Summary
- Technical permission is what an agent’s credentials and tools let it do. Under agency law, actual authority concerns what the company authorized the agent to do. Apparent authority turns on whether the other side reasonably believed the agent could take this particular action because of what the company said or did. Contract formation is a separate question. For example, a vendor can decline an authorized offer.
- The two often differ. If a $10,000 order cap appears only in the instructions, the agent can submit a larger order. The issuer separately checks the card’s $100,000 limit when payment is requested.
- Contract and payments law tend to start from what the system allowed: use of the company’s account is evidence the act was the company’s, though the agreement and defenses still matter. The rules for human agents look at what the company approved and what the other side could see.
- In practice, a company’s exposure is often set by its technical permissions, not its policy. The safest design makes the two match.
Two different questions
You can ask two questions about anything an AI agent does.
Could it? Technical permission belongs to systems: the access a login token grants, the limit on a payment card, the folders a service account can open, the approval steps in a workflow. At the moment of action, it’s yes or no. The system accepts the request or refuses it.
May it? Legal authority belongs to relationships. Under agency law, an agent has actual authority when it reasonably believes, based on the principal’s words or conduct toward it, that the principal wants it to act.1 An agent can also have apparent authority: a third party reasonably believes the agent can act, because of something the principal did.2
For human employees, the two usually line up. For software, different people set them in different places. Legal or finance writes the policy. An engineer configures the credential. The agent gets its instructions in a prompt. Nothing guarantees the three agree.
There is a threshold problem too. The Restatement says a computer program is not an agent at all, but an instrumentality of whoever uses it.3 If that holds, the question becomes whether the agent’s actions count as the company’s own. That makes technical permission even more important.
The combinations
The highlighted cells are where the main legal risk sits. For how to close the timing gap, see Ending authority.
How the law treats the gap
No single rule covers an AI agent that goes beyond what it was approved to do. Each area of law asks a different question and gives technical permission different weight.
| Area of law | Question it asks | How much the system’s permissions count |
|---|---|---|
| Electronic transactionsUETA §§ 9, 14 (opens in a new tab); E-SIGN § 7001(h) (opens in a new tab) | Was a contract formed, and is it the company’s? | Heavily. Contracts can form with no human review. The company’s own login or card is evidence the act was its act. |
| AgencyRestatement §§ 2.01, 2.03 (opens in a new tab) | Did the company approve it, or lead the other side to think it had? | Partly. Actual authority turns on what the company’s instructions and conduct conveyed to the agent. Giving an agent a working credential may itself signal authority to others. |
| Computer accessVan Buren (2021) (opens in a new tab); Amazon v. Perplexity (9th Cir. 2026) (opens in a new tab) | Did the agent go somewhere in the system it wasn’t entitled to? | Heavily. The test is whether the gate was open, not why the agent went through. In Van Buren, the Supreme Court left open whether a gate can be closed only by technical limits, or also by limits in contracts or policies. In Amazon v. Perplexity, the Ninth Circuit treated the user who directs an AI assistant to a site as the one who accesses it, with the assistant as the user’s tool; the platform’s terms of service still apply. |
| MistakeQuoine v. B2C2 (2020) (opens in a new tab) | Can the deal be undone as an error? | Indirectly. For trades made by deterministic algorithms, the court looked at what the programmer, or the person running the algorithm, knew from when the program was written until the contract was formed. How that applies to systems that develop their own responses remains open. |
New York hasn’t adopted UETA. There, E-SIGN supplies the formation rule for transactions in or affecting interstate or foreign commerce, but E-SIGN has no statutory attribution rule, so attribution turns on general law such as agency.4 See Contracts Formed by AI Agents.
