An Authority Review, Step by Step
A review lines up three things that are easy to treat as one: what the company authorized its agent to do, what the software can actually do, and what the companies it deals with are led to believe. Disputes happen where they diverge. The review leaves a record of all three. Pick the part that fits where you are.
Before an agent goes live
Five steps take one agent from “we want to turn this on” to a signed review file. Steps 1 and 3 can often be done by one or two people; steps 2, 4 and 5 usually need engineering in the room. Each step says who should be involved, where to record the answers and what you leave with.
1. Scope the agent
- Gather what you have. You’ll need whatever describes the agent’s job now (its instructions, a mandate, or the purchasing or approval policy it works under), the provider agreement if you didn’t build it, the agreements with the main vendors or customers it will deal with, and the company’s delegation of authority policy.
- List the acts with legal consequences. These include orders, acceptance of terms, payments, subscriptions, signatures, statements to customers, account changes and messages that make commitments. Leave out work that commits the company to nothing by itself, such as drafting for a person to review or summarizing documents.
- Set the authority level. The scale runs from 0 (recommend) to 4 (execute within limits). Note which acts can’t be undone.
- Name the principal and the other roles. Record which legal entity the agent acts for, and who else is involved: the provider, the person who configures it, the counterparties. For regulation, also note who is the provider and who is the deployer, which may differ.
- Trace the approval. Who approved the deployment, under which delegation, and does that person’s own authority cover everything the agent will do? Identify any second approval required by the delegation policy, including for card payments or signatures.
Background: Tool or Agent? (authority levels), Corporate Authority for Agents and The Regulatory Overlay.
2. Find out what it can actually do
- Start with the business owner. Engineers know what the system can do, but often not what vendors and customers have been told about it.
- Ask what counterparties see. Find out what name the agent uses, what its messages say, which account it uses, and whether earlier orders placed the same way were paid. This is where apparent authority comes from.
- Map the system. Record the agent’s identity, its credentials and their scopes, its tools, the payment instruments it can use, what outside systems will accept from it, whether it can start other agents, what outside content reaches the model through retrieval or tool results, and what it keeps in memory.
- Ask about commitments that aren’t payments. Can it click through terms, order on credit or on account, start or upgrade a subscription, or promise something to a customer?
- If a provider supplies the agent, send it the Agent Diligence Questions, or ask for its Agent Authority Profile. Record its answers alongside the internal interview answers.
Background: Inside an Agent System and Identity, Credentials & Standards.
3. Write down the authority and find the gaps
- Fill in the mandate. Cover the amount per order, totals per period and per counterparty, categories, counterparties, geography, terms it may accept, approvals, subagents and duration. Say which limits are limits on authority and which are internal checks.
- For each line, name what enforces it. A limit that exists only in the agent’s instructions or in policy is a gap: the agent can do more than the company approved. Compare the approved commitment limit with the lowest limit the systems check before a commitment is sent, whether in the software or an approval step. Check payment controls separately. A card cap limits what the issuer will authorize. A charge can still post after a decline, and the company may already have committed to the order. See Three places a limit lives.
- Check the commitments that aren’t payments. For terms, credit, subscriptions and customer promises, record the limit and what stops the agent from going past it.
- Check what counterparties know. A limit the other side knew about protects the company far more than one it didn’t.
- Rank the gaps. The most serious are those that would let the agent make a large or irreversible commitment, or one a counterparty could reasonably rely on. Close those first in step 4.
- Run the regulatory review separately. Is any use prohibited, high-risk, or a consequential decision? Which existing rules apply regardless of AI, such as privacy, export controls or payments rules? An agent can comply with every regulation and still act without authority, and an agent acting within its authority can still violate a regulation. The Regulatory Overlay’s issue spotting sequence lists questions to work through.
Background: Technical Permission vs. Legal Authority, Identity, Credentials & Standards (mandate to control) and The Regulatory Overlay.
4. Close the gaps
- Move the limits that matter into the system. Have the orchestrator or ordering tool check each commitment before it is sent. Identify which receiving service checks each credential and what it will refuse. Card caps are a payment backstop, not a commitment limit. Split powers so no agent can both commit the company and hold independent payment powers: approving invoices, initiating transfers, or changing where payments go. A capped card that settles the agent’s own orders at an approved vendor’s checkout can be an express exception.
- Add approval steps that fail closed. Block prohibited actions outright. For an action the mandate allows with approval, show the authorized reviewer the counterparty, amount and terms. Release only the action approved. A refusal ends the request, and an unanswered request stays blocked.
- Give the agent its own identity, not an employee’s login or a shared key.
- Adopt the policy. The Autonomous Procurement Policy turns what this review checks in one agent into company-wide rules: who approves a deployment, which limits must be enforced, when a person must approve, and who reviews the agent’s activity. Adopt it, or update your own, so the next agent starts from the same rules.
- Align the contracts. Vendor agreements should name the designated agents and their limits, deal with clauses that make every credentialed order binding, allocate compromise risk, and give a cancellation window. The Agent Transaction Addendum is one way to tell vendors the limits. Provider agreements should cover enforcement of configured limits, logs, shut-off, and indemnities.
