The 'my agent did it' defence: disputing commitments nobody authorised

A supplier receives an order from a company's purchasing agent, ships the goods, and is told the order was placed by an AI that misunderstood its instructions. Which body of law resolves that, and what evidence does it turn on?

Part of Agent tool-call gating

The doctrines, briefly

Agency law asks whether a principal is bound by acts done in its name. Three concepts do most of the work.

Each of those is normally established through evidence about what people intended, said and did. A human agent can testify about their instructions; the principal can testify about what they conferred.

What breaks when the agent is software

Three problems, and the third is the hardest.

1. No testimony. The agent cannot be examined about what it understood its authority to be. The nearest equivalent — logs and model outputs — is produced and retained by the party relying on it.
2. No stable scope. A human agent's authority is bounded by a role that changes slowly. An agent's effective authority is the union of its credentials, which changes whenever anyone updates a permission.
3. Apparent authority is nearly unbounded. A company that deploys an agent on its systems, using its credentials, on its domain, has held it out to the world. A counterparty transacting with it has few grounds to doubt authority.

The third point matters commercially. Companies imagining they can disclaim agent commitments are probably wrong about apparent authority, and the resulting exposure is larger than the one they are worried about.

Why the defence will be attempted anyway

Because the alternative is worse and the evidentiary position is soft. Faced with a commitment nobody intended, a company's options are to honour it, to dispute formation, or to argue the agent exceeded its authority.

The third is attractive precisely because it cannot easily be disproved. Nobody can produce a record showing which human authorised what scope, so the assertion sits in an evidentiary vacuum.

That vacuum is bad for everyone. Counterparties cannot rely on agent transactions, which slows adoption of a technology with real value; and companies that genuinely did authorise something cannot demonstrate it either.

What a delegation chain supplies

An authority record that exists before the dispute and does not depend on either party's logs.

human            [named officer, credential]
  └─ delegation  scope: purchase_order
                 counterparties: [approved supplier list]
                 max_value: 50,000 GBP
                 notAfter: 2026-07-31
                 depth: 0
       └─ agent   [action] purchase order to Acme Ltd, 12,400 GBP
                    └─ receipt verifies to the human's signature

A counterparty presented with that chain can verify it before performing. A commitment outside the bounds produces no valid chain, which puts the counterparty on notice rather than leaving them to discover the problem at payment time.

What this does to each doctrine

The middle row is the most consequential and the least obvious. Where a mechanism exists for verifying authority, a counterparty who does not use it is in a different position from one who had no means of checking.

What to do before this is litigated

1. Inventory the commitments your agents can make: orders, contracts, communications that could constitute an offer or acceptance, payments.
2. For each, establish who currently authorises the class of commitment and whether that authorisation is recorded anywhere.
3. Bound the agent: value ceilings, counterparty scope, expiry — enforced rather than described in a policy document.
4. Make the chain verifiable by counterparties for the commitment classes that matter. That is the step that changes the legal posture rather than merely the internal control.

This article describes doctrine at a general level and gives no legal advice. How agency principles apply to autonomous systems is unsettled and jurisdiction-specific, and that unsettledness is itself the argument for building the records now.

Why the defence will be attempted anyway

Because the evidentiary position is soft. Nobody can produce a record showing which human authorised what scope, so the assertion sits in a vacuum. That vacuum is bad for everyone: counterparties cannot rely on agent transactions, and companies that genuinely did authorise something cannot demonstrate it either.

Objections and honest limits

“Our terms disclaim agent commitments.” Disclaimers rarely cure a representation that induced a transaction, and apparent authority is about how you held the agent out rather than what your terms say.

“This is speculative until courts rule.” The doctrines are old and settled; their application to software is not. That unsettledness is the argument for building the records now, because they cannot be created retrospectively.

Before this is litigated

1. Inventory the commitments your agents can make. Orders, contracts, communications that could be an offer or acceptance, payments.
2. Establish who authorises each class today. And whether it is recorded anywhere.
3. Bound the agent at the credential. Value ceilings, counterparty scope, expiry — enforced, not described.
4. Make the chain verifiable by counterparties. That is the step that changes the legal posture.

Terms used here

Actual authority

Authority the principal conferred, expressly or by implication.

Apparent authority

Authority a third party reasonably believed existed because of how the principal held the agent out.

Ratification

Adopting an unauthorised act after the fact, including by retaining its benefit.

Frequently asked questions

Is this legal advice? No. It describes agency doctrine at a general level. How those doctrines apply to autonomous software is unsettled and varies by jurisdiction; take it to counsel.

Can a company disclaim its agent's commitments? Apparent authority makes that harder than companies expect. Deploying an agent on your systems with your credentials is holding it out, and counterparties may reasonably rely on that.

Does a delegation chain bind a counterparty? It gives them a means of verifying authority before performing. Whether a court treats failure to verify as affecting reasonable reliance is exactly the kind of question that will be litigated.

What if the agent acted within scope but produced a bad outcome? Then authority is not the issue and the dispute is about performance or mistake, which is ordinary contract territory. The chain removes the authority question from the argument.

Can a business disclaim its agent's commitments? Apparent authority makes that harder than expected. Deploying an agent on your systems with your credentials is holding it out.

Sources and further reading

• Restatement (Third) of Agency — overview (Cornell LII, Agency)
• Uniform Electronic Transactions Act (ULC)
• Published commentary on contract formation by autonomous systems.
• Reported disputes involving automated ordering and commitment systems.