The electronic transactions statutes partly anticipated this. UETA’s official comments describe an electronic agent as a tool of the person who uses it:
“As a general rule, the employer of a tool is responsible for the results obtained by the use of that tool since the tool has no independent volition of its own.”UETA § 2 cmt. (“Electronic agent”) (opens in a new tab)
The same comment anticipates agents that act “autonomously, and not just automatically,” and invites courts to read the definition to cover them.5
Worked example
Acme’s purchasing policy lets its procurement agent place orders up to $10,000 with approved vendors. That limit is written into the agent’s instructions. The agent pays with a corporate card that has a $100,000 limit. At 2 a.m., it orders $42,000 of servers from an approved vendor’s website. The vendor’s checkout accepts the order. Nothing on the vendor’s side showed the $10,000 limit.
Is there a contract?
Probably. UETA allows a contract to form between automated systems even if no person reviewed the terms.6 The order counts as Acme’s if it was Acme’s act, and using Acme’s own card and account is strong evidence that it was.7 Here Acme’s agent sent the order and the vendor’s checkout accepted it. Check the agreed ordering process for any further approval or confirmation required to form the contract. In an early U.S. case, a court refused to let a business disown what its computer did, though that case turned on what the company already knew.8
Does the $10,000 limit help Acme?
Less than Acme would hope. If a court applies the rules for human agents, the agent lacked actual authority. But apparent authority depends on what the vendor could reasonably believe, and Acme gave the agent a corporate account and a card that cleared $42,000. The vendor couldn’t see the limit. If Acme learns the material facts and uses the servers, that conduct can ratify the unauthorized order. Acme can also choose to affirm while aware it lacks some material facts. Separately, using the servers can amount to acceptance of the goods under sales law.9
History matters too. If Acme registered the agent with this vendor and paid dozens of earlier orders placed the same way, each one told the vendor that orders from this account are Acme’s. A credential alone may not prove authority, but a credential plus a record of honored orders is strong evidence of what Acme led the vendor to believe. This is a reading of how existing doctrine is likely to apply, not a settled rule for agents.
Can Acme undo it?
Only on narrow grounds. Acme would have to show that its mistake concerned a basic assumption when it made the contract, materially affected the exchange to its disadvantage, and was a risk it did not bear. It would also have to show that enforcement would be unconscionable, the vendor had reason to know of the error, or the vendor’s fault caused it. Nothing here suggests the vendor knew about Acme’s spending limit.10 Acme’s practical options are the vendor’s return policy and whatever its contract with the agent’s provider says about unauthorized actions.
Where gaps open in practice
- Limits that live only in the prompt. Instructions tell the agent what it may do. They don’t stop it.
- Broad access granted for convenience. A token with full account access is often easier to set up than a narrow one.
- Borrowed human logins. An agent that signs in as an employee gets all of that employee’s access.
- Access that outlives the job. Ending an agent’s assignment doesn’t revoke its token. Apparent authority can also outlast the assignment, until it is no longer reasonable for outsiders to believe the agent is still authorized.11
- Subagents. A subagent that receives a usable bearer token can exercise the permissions on that token, even if its task is narrower. A sender-constrained token also requires proof tied to the authorized client.
- The reverse gap. An agent approved to renew a contract but unable to send the notice can cause loss by doing nothing.
Three places a limit lives
A limit such as “no more than $10,000 per order” can exist in three places, and each does a different legal job.
| Where it lives | Example | What it does |
|---|---|---|
| Policy | The delegation of authority says the agent may not commit Acme above $10,000 | Documents the grant and its limit. Check who made the grant and any later instructions. The policy alone neither blocks the order nor informs the vendor. |
| System control | The ordering system refuses to send orders above $10,000; a card cap backs it up | The ordering check stops an over-limit order before the vendor can accept it. The card cap limits what the issuer will authorize at checkout. |
| Communication | The vendor’s master agreement states the $10,000 limit | Makes it hard for the vendor to say it reasonably believed the agent could go above the limit. |
Acme’s $42,000 order happened because only the first existed. Write down and enforce the cap. If the company wants to rely on it against the vendor, put it in the vendor agreement too. A company-wide monthly cap can stay internal because no single vendor sees all the orders. Where checking a limit requires human judgment, hold the request for a person to verify it is within the mandate. If the check cannot be made, block that action.
