. That is a deal slipping a quarter, not because the buyer said no, but because nobody on your side could resolve a liability question inside a week.

Now compare that with vendors who decided their position in advance. Microsoft's Copilot Copyright Commitment applies to covered enterprise SKUs without per-contract negotiation. OpenAI introduced its Copyright Shield in November 2023 with substantively similar protection for enterprise customers, and Anthropic, Google, and IBM offer their own versions, generally conditioned on using guardrails, content filters, and generally available features.

Whatever you think of the scope of those programs, and the conditions attached to them mean most real deployments break at least one requirement, notice the structural advantage. The buyer's question is answered before it is asked. Their legal review becomes verification rather than negotiation. A standardized protection program is a cycle-time asset, and it happens to be written by lawyers.

When the output stops being advice and starts being an action

Everything above describes the easier version of the problem, where AI produces something a human reviews before it matters. That is not where the market is heading, and buyers know it.

Fifty-three percent of executives expect AI agents to negotiate customer and supplier deals autonomously within twelve months. In the same research, 55% named output quality as a significant concern and 44% said they do not yet have enough trust in autonomous capability. Those two findings are not in tension. They describe a buyer who intends to deploy agents on a timeline shorter than their own confidence supports, which is precisely the buyer who negotiates hard on liability.

Human review is the load-bearing assumption in most AI contract terms today. Accuracy risk sits with the customer because the customer is presumed to check the work. Take the human out of the loop, as agentic deployments do by design, and that assumption collapses. An agent that misprices a quote, sends the wrong data to the wrong party, or approves something it should have flagged does not produce an output someone can catch. It produces a transaction.

Separately, 74% of businesses report concern about AI privacy risk while buying AI tools anyway, often through a procurement process built for conventional software. The gap between what buyers are deploying and what their contracts contemplate is exactly the gap their legal teams are trying to close in your redline. If your product takes autonomous action, expect the liability conversation to arrive earlier in the cycle and to be less negotiable than it is for vendors shipping a copilot.

Price that in. Some vendors now tier commercial terms by deployment mode, offering different caps and review obligations for supervised versus autonomous use. It is a more honest conversation than pretending one set of terms fits both, and it gives your rep something to trade other than discount.

What a defensible position actually looks like

You do not need to indemnify everything. You need a position you can state in a sentence, defend with evidence, and repeat in every deal without escalation.

Decide it before the redline, not during. Write it down, version it, and give one person ownership. Tier it by use case, because the risk profile of an AI that drafts an internal summary and one that takes an action against a customer record are not the same thing, and pricing them identically means you are either overexposed on one or uncompetitive on the other.

Separate the three risks buyers tend to bundle into one paragraph. IP infringement from outputs is often insurable and increasingly indemnifiable, which is why the large model providers went there first. Data handling and security is familiar ground with established market terms. Output accuracy is the hard one, and a blanket accuracy warranty is not something most companies can responsibly write. Saying that plainly, then offering something real on the parts you can cover, lands better with a risk team than a wall of disclaimers. Elevated caps for IP and data claims, notification when a model changes materially, and human review obligations framed as shared duties rather than customer-only disclaimers all give the other side something to work with.

Make the position provable. Indemnities get underwritten on evidence, and so does buyer trust. Logs, audit trails, evaluation results against your own error thresholds, and documentation of model provenance are what turn a promise into something a risk committee can approve. Vendors who cannot show how they detect a bad output are asking buyers to accept a commitment with no instrumentation behind it.

Have a real answer for the cases where you cannot give ground. Plenty of companies, especially early stage ones, genuinely cannot write an uncapped indemnity against output error without betting the company on it. Saying no is fine. Saying no with nothing attached is what kills deals. Offer the things that cost you less than the indemnity does and still reduce the buyer's exposure: a scoped initial deployment with defined use cases, published error rates from your own testing, contractual notice before model changes, an agreed process for pulling the system out of a workflow that starts producing bad results, and a review point where terms get revisited once both sides have production data. Risk teams accept limits far more readily than they accept vagueness.

Then publish it. A standard AI addendum and a public terms page that your own legal team has pre-approved removes the slowest step in the cycle, which is usually the gap between the buyer's question and your first substantive response. Give reps a one-page brief covering what you indemnify, what you cap, what you will not agree to under any circumstances, and who can approve exceptions at what deal size. Reps do not need to be lawyers here. They need to stop saying "let me check with legal" to a question that gets asked in every deal.

The vendors this sorts

Buyers are not going to stop asking. Two thirds of enterprises have already had an AI incident, their own filings call it a material risk, and their insurance is quietly retreating from the exposure. The question is now standard, and standard questions reward preparation.

What it sorts on is response time. Two vendors with comparable products will get the same redline in the same week. One answers in a call with a document their GC signed off on last quarter. The other opens a Slack thread that ends in a meeting that ends in a legal review that ends in an escalation, and their deal closes a quarter later, if it closes.

That difference does not show up in a product comparison. It shows up in the close rate.