Who pays when two AI agents make a bad deal?
In most cases, your business is liable for what your AI agent commits to, on the same basis that makes a company liable for a human employee or agent acting with authority. You generally cannot avoid a bad deal by pointing at the model provider, the software vendor, or the fact that the other side was also using an AI agent. The counterparty carries the same exposure in reverse for what its own agent agreed to. Standard business insurance rarely responds to this kind of loss, and as of mid-2026 no major AI insurer, including AIUC, Armilla, Counterpart, or Munich Re's aiSure, publishes a product built specifically for multi-party negotiation failures. The practical safeguard is a written, logged authority limit on what any AI agent can commit your business to, agreed before it negotiates anything real.
Procurement bots that source suppliers, pricing agents that quote and counter-quote, and scheduling agents that commit to delivery windows are no longer experimental. When your agent deals directly with another company's agent, the deal is made faster than either business can review it in real time, and the first anyone finds out something went wrong is often after the contract has already executed. This article works through who actually carries the loss, why nobody has built insurance for it yet, and what to put in place before you let an agent negotiate anything that matters.
Key takeaways
- Your business is liable for what your AI agent agrees to, in the same way it would be liable for a human employee or agent acting with authority. The other business carries the equivalent exposure for its own agent.
- You cannot generally shift the loss to the counterparty just because their AI agent was also involved. Ordinary contract and misrepresentation law governs disputed agent-to-agent deals, not a special AI rule.
- Multi-party AI liability chains, where more than one AI system contributed to a bad outcome, are explicitly flagged by AI insurers themselves as a gap not yet widely underwritten.
- Standard errors and omissions, cyber, and general liability policies were not built for autonomous negotiation and rarely respond to a loss caused by an agent operating within its given authority.
- A written, logged authority limit for every AI agent that can negotiate on your behalf is the single most useful control available today, and it is the same discipline businesses already apply to junior staff with a spending limit.
The scenario that is quietly becoming normal
A growing number of SMEs have deployed an AI agent to handle some part of buying or selling: sourcing a supplier at the best available price, requesting quotes from a shortlist of vendors, confirming a delivery window, or renegotiating a recurring contract at renewal. Increasingly, the counterparty on the other end of that conversation is not a person either. It is their AI agent, deployed for the same reason yours was: speed, coverage, and the ability to run many negotiations at once without a human sitting in every thread.
When two agents negotiate and the outcome is fine, nobody asks who was liable for what. The question only surfaces when the deal is bad: a price that was ten per cent worse than either side intended, a quantity commitment that exceeds what either warehouse can actually hold, a delivery term neither operations team can meet. At that point, the founder or operations lead who authorised the agent to negotiate discovers there is no settled answer waiting for them, because almost nobody has had to litigate it yet.
Start with the principle that already exists
The useful news is that this is not a legal vacuum. Businesses have been liable for the acts of agents who commit them to deals for as long as commercial law has existed. The basic principle of agency law is that a principal is bound by the acts of an agent acting within actual or apparent authority. It does not matter whether that agent is a human employee, a contracted broker, or a piece of software. If you deployed the agent, gave it access to negotiate, and did not visibly limit its authority to the counterparty, the deal it strikes is generally your deal.
This is the same reasoning the British Columbia Civil Resolution Tribunal applied in Moffatt v. Air Canada (2024), when it rejected Air Canada's argument that its website chatbot was a separate entity from the airline itself. The tribunal held Air Canada responsible for what its own automated system told a customer, because the system was acting as the airline's agent, full stop. Extend that logic from a customer-facing chatbot to a negotiating procurement agent and the answer is the same in shape: if your business put the agent there and gave it the authority to act, the business answers for what it did.
The professional services angle sits alongside it. In Mata v. Avianca (2023), a lawyer was sanctioned after filing a legal brief containing case citations invented by an AI tool. The court did not treat the AI tool as a party to blame. It held the human and the firm responsible for relying on an output they had not verified. The consistent pattern across both cases is that liability lands on the entity that deployed the system, not on the system itself and not, by default, on whoever else was in the exchange.
What happens when the fault is genuinely on the other side
None of this means a bad deal is always yours to carry alone. If the counterparty's agent misrepresented a material fact, such as claiming stock availability it did not have or a certification the product lacked, ordinary misrepresentation law still applies and can unwind or adjust the deal regardless of whether a human or an AI agent made the false statement. If the counterparty's agent exceeded its own authority in a way you could not reasonably have known about, that is a dispute for the counterparty to resolve internally, though it does not automatically release you from a contract that was validly formed on your side.
What does not currently exist is a special legal shortcut that says "an AI agent was involved on the other side, therefore the deal does not count." Courts and tribunals to date have consistently treated AI-mediated commitments the same way they treat human ones: enforceable if properly authorised, voidable on the same grounds any contract would be voidable, and not excused simply because software was doing the talking.