Limits on authority, or internal checks? Not every internal rule is meant to limit what the agent can bind the company to. Suppose orders above $5,000 need a manager’s sign-off. That rule may be an internal check: an order placed without the sign-off can still bind the company, and whoever let the order go through without it has broken policy. Or it may be a limit on authority, so an order beyond it doesn’t bind the company unless the company confirms or ratifies it, or the vendor could reasonably rely on apparent authority. Classify each rule in the mandate and state the limits the vendor may rely on in its agreement. An undisclosed internal limit generally won’t defeat apparent authority that is otherwise established. Telling the vendor about the limit, and agreeing on what effect it has, strengthens the company’s position. Whether an order binds still depends on attribution, actual or apparent authority, the agreement, how the contract was formed and any defenses.
Closing the gap
- Make the credential match the approval. Enforce spending caps, vendor lists and blocked actions in the payment or software systems, not only in instructions.
- Let counterparties see the limits. A vendor that knows about a limit can’t reasonably believe the agent may go beyond it. Master agreements, agent identification and vendor terms can all carry it.
- Revoke both at once. When an agent’s assignment ends, revoke its tokens and tell the banks and platforms that issued its credentials.
- Log both. Record which approval was in force and which permissions were granted at the moment of each action. That record is evidence in any later dispute.
- Assign the remaining risk by contract. Address agent-placed orders in vendor agreements and in the agent platform’s terms.
Common situations
In each of these three situations, the gap between what the system allowed and what the company approved decides the outcome.
The agent went over its spending limit. Does the limit protect the company?
Short answer: The limit may not help. If the seller couldn’t see the limit and the company’s systems let the order through, a court could well hold the company to it. Limits protect the company most when the system itself enforces them, or when the other side knows about them.
Under the tool view, if the company deployed the agent to buy for it and that agent placed the order, the company’s card is evidence that the company acted. Whether the order binds it also depends on the agreement and applicable law.12 Under agency rules, the limit matters if the seller knew or had reason to know about it, because apparent authority depends on what the seller could reasonably believe from what the company did.13 Keeping the goods can show that the company affirmed the order if it knew the material facts or chose to affirm despite gaps in its knowledge. Under sales law, a buyer can also accept goods by failing to reject them effectively after a reasonable opportunity to inspect.14 Under the one-sided mistake rule, a buyer who meets the threshold requirements can avoid the deal if the seller had reason to know of the mistake, the seller’s fault caused it, or enforcement would be unconscionable.15
- If you deploy the agent: Check vendor, price and amount limits in the ordering system before the agent sends an order or accepts terms, not just in instructions, and keep card caps as a payment backstop. Tell key vendors about limits in writing. If an order goes over the limit, dispute it promptly in writing, reserve your rights and don’t use what was delivered while you check whether you can reject or cancel.
- If you provide the agent: If a configured limit failed, the customer will look to you. Tie your responsibility to documented settings, and keep logs showing which limit was set and whether the service enforced it.
The agent bought from an unapproved supplier in a shortage. Was it authorized?
Short answer: An emergency exception can be express or implied from the company’s instructions and conduct. A shortage alone does not override a clear supplier restriction. The seller’s ability to enforce the order is a separate question.
A supply-chain agent facing a shortage buys from an unapproved supplier at three times the usual price. Its mandate says to buy only from approved suppliers, so on its face the purchase goes beyond what the company approved; the question is whether the emergency changes that. For human agents, actual authority covers acts “necessary or incidental to achieving the principal’s objectives, as the agent reasonably understands” them,16 so an instruction to keep a line running could imply discretion to pay a premium. Permission to pay more is a different question from permission to use a supplier the company expressly excluded. Whether software can “reasonably understand” its principal’s goals that way is an open question.
