An AI agent sends an agreement with the wrong price, the wrong terms, or to the wrong person, and someone signs it. Who’s responsible? The law already has an answer. The good news: with the right guardrails, this risk is very manageable.
Key takeaways
- Under US law, the actions of an “electronic agent” generally bind the business that deployed it. “The AI did it” is not a defense.
- Leading law firms warn that today’s standard technology contracts often leave the customer, not the AI vendor, holding the risk of an agent’s actions.
- UETA contains a little-known error provision: a person may be able to undo an automated transaction if the system gave them no chance to catch the mistake. Clear signing ceremonies protect both sides.
- The practical answer is guardrails, not avoidance: scoped authority, locked templates, human checkpoints, real-time visibility, and evidence by default.
- BoldSign gives AI agreement workflows those guardrails out of the box.
A scenario every operations leader should think through
Imagine your company runs an AI assistant that handles customer renewals. One afternoon it prepares a renewal agreement, quotes last year’s discounted price by mistake, and sends it for signature. The customer, delighted, signs immediately.
Are you bound to that price? Can you unwind it? And whose problem is this: yours, the AI vendor’s, or the customer’s?
These questions used to be hypothetical. They aren’t anymore. Deloitte’s 2026 Technology, Media & Telecommunications Predictions report expects SaaS applications to become more adaptive and autonomous, and forecasts that by the end of 2026 “as many as 75% of companies may invest in agentic AI”. When software starts acting on its own, agreements follow. So let’s answer the question properly.
A quick note before we start: this article is general information, not legal advice. Talk to your counsel about your specific situation.
The short answer: liability follows the deployer
US law addressed automated contracting decades ago. The Uniform Electronic Transactions Act (UETA), adopted in nearly every state, recognizes “electronic agents”: software that initiates actions or responds to records without human review. Its rule is direct. A contract can be formed through an electronic agent’s actions even if no individual was aware of or reviewed those actions, and the Uniform Law Commission’s commentary makes the attribution clear: an agent set up and used by a person binds that person. The federal ESIGN Act contains a matching provision, and Stanford Law School’s CodeX center confirms this framework applies to today’s AI transactional agents.
In plain English: the law treats your AI agent like your employee’s keyboard, not like an independent decision-maker. If your agent sends it, your business generally stands behind it.
That sounds scary until you notice what it really means. The law isn’t punishing you for using AI. It’s telling you where to invest: in controlling what your agent is allowed to do, and in keeping proof of what it actually did.
The four building blocks still apply
When considering this, it’s key to note that the elements from contract law still apply.
Offer. A commitment, with definite terms, that invites the other side to close the deal by saying yes. Quotes and marketing usually aren’t offers, but a document that looks like a real commitment often is.
Acceptance. The other side’s yes, on the offer’s terms. Send it, and under the “mailbox rule” it can take effect the moment it’s dispatched, before it even lands in your inbox.
Consideration. Each side gives something of value. It’s a low bar, but without it there’s no enforceable contract.
Revocation. You can usually pull an offer back, but only until it’s accepted, and your revocation only counts once the other side receives it. Acceptance, meanwhile, counts on sending. That timing gap decides who wins the race.
None of this bends for software. An AI is a tool you chose to use, not a legal actor that absorbs the blame.
What the top law firms are telling their clients
Three of the world’s leading firms have published guidance on agentic AI, and they converge on the same themes.
The clock is brutal. To escape, your revocation has to reach the counterparty before their acceptance goes out. Miss that window and you’re left arguing for rescission, which is slower, costlier, and far from guaranteed.
The liability gap sits in your vendor contracts. Clifford Chance warns that under most current technology agreements, the customer, not the AI provider, bears the risk of an agent’s actions. Standard SaaS terms were written for software that waits for humans, not software that acts. If your agent misfires, don’t expect the vendor’s contract to catch you.
The contracting model itself is changing. Mayer Brown advises that agreements for agentic AI should shift from a SaaS model toward a services model, and should explicitly define the agent’s delegation of authority: precisely what it may do, under what policy guardrails, with what audit rights.
Buyers are demanding oversight rights. Morgan Lewis reports that sophisticated buyers now negotiate governance and audit rights into AI deals as standard terms.
Notice the pattern. The legal profession’s answer to AI agent risk is not “don’t use agents.” It’s authority, guardrails, and evidence. Those are contract concepts, but every one of them has a direct technical counterpart in how you build the workflow.
Your letterhead works against you. Ambiguities are read against the drafter (contra proferentem), and sending under your name reinforces that you’re the drafter. Apparent authority can bind you even if the AI blew past internal limits.
Contract law addresses a wrong or unintended term not through any general right to fix mistakes, but through the doctrines of unilateral mistake and mutual mistake. A mutual mistake, where both parties are wrong about a basic assumption that materially affects the exchange, can make a contract voidable by the disadvantaged party, unless that party bore the risk of the mistake. A unilateral mistake, where only one side is wrong, is far harder to escape: the mistaken party generally cannot avoid the contract unless the other party knew or had reason to know of the mistake (the “snap-up” limitation), the mistake makes enforcement unconscionable, or the other side hasn’t relied. Even then, the usual remedies are rescission or reformation, not a unilateral do-over. So when an AI sends the wrong contract, the company has no “error” defense to invoke; it must fit its situation into the mistake doctrine, almost always as a unilateral mistake, and, because the counterparty typically had no reason to know the terms were wrong, the realistic path is a timely revocation before acceptance rather than a hoped-for right to undo a deal already formed.