- Staff the contract deadlines. Counterparty confirmations and notices should go to a monitored inbox with a named owner and backup, triaged in time to act before each applicable deadline. If shipment cuts off rejection for custom or perishable goods, flag the order when it is placed and contact the seller before shipment. A weekly review is oversight, not triage; it can miss these deadlines.
- Log the authority in force, not only what the agent did. Version the mandate so you can show what the agent was allowed to do on any past date.
- Plan the end. Decide which event ends authority, how fast credentials stop working, what happens to orders and offers in progress, how subagents are shut off, and who tells counterparties. Name an owner for each phase of the Incident Playbook.
Background: Building Limits into Agent Systems and Identity, Credentials & Standards (per-agent credentials).
5. Test and sign off
Try to break the finished setup before the agent goes live:
- Over the limit. Try an order above the hard cap and another that is within the cap but requires approval. Does the first stop outright? Does the second wait for approval of that exact order?
- Split orders. Place several orders, each under the cap, that together exceed a total. Is the total caught?
- Other commitments. Have the agent accept terms, order on account, or start a subscription outside what the mandate allows. Is each one stopped?
- Planted instructions. Test external content that reaches the model through a document, email or tool result. If that content steers the model toward an unauthorized order, does the order check stop it?
- Compromised credential. Can authority be cut off quickly, and what can an attacker reach in the meantime?
- Subagents. Can an agent started by the first one do anything the first agent isn’t allowed to do? To prevent that, a separate service, such as payments, should check each request against its own record of what the requesting agent may do, rather than rely on what the first agent passes along.
- Approval service down. Does the agent wait, or proceed?
- Expired mandate or departed approver. On expiry, do orders stop and credentials cease working? If the approver leaves, does the agent pause for reapproval, and does someone handle any notice due to vendors? Test the backup owner’s access to alerts as well.
- Conflicting instructions. When the prompt, an API field and the master agreement disagree, which one controls?
- Reconstruction. Pick a past date and show exactly what the agent was authorized to do then.
Fix what fails and run the test again. Record all required approvals on the signed mandate and finalize the registry entry with that version before issuing production credentials or a card.
The review file
When you’re done, the file for this agent holds:
- the Agent Registry entry, with the mandate version and all required approvals;
- the internal interview answers, with who gave each one, and the provider’s diligence answers if a provider supplies the agent;
- the signed Agent Mandate, with what enforces each line, including the lines for commitments that aren’t payments;
- the controls changed and the contract terms agreed with vendors and the provider;
- the result of the regulatory review;
- the test results;
- the Incident Playbook, with an owner for each phase;
- a date for the next review, and the changes that trigger one sooner: a new tool, a new credential, a higher limit, or a new kind of counterparty.
After an agent has acted
The agent has placed an order, made a payment or a promise, or accepted terms that nobody approved. In the first hours, use the Agent Incident Playbook to contain the agent and preserve the evidence. Then work through the transaction in this order:
- Principal. Whose legal position could the act affect?
- The act. Was it a message, a negotiation, an acceptance, a signature, a purchase, or a payment?
- Actual authority. What had the company’s policy, mandate, approvals and instructions conveyed to the agent?
- What the counterparty saw. What did the company say or do that the counterparty could reasonably rely on, including credentials and past dealings?
- Attribution. Is the agent’s electronic act treated as the company’s act? In most states, UETA’s attribution rule applies. New York hasn’t adopted UETA, and E-SIGN has no attribution rule, so there general law such as agency decides. Does the contract also address orders placed with the company’s credentials?
- Formation. Did a contract form, and on whose terms?
- Mistake or malfunction. Did the agent misread, err, or get manipulated? Is there an error rule or cancellation right?
- Ratification. When the company kept, used or paid for what the agent obtained, which material facts did it know? Did it choose to affirm despite missing facts? Pause further use and check any deadline for rejection.
- Corporate delegation. Did whoever deployed the agent have the authority to delegate?
- The record. What credentials, scopes and approvals were in force, and what do the logs show?
Answers can differ at each step. The common pattern is an agent acting within its credentials but outside its intended scope, a counterparty that relied on what it was shown, and a company arguing that its limit was internal. Whether the software is treated as a tool or as an agent matters mainly at questions 3, 4 and 8; the Tool or Agent? guide sets out where the answer changes the result.
For the situation in front of you, start with the closest worked example:
- The agent bought something, clicked through terms, or bought cloud services or software
- It went over its spending limit, bought outside the approved list, or made a payment
- It told a business customer something wrong or issued a refund or changed an account
- It handed work to another agent
If you provide agents
If you build agents that other companies use, the counterparty will usually look to your customer, and your customer will look to you. Settle four things before a customer asks:
- Where limits are enforced. Know which limits your service enforces in code and which are only instructions to the model. See Inside an Agent System.
- Your contract position. Tie your responsibility to configured controls that fail to work as documented, not to the agent’s judgment. See The provider’s agent went past a customer’s limits and the provider clauses in Agent Use Clauses.
- Your answers to diligence. The Agent Diligence Questions are what customers will ask you. Answer them once in the free Agent Authority Profile and attach it to your questionnaire responses.
- Your records. Keep logs that show which limits the customer set and whether the service enforced them.
This is general information, not legal advice. The steps draw on the guides, which list their sources.