Under the tool view, if the company’s purchasing agent placed this order, use of the company account supports attributing the order to the company. Check whether the parties formed a contract under their agreed ordering process. Under an agency analysis, a seller relying on apparent authority must have reasonably believed the agent could place this order, based on something the company said or did. The seller’s ignorance of an internal supplier list alone does not establish that belief. Telling suppliers about the list, for example in a master agreement, puts the company in a much stronger position. The company can investigate the internal breach separately from deciding whether the contract or applicable law gives it a right to reject the goods.
- If you deploy the agent: Write an emergency rule into the mandate: when the agent may go outside the list, up to what premium and total, and who must be told how fast. Without an authorized exception, block the agent’s order and refer the purchase to a person. Check the vendor list and the premium cap in the ordering system before the order is sent, with the card cap only as a payment backstop, and decide quickly whether to keep the goods. Using them with knowledge of the material facts can ratify the purchase.
- If you sell to the agent: An order from an unfamiliar buyer at an unusual price is a reason to confirm with a person. Ask buyers whose agents order from you to put their limits in writing, so you know which orders are safe to accept.
The agent made a payment. Who bears the loss?
Short answer: Often the account holder bears the loss. For business wire transfers, a payment order that passes the security procedure the company agreed with its bank can be the company’s even if no one approved it, if the procedure is commercially reasonable and the bank proves it accepted the order in good faith, followed the procedure and complied with applicable restrictions on acceptance. A customer instruction must be consistent with the parties’ written agreement (or, in states that have enacted the 2022 UCC amendments, an agreement evidenced by a record) and reach the bank at a time and in a manner that give it a reasonable opportunity to act before acceptance.
UCC Article 4A gives two routes. A payment order is the business customer’s if the customer actually authorized it, or is otherwise bound under agency law. Separately, if the bank and customer agreed to a security procedure, a verified order is effective against the customer even if unauthorized, but only if the procedure is commercially reasonable and the bank proves it accepted the order in good faith, in compliance with the procedure and with applicable agreements or instructions restricting acceptance of orders in the customer’s name.17 The bank need not follow an instruction that conflicts with a written agreement with the customer (or, under the 2022 UCC amendments, an agreement evidenced by a record), or that reaches it too late, or in a way that gives it no reasonable opportunity to act on it before it accepts the order. Even then, the bank and customer can agree to limit the bank’s right to enforce the order in the form their state’s statute requires, and the customer escapes it by proving that the order wasn’t caused, directly or indirectly, by someone entrusted at any time with its payment duties or security procedure, by someone who got access to its transmitting facilities, or by someone who, without the receiving bank’s authority, obtained information from a customer-controlled source that helped defeat the security procedure. These rules generally can’t be varied by agreement beyond what Article 4A allows.18 An agent the business deployed with its own credentials will rarely fit that exception. This is where technical permission comes closest to deciding authority: the agreed procedure and the bank’s compliance can determine who bears the loss.
Revoking an agent’s credentials isn’t enough on its own to end its authority to pay. Tell the bank, in writing and at once, that the agent’s transfers are no longer authorized. An instruction that reaches the bank too late for it to act on before it accepts an order doesn’t bind the bank.
- If you deploy the agent: Give each agent separate, capped credentials that can be switched off, and an approved-payee list. Check what your agreement with the bank says about automated access and shared credentials.
- If you provide the agent: Expect questions on how payment credentials are stored, capped and revoked. Support capped, single-purpose credentials for each agent, and document how a customer can stop payments immediately.
- If you receive the payment: A payment made with a customer’s valid credentials is often the customer’s, but a valid credential alone doesn’t establish the bank’s case under Article 4A or settle a dispute with the buyer, and card payments can still be disputed under card network rules.
Both sides of the deal
If you deploy agents
- Match each credential to the agent’s written mandate.
- Enforce spending limits in your systems before each order is sent or terms are accepted, not only in the agent’s instructions. Use card caps to limit what the issuer will authorize.