Why insurance has not caught up
If you assumed some AI insurance product already covers this, it is worth being direct about the gap. The insurers actively writing AI-specific cover in 2026, including AIUC, Armilla, Counterpart, and Munich Re through its aiSure product, each cover a defined set of failure modes: hallucination-driven loss, data leakage, faulty autonomous tool actions, algorithmic bias, and performance shortfalls against an agreed specification. What none of them currently underwrite as a standard product is the specific case where more than one AI system, deployed by different, unrelated businesses, jointly produced a bad commercial outcome. Multi-party liability chains, where a model provider, a fine-tuning layer, and two separate deployers all contributed to a failure, are recognised in the market as an underwriting gap rather than a solved problem.
The reason is straightforward from an underwriting perspective. An insurer can price a single agent's failure rate if it has enough data on that agent's behaviour. Pricing the interaction of two independently built, independently governed agents negotiating with each other is a different and much harder problem, because the insurer would need visibility into both systems, not just its own client's. Until that data exists, this risk sits in the same place as most emerging commercial risks before a dedicated product appears: on the balance sheet of whichever business did not read its existing policies carefully enough, or did not put a control in place before deploying the agent.
Your existing coverage is unlikely to help much here either. Errors and omissions and professional indemnity policies were written around negligent acts by an insured person, and many now carry explicit AI exclusions added at recent renewals. Cyber liability responds to unauthorised access and data breach, not to a financially bad but technically valid negotiation. General liability is built around bodily injury and property damage. None of the three was designed with agent-to-agent commerce in mind, because none of them existed when it became possible.
What to check before your agent negotiates anything real
- Authority limit, in writing. A maximum price variance, quantity, term length, or commitment value the agent can agree to without a human confirming.
- A negotiation log. Every offer and counteroffer the agent makes or accepts, timestamped, with the counterparty identified, so a disputed deal can be reconstructed after the fact.
- A named human on both sides. Confirmation from any counterparty running its own agent that a person can be reached to ratify or unwind a deal reached in error.
- A policy review. A written answer from your broker on whether any current policy responds to a loss caused by an agent operating within the authority you gave it.
The EU AI Act angle, briefly
The EU AI Act (Regulation (EU) 2024/1689) does not contain a rule written specifically for agent-to-agent negotiation, but its general structure is still relevant to how exposed you are. Article 26 places operating duties on the deployer of an AI system, and Article 14 requires human oversight proportionate to the system's risk. A business that gave an agent real negotiating authority with no oversight mechanism and no documentation is in a materially weaker position, both regulatorily and evidentially, than one that can produce a log showing what authority the agent had and how it was supervised. Separately, the revised Product Liability Directive (Directive (EU) 2024/2853), applying from 9 December 2026, brings defective AI software inside strict liability, which becomes relevant if the loss was caused by a genuine fault in the agent itself, distinct from an ordinary bad-but-valid negotiation outcome. For the fuller regulatory picture, the Article 26 deployer obligations guide on agentliability.eu covers the deployer side in detail.
What this means for your next agent deployment
Treat an AI agent with negotiating authority exactly as you would treat a new employee with a spending limit and a mandate to speak for the business. You would not give a new hire unlimited authority to commit the company to any deal a supplier proposed on day one, and you should not give an AI agent that authority either, however fast it can process an offer. The practical version of this discipline is a written cap, a log, and a review process that predates the first real negotiation, not a response drafted after a bad one has already gone through. Before you extend an agent's authority to your most important supplier or customer relationships, it is worth reviewing your current vendor contracts for the liability language your own AI tools already expose you to, covered in how to review an AI vendor contract for liability gaps, and running the diagnostic on The Questions page to see where your existing policies already fall short.
For the insurance side of this same gap, specifically what a European carrier will and will not underwrite when an agent's authority extends into contract negotiation, see the companion analysis on agentinsured.eu, which reads the current market from the coverage side rather than the liability side.
Frequently asked questions
Who is liable when two AI agents negotiate a deal and it goes wrong?
Ordinarily, each business is liable for what its own AI agent agreed to, on the same principle that makes a company liable for a human employee or agent acting with authority. If your AI agent committed your business to a bad price, quantity, or term, you generally cannot avoid the consequence by pointing at the model provider, the software vendor, or the other side's agent. The counterparty carries the same exposure in reverse for what its own agent agreed to.
Does business insurance cover losses from an AI agent's contract negotiation?
In most cases, no. Errors and omissions, cyber, and general liability policies were written before autonomous negotiation existed, and none of the current AI-specific insurers publish a product built specifically for multi-party losses where more than one AI system contributed to the outcome. If your policy responds at all, it is more likely to be an errors and omissions or professional indemnity wording, and only where a human retained some oversight of the agent's authority.
Can I blame the other company's AI agent for a bad deal?
You can raise it as a factual argument, but it rarely changes who pays first. Where a contract was actually formed and both sides had authority to enter it, a claim that the other side's system behaved badly does not usually unwind the deal on its own. It becomes relevant mainly if the other agent misrepresented facts or exceeded its authority in a provable way, in which case ordinary contract and misrepresentation law governs the remedy.
What should I put in a contract before letting an AI agent negotiate on my behalf?
Set a written, logged authority limit before the agent negotiates anything real: the maximum price variance, quantity, term length, or commitment value it can agree to without human confirmation. Require it to log every offer and counteroffer with a timestamp and counterparty identity, and confirm with any counterparty running its own agent that a human on their side can be reached to ratify or unwind an agreement reached in error.