Agent Authority
Home/Guides/Practice

Building Limits into Agent Systems

Practice · Guide 7Last reviewed: November 2026About 14 minutes

Summary

  • A company’s legal exposure from an AI agent depends heavily on what its systems allow and what counterparties are told, not only on what its policies say. Limits that matter have to be built in.
  • Think of authority as a stack, from board approval down to the agent’s credentials and logs. Each layer should match the one above it.
  • Five controls do most of the work: separate agent identities, narrow credentials, limit checks before the agent commits (backed by payment caps), human approval above set thresholds, and complete logs.
  • Ending an agent’s authority means revoking its access and telling the banks and vendors that relied on it.

The authority stack

Authority flows down through several layers before an agent acts. A gap at any layer is where disputes start.

  1. Corporate authorityBoard or officer approval for the kind of spending or commitmentLegal, finance
  2. PolicyProcurement, payment and contracting policiesLegal, finance
  3. Agent mandateThis agent’s written task, limits and prohibited actionsBusiness owner, legal
  4. Identity and credentialsWhich accounts, tokens and cards the agent holds, and their scopeEngineering, security
  5. Transaction limitsCaps, vendor lists and blocked actions enforced by the systemsEngineering, finance
  6. Approval gatesActions that pause for a person’s approvalEngineering, business owner
  7. ExecutionThe agent actsThe agent
  8. LogsWhat was allowed, what was done, and by whomEngineering
  9. ReviewAccept or reject what the agent did, promptlyBusiness owner

The highlighted layers are where technical permission is set. They should match the mandate above them. For where each sits inside the software, see Inside an Agent System.

Layer by layer

Mandate: write it down

Each agent should have a short written mandate: its task, its limits, what it must never do, and when it must ask a person. This is the document a court would look to for what the company authorized the agent to do, and the source for every technical setting below.1 Version it, so you can show which mandate was in force at any moment. The Agent Mandate template sets one out line by line.

Identity: give each agent its own

Don’t let agents sign in as employees. An agent using a person’s login inherits all of that person’s access, and its actions are hard to tell apart from theirs. A separate identity for each agent makes attribution clearer, though on its own it doesn’t confine the agent to the company’s account: bind each identity and its grants to the company’s account or tenant (the company’s own space within a shared service) at each service. Under UETA, a record is attributable to a person if it was that person’s act, including an act done through its electronic agent, and that can be shown in any manner, including by the security procedure used. When an agent signs in as an employee, the login points to the employee, not the agent, which makes it harder to show who a record is attributable to.2

Credentials: as narrow as the task

Grant only the access the mandate requires: read-only where the agent only needs to read, specific folders rather than whole drives, one vendor account rather than all of them. Broad tokens can create a large gap between what an agent can do and what it’s approved to do. See Technical Permission vs. Legal Authority.

Transaction limits: check before the agent commits

A limit in the agent’s instructions is advice to the agent. The issuer can enforce a payment limit when it receives an authorization request. Check each order against the limits in the orchestrator or ordering tool before the agent sends an order or accepts terms. A card cap is a payment control, not a contract control: a vendor can accept an order before it charges the card, and a declined charge can leave an unpaid order the vendor says is binding. Use the card as a backstop, with single-purpose or virtual cards, per-transaction and monthly caps, merchant restrictions and expiry dates. Purchases on account or invoice, subscriptions and renewals often involve no charge at the moment of commitment, so for those the pre-commitment check is the only limit that applies before the company is bound. For purchasing, restrict orders to approved vendors, ideally ones under a master agreement that states your agents’ limits.

A per-order cap on its own can be beaten by splitting. Ten $9,900 orders to one vendor in a minute each pass a $10,000 cap. Add caps per period and per counterparty, and require confirmation for orders far outside the agent’s normal pattern, such as several times its average order. A useful exercise: list every dimension the mandate limits (amount, period, counterparty, category, geography, term length) and ask engineering which of them the system actually enforces. Use a mandatory review for actions the mandate allows with approval. Block prohibited actions before submission. See Three places a limit lives.

Separation of duties: split the powers

Companies don’t let one employee add a new vendor to the payment system, approve it, order from it and pay it. An agent can hold all four powers unless you split them. Give a purchasing agent ordering access but not independent payment powers or vendor record access. If it settles its own orders with a capped, vendor-restricted card at an approved checkout, say so expressly as an exception; approving invoices, initiating transfers and releasing payments stay with a person or a separately authorized payment service with its own checks. Never let an agent change a vendor’s bank details or payment instructions, or approve its own orders; send those to a person. Changed payment details are a common fraud route, and agents read the emails that carry them.