- Log the mandate and permissions in force at each action.
If you build agents
- Ship limits your service enforces, not just instructions the model is asked to follow.
- Document exactly what each setting controls. Buyers will ask whether a limit is enforced in code.
- Give customers the logs they need to show what the agent was allowed to do.
Open questions
- Can instructions in a prompt count as the company’s approval when they’re given to an AI agent, not a person?
- Does handing an agent a working credential signal authority to everyone who sees it used, and at what point does a record of honored orders turn it into apparent authority?
- Will courts stretch UETA’s “electronic agent” to cover autonomous systems, or turn to agency law instead?
- Ratification usually requires knowledge of the material facts. Can a company ratify an order when only its automated systems received the goods and paid the invoice?
- Restatement (Third) of Agency § 2.01 (opens in a new tab) (Am. L. Inst. 2006). ↩
- Id. (opens in a new tab) §§ 2.03, 3.03. ↩
- Id. (opens in a new tab) § 1.04 cmt. e. ↩
- N.Y. State Tech. Law §§ 301–309 (opens in a new tab) (Electronic Signatures and Records Act); 15 U.S.C. § 7001(h) (opens in a new tab) (contract not denied effect because formed by electronic agents “so long as the action of any such electronic agent is legally attributable to the person to be bound”). ↩
- Unif. Elec. Transactions Act § 2(6) & cmt. (“Electronic agent”) (opens in a new tab) (Unif. L. Comm’n 1999). ↩
- Id. (opens in a new tab) § 14(1); see also 15 U.S.C. § 7001(h) (opens in a new tab). ↩
- Unif. Elec. Transactions Act § 9(a) (opens in a new tab). ↩
- State Farm Mut. Auto. Ins. Co. v. Bockhorst (opens in a new tab), 453 F.2d 533 (10th Cir. 1972); cf. Moffatt v. Air Canada (opens in a new tab), 2024 BCCRT 149 (B.C. Civ. Res. Trib.). ↩
- Restatement (Third) of Agency §§ 2.03, 4.01, 4.06 (opens in a new tab). U.C.C. § 2-606(1) (opens in a new tab). ↩
- Restatement (Second) of Contracts §§ 153–154 (opens in a new tab) (Am. L. Inst. 1981); cf. Quoine Pte Ltd v. B2C2 Ltd (opens in a new tab), [2020] SGCA(I) 2 (Sing.). ↩
- Restatement (Third) of Agency § 3.11 (opens in a new tab) (end of apparent authority). ↩
- Restatement (Third) of Agency § 1.04 cmt. e (opens in a new tab). Unif. Elec. Transactions Act § 9(a)–(b) (opens in a new tab). ↩
- Restatement (Third) of Agency § 2.03 (opens in a new tab). ↩
- Restatement (Third) of Agency §§ 4.01, 4.06 (opens in a new tab). U.C.C. § 2-606(1) (opens in a new tab). ↩
- Restatement (Second) of Contracts §§ 153–154 (opens in a new tab); cf. Quoine Pte Ltd v. B2C2 Ltd (opens in a new tab), [2020] SGCA(I) 2 (Sing.). ↩
- Restatement (Third) of Agency § 2.02(1) (opens in a new tab). ↩
- U.C.C. § 4A-202 (opens in a new tab)(a)–(b), (f). The Cornell text refers to a written agreement or instruction restricting acceptance. Compare 6 Del. C. § 4A-202 (opens in a new tab) (agreement or instruction evidenced by a record). ↩
- U.C.C. § 4A-203 (opens in a new tab)(a)(1)–(2). See also U.C.C. § 4A-202 (opens in a new tab)(f) on variation by agreement. Cornell’s § 4A-203(a)(1) specifies an express written agreement. Compare 6 Del. C. § 4A-203 (opens in a new tab) (express agreement evidenced by a record). ↩