The five-guardrail playbook for AI agreement workflows
Put the legal guidance and the technology together, and a practical playbook emerges. We call it the AI Agent Agreement Authority Checklist:
1. Scope the authority. Decide, in writing, exactly which document types your agent may send, to whom, and up to what value. This mirrors the “delegation of authority” that Mayer Brown says belongs in your contracts, and it becomes your configuration spec.
2. Lock the documents. An agent should assemble agreements from approved, protected templates, never free-form. With BoldSign, template usage controls lock files, fields, recipients, and messages, so the agent can fill in the blanks but cannot rewrite the deal.
3. Add human checkpoints where stakes are high. Below a threshold, let the agent run. Above it, require a human approval before the document completes. BoldSign makes this a built-in workflow step through the Reviewer role in templates: a recipient who must review and approve the document, before or after the signer, without signing it themselves. Add a Reviewer to the template your agent uses, and every agent-sent agreement automatically passes through a human check as part of the signing order. And on the receiving end, the signer’s review-and-sign ceremony is itself a checkpoint, backed by signer identity verification when the agreement warrants it.
4. Watch everything in real time. A webhook is simply an automatic notification one system sends to another the moment something happens, so your software learns about events instantly instead of having to keep asking. In an agreement workflow, that means every send, view, sign, decline, and expiry can trigger an immediate alert to your own systems, so a misfire is caught in minutes, not at month-end. BoldSign supports webhooks at both the app level and the account level, and choosing the right scope for your setup is worth a few minutes of reading: see the guide to BoldSign webhook levels and how to choose between them. If something does go out wrong, the agent (or a human) can revoke the document immediately through the API before it’s signed.
5. Keep evidence by default. Every agreement should carry a complete audit trail and a cryptographic tamper-evident seal, automatically. When a dispute comes, you don’t reconstruct what happened. You hand over the record. This is exactly the “audit rights” the law firms are negotiating for, delivered as a product feature.
Run an agent inside these five guardrails and the liability question transforms. Yes, you stand behind your agent’s actions. But its actions are scoped, locked, supervised, observable, and provable. That’s a risk profile most legal teams can approve.
Try the pattern live
Our public demo shows the playbook in miniature. An AI warranty claim agent chats with a customer, assembles the warranty agreement from a template, sends it for signature through BoldSign MCP, tracks status in real time, and returns the completed, sealed document with its audit trail. (Because the demo is open for anyone to try, signer verification is switched off; in production you’d enable the layers your agreements warrant.)
Chat with the AI warranty claim agent at demos.boldsign.com.
Frequently asked questions
Is “the AI made a mistake” a valid legal defense?
Generally no. Under UETA, an electronic agent’s actions bind the business that deployed it, even if no person reviewed them. The agent is treated as your tool.
Can a recipient undo a contract an AI agent sent by mistake?
Sometimes. UETA’s error provision may let an individual avoid an automated transaction if the system gave them no opportunity to catch the mistake. A clear review-and-sign ceremony reduces this risk for both sides.
Does the AI vendor share liability?
Usually only to the extent your contract with them says so. Clifford Chance warns that standard technology agreements often leave the customer holding the risk, which is why firms like Mayer Brown recommend negotiating explicit authority definitions, guardrails, and audit rights.
How do I make AI agreement workflows safe to approve?
Scope the agent’s authority, lock it to protected templates, add human checkpoints above a value threshold, monitor every action through webhooks (automatic real-time notifications sent to your systems), and keep audit trails and tamper-evident seals on every document.
The bottom line
When an AI agent sends the wrong contract, the law points at the business that deployed it. That’s not a reason to keep agents away from agreements. It’s a design brief. Give your agent narrow authority, locked templates, human checkpoints, real-time visibility, and automatic evidence, and you get the speed of automation with a risk profile your legal team can sign off on.
That combination, agent autonomy on rails, is exactly what BoldSign was built to provide.
Start here:
- Explore BoldSign MCP
- Watch the AI warranty claim agent in action
- Create a free API sandbox and build your first guardrailed agent workflow.
This article provides general information about laws affecting automated transactions and is not legal advice. Consult qualified legal counsel for guidance on your specific situation.
Sources
- Uniform Electronic Transactions Act (1999), Uniform Law Commission, Sections 2(6), 9, 10 and 14. Section text as enacted in California, Civil Code sections 1633.1 to 1633.17
- New York Electronic Signatures and Records Act, State Technology Law Article 3, New York State Senate
- Regulation (EU) No 910/2014 (eIDAS), consolidated text, Articles 3 and 25. EUR-Lex
- Regulation (EU) 2024/1183, European Digital Identity Framework (eIDAS 2.0), in force 20 May 2024. EUR-Lex
- Stanford Law School CodeX, “From Fine Print to Machine Code: How AI Agents are Rewriting the Rules of Engagement”, Part 2, January 2025
- United States v. Heppner, No. 25-cr-00503 (S.D.N.Y.), Judge Rakoff, bench ruling 10 February 2026 and written opinion 17 February 2026. Case note, Harvard Law Review Blog
- Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, 14 February 2024. CanLII
- Contra proferentem and attorney-client privilege, Legal Information Institute, Cornell Law School
- BoldSign, Electronic signature legality by country, including the United States, European Union and Canada guides