Approval gates: set thresholds, then honor them

Require a person’s approval above set amounts, for new counterparties, and for listed terms such as auto-renewal or arbitration. The gate can release an otherwise permitted action once someone with the required authority approves the specific transaction. A reviewer’s approval leaves the agent’s hard limits in force. It only helps if reviewers actually review: an approval clicked through in seconds is evidence of authority, not a safeguard.

Know which of three setups you have, because products call all of them “agents.” In an advisory setup, the agent recommends and a person acts, so the person’s own authority governs. In an approval-gated setup, the agent proposes, a person approves and the system executes, so the approval is part of the authority. In an autonomous setup, nobody approves in between, so authority, attribution and ratification are all in play.

For each approval step, say what approving authorizes: sending a message, accepting terms, signing or paying. Record what the approver saw and the approver’s own authority, since an approval is only as good as the approver’s power to give it. Two failures are common. A notice sent to someone after the agent acts is not approval. And if the approval service is down, the agent should stop and wait, not proceed.

Logs: record both sides

For each action, record the mandate version, the instructions in force, the input details needed to explain the decision, the identifiers of the credentials used, the counterparty, the output, and any approval. Record a credential by an identifier or fingerprint, its issuer and its scopes; never log live tokens, API keys, passwords or full card numbers, because a log that holds them becomes a way in. Keep transaction evidence tamper-evident and available for disputes. Set a separate retention period for raw inputs. Redact or remove personal details when the need ends, and preserve evidence under a legal hold. If a provider holds them, secure access by contract. In a dispute, the logs are the evidence of what was authorized and what happened. Where a law requires a contract or record to be in writing, E-SIGN lets an electronic record of that contract or record be denied legal effect if it can’t be kept and accurately reproduced for everyone entitled to retain it—one more reason to keep transaction records and logs that hold them in a form you can produce.3

Subagents: derived, narrower access

If an agent hands work to other agents, give each subagent its own credentials, limited to its piece of the task. Never pass on the parent’s full access. Record who delegated what. Agency law asks whether the first agent was allowed to appoint a subagent, so the mandate should say.4 The total across all subagents must also stay within the parent’s limits. An agent with a $10,000 limit that starts five subagents, each with its own $10,000 card, has created $50,000 of spending power that nobody approved.

Review: accept or reject quickly

Someone should review agent activity on a regular schedule, but periodic review is oversight; it won’t catch urgent problems in time. Actions outside the mandate and confirmations from counterparties need attention the same business day: contract rejection and dispute windows can be one or two business days, and silence past them can bind the company. Send counterparty notices to a monitored inbox with a named owner and backup. If the agent did something outside its mandate, give written notice before the deadline disputing it and reserving the company’s rights, then exercise any rejection or cancellation right before it lapses, and don’t use what the agent obtained while you decide. If the deal turns out to bind the company, refusing to perform because it broke an internal limit can itself be a breach. Keeping the benefit of an unauthorized deal can amount to ratification. A company can affirm with knowledge of the material facts or by choosing to affirm despite knowing facts are missing.5 Whether automated receipt and payment alone can show that choice before anyone reviews the order is unsettled. Don’t rely on them not counting: route orders outside the mandate to a person before receipt and payment run automatically.

Ending authority: revoke and notify

When an agent is retired or its mandate changes, revoke its tokens and cards, and tell the banks and vendors that dealt with it. An unauthorized payment order can bind the company under UCC Article 4A if the agreed security procedure was commercially reasonable and the bank proves it accepted the order in good faith, followed the procedure and complied with applicable restrictions. Tell the bank promptly and replace credentials the agent could still use.6 Under agency law, apparent authority can last until it is no longer reasonable for third parties to believe the agent still has authority, which in practice usually means until they are told.7

Revocation runs on several clocks at once, and they rarely line up. Suppose Acme’s board withdraws its procurement agent’s authority at 9:00. The policy is updated at 9:05, IT gets the ticket at 9:30 and the agent’s token is revoked at 11:00. At 10:15 the agent places an order, and nobody has told the vendor. Actual authority may have ended at 9:00. The token still worked at 10:15 and could keep working after 11:00 until the vendor enforced revocation or the token expired. Apparent authority may last until the vendor has reason to know of the change. Five decisions close most of that gap:

  • Name the event that ends authority. A contract can say authority ends when a notice is received, when the credential is disabled or at a stated time. Pick one and write it down before you need it. The Agent Transaction Addendum uses receipt of notice plus a fixed delay.
  • Make the technical step fast. Ask how long revocation takes to reach every system, including cached sessions and long-lived keys. A token that stays valid for eight hours after the decision leaves eight hours of access, and exposure, that nobody approved.
  • Decide what happens to work in progress. List what might be pending when the agent stops: orders sent but not confirmed, offers the counterparty hasn’t accepted, scheduled payments. Say which to cancel, which to complete, and who tells the counterparty.
  • Revoke down the chain. Credentials a subagent received from the parent should end with the parent’s, automatically.
  • Decide where the agent’s authority comes from. State in the mandate whether the company granted it directly or whether it rests on the authority of the person who set it up. If the company granted it, the approver’s departure alone does not end the grant. Acme’s policy still pauses the agent until an authorized successor approves the mandate again. If the grant depends on the departing person’s delegation, check whether that delegation has ended before allowing another action. Either way, a departure should trigger a review of every agent the person set up or was responsible for.

Control map

ControlLegal purposeOwner
Written, versioned mandateRecords what the company authorized the agent to doBusiness owner, legal
Separate agent identityHelps distinguish requests made with the agent’s credentials from employee requestsSecurity
Narrow credentialsEach receiving service checks the grant and rejects requests outside it, limiting access to what was approvedEngineering
Limit checks before commitment, plus payment capsStops orders outside the mandate before they are sent; caps card authorizations as a payment backstop, though later captures and fees can still exceed a capFinance, engineering
Approved vendors under contractGives vendors notice of the limits stated in their agreementsProcurement, legal
Approval thresholdsRecords who approved each high-risk action and what the approval coveredBusiness owner
Complete, retained logsEvidence of authorization and conductEngineering
Prompt review and written disputeDocuments the objection and helps preserve timely rejection rightsBusiness owner
Revocation and noticeEnds actual and apparent authority; puts banks and card issuers on noticeSecurity, finance

Acme, rebuilt

Return to the Acme example: a $10,000 policy, a $100,000 card and a $42,000 order. With the controls above:

  • The procurement agent has its own identity and a virtual card capped at $10,000 per transaction, restricted to approved vendors.
  • The ordering tool checks every order against the mandate before it is sent.
  • Otherwise permitted orders above $5,000 and up to $10,000 pause for an authorized buyer’s approval. The first otherwise permitted order with a vendor or under the Addendum also pauses, whatever its amount. Above $10,000 the ordering tool rejects the request.
  • The vendor’s master agreement names the agent and states its limits.
  • The ordering tool blocks the $42,000 order before it reaches the vendor, so there is no order for the vendor to accept. Had that check failed, the card issuer would ordinarily decline the authorization, but a declined authorization alone doesn’t undo an order the vendor has already accepted. The log records the attempted order, the mandate version in force and the cap that blocked it, and the buyer is alerted.

The same facts that produced a dispute now produce a log entry.

Launch checklist

Before an agent goes live, confirm:

  1. The written mandate records the approval of a person whose authority covers it and any additional approval required by policy. The Agent Mandate is a template. Schedule A to the Agent Transaction Addendum shares its limits with counterparties.
  2. The agent has its own identity and credentials.
  3. Each credential’s scope matches the mandate.
  4. Per-order, per-period and per-vendor limits are checked before the agent sends an order or accepts terms, against a running total of all its commitments, including orders on invoice. Card caps back this up but don’t replace it.
  5. Approval thresholds are set, and the reviewers know their role.
  6. Logs document the mandate version, relevant input details, actions and approvals. Retention rules cover transaction evidence, personal data and legal holds.
  7. Key counterparties have agreements that address agent activity.
  8. There’s a tested way to shut the agent off and revoke its access.
  9. You’ve walked through the internal interview script with the team that built it.
  10. A named owner and backup triage alerts and counterparty confirmations from a monitored inbox, faster than the shortest rejection or dispute deadline in your agreements, and know to dispute unauthorized deals in writing, and exercise any rejection or cancellation right, before the deadline passes.
For legal: the checklist is also a record. Keeping a completed copy for each agent documents the review and can support an argument that the company acted reasonably, which matters if a negligence claim follows an agent’s error. Whether the controls were reasonable, and whether they actually worked, remain questions of fact.

Both sides of the deal

If you deploy agents

  • Complete the launch checklist for each agent and keep a copy.
  • Assign someone to review activity and dispute unauthorized deals promptly in writing.

If you build agents

  • Expect enterprise buyers to ask for every control on the checklist. Map each one to a product feature.
  • Don’t claim a control you can’t demonstrate in the logs.
  • Offer separate identities and scoped credentials per agent so customers can meet their own obligations.

Open questions

  • Will courts treat a company’s failure to use available technical limits as negligence?
  • How long must logs be kept, and will courts expect specific content in them?
  • Can a company ratify an agent’s unauthorized order when its own systems receive the goods and pay the invoice before any person sees it?