Approving AI agent actions: segregation of duties, evidence and signed grants
How to approve what AI agents pay, refund, deploy or change: segregation of duties, four eyes, evidence, signed grants and a tamper-evident audit trail.
Updated · 10 min read
In short
- An agent that moves money, changes production or grants access acts with your organization's authority, so it needs the controls a person doing the same job would have: someone else approves, and the record shows who and why.
- Bind each approval to the exact action: a signed, single-use grant over a hash of the payload, which expires and which the executing system checks before it acts.
- For most organizations this is internal control and governance, not a question tied to any one regulation: segregation of duties and a tamper-evident record apply whatever the organization does or where it operates.
- Where an agent's action is itself a decision about a person with legal or similarly significant effects, a human review only counts if it is real: a token click that never changes the outcome is not oversight.
An AI agent that pays invoices, issues refunds, deploys code or changes who can access what acts with your organization's authority. It needs the controls you would apply to a person doing the same job: the action is approved by someone other than the one who asked for it, the approval covers exactly what is executed, and the record shows who decided, on what evidence and when. For most organizations this is internal control and good governance, not a question tied to any one regulation.
Why put a human approval step in front of an agent?
Three reasons hold regardless of what, if anything, a regulation says.
- Errors are cheap to make and expensive to undo. A model can misread an invoice, act on stale context or follow instructions planted in a document it was asked to process. Money sent, data deleted and production changed are hard to reverse.
- Someone must answer for it. When an auditor or a customer asks who approved an action and why, "the agent did it" is not an answer. The organization needs a named person and a reason.
- Delegation needs limits. People in finance and engineering work within approval limits and change processes. An agent that bypasses them opens a hole in controls the organization already relies on.
Not every action needs a person. Set thresholds by risk: small refunds within a policy can be approved automatically, larger ones by one person, the largest by two people in sequence. What matters is that the rule is written down in advance, applied every time, and fails closed, so that an action no rule covers is refused rather than executed.
What do segregation of duties and the four-eyes principle mean for agents?
Segregation of duties means that no single person or system controls an action from start to finish. The one who asks for an action is not the one who approves it, and the one who approves it does not execute it unchecked. The four-eyes principle is its simplest form: a consequential action needs a second person.
Applied to agents:
- the agent is the requester and never approves its own request;
- a person who triggered or configured the request should not be its approver;
- where two approvals are needed, they come from two different people, and one person counts once across all steps;
- the system that executes the action, a payment script or a deployment pipeline, acts only on a valid approval and cannot approve.
Few governance frameworks mandate a four-eyes check by name, but many organizations already require one for consequential actions as a matter of internal control design. An agent that can trigger and, in effect, approve its own request closes that gap on paper only; applying the same control to agents that a person doing the same job would face closes it in practice.
What should the approver see?
An approval is only as good as what the approver was shown. Each request should carry:
- the reason, written by the agent and presented as the agent's claim, not as a fact;
- the payload as it will be executed: amount, currency, payee and account, or service, environment and commit;
- context: the invoice, the purchase order, the pull request, the ticket;
- who is asking: the agent's name and model, the run, the credential it used and the person or system that started it;
- the rule that applied: why this request needs these people.
What evidence should an approval carry?
Match the evidence to the risk. A routine approval can be one click. A larger one should carry a written reason. For the highest risks, require a document or a photo: the signed invoice, the supplier's payment details confirmed by phone, the change risk assessment. Store each file with a cryptographic hash, so you can later show that the file you hold is the one the approver attached.
How do you bind an approval to the exact action?
The weak point in most approval flows is the gap between approving and executing. A person approves a payment of EUR 1 200 to a supplier; the system then executes a different amount or pays a different account, by error or by manipulation. Closing that gap takes four properties:
- A signed grant. When a request is approved, the approval system issues a token signed with its private key that names the request, the action, the approvers and a hash of the exact payload that was approved.
- Verification by the executor. Before acting, the system that executes the payment or deployment verifies the signature against the published public key and recomputes the hash of the payload it is about to execute. If one field differs, it refuses.
- Single use. The executor reports that it used the grant, and a second use is refused, so one approval cannot pay twice.
- Expiry. Requests expire if nobody answers, and grants expire shortly after approval. An approval given on Monday for Monday's facts should not authorise Friday's payment.
The approval system should never hold the payment credentials or the deployment keys. It decides whether an action may happen and proves it; the executor, which holds the keys, checks the proof.
What should the audit trail contain?
Record every event: the request with its payload and reason, each approver's answer with their comment and evidence hashes, the issuing of the grant, the execution report, and any rejection, cancellation or expiry. Make the record tamper-evident, for example by including in each entry a hash of the previous one, so that a deleted or altered entry breaks the chain. Keep it as long as the obligations behind the action require, such as accounting, tax or supervisory record-keeping rules.
When does this go beyond internal policy?
For most organizations, a human approval step in front of an agent is internal control, not a specific legal duty. One case turns it into a data-protection requirement: if the agent's action is itself a decision about a person with legal or similarly significant effects, such as approving or denying something they asked for, Art. 22 of the General Data Protection Regulation limits decisions taken without meaningful human involvement. A human step only takes such a decision out of "solely automated" processing if it is meaningful: the Article 29 Working Party's guidelines, endorsed by the European Data Protection Board, say oversight must not be "just a token gesture" and must be carried out by someone with "the authority and competence to change the decision".
A checklist before you give an agent the keys
- List the actions each agent can take and the systems it can reach.
- For each action, write the rule: who approves, at what thresholds, with what evidence, and what happens when no rule matches (refuse).
- Give each agent its own identity and credentials, never a person's or a shared account.
- Keep the credentials in the executor, not in the agent or the approval system, and make the executor verify a signed, single-use, expiring grant against the payload.
- Keep a tamper-evident record of requests, approvals, evidence and executions.
- Review the whole flow against your organization's own change-management and access-control policies before the first production run.
DueReason's templates for agent payments and production deployments start from this model: thresholds by amount or environment, Finance with the invoice and then management for large payments, an engineering lead and then security for production, and a signed grant bound to the payload that the executor verifies before it acts.
Questions
- Is a human click before every action enough?
- Only if the person sees what will be executed, is not the requester and can refuse. Approving a description and then executing something else proves little; bind the approval to a hash of the exact payload and have the executor check it.
- Can low-value actions be approved automatically?
- Yes, if a written rule says so, for example refunds under a set amount. That is the organization's decision, taken in advance, not the agent's. Keep automatically approved actions in the same record as the others.
- Do we need two approvers for every production deployment?
- No specific number is required by default; match the number of approvers to the risk of the change. What matters is that the approving person is independent of whoever requested and implemented it, and that the decision is recorded.
- What if no approver is available and the action is urgent?
- Define an emergency route in advance: who may approve, what gets recorded and how the action is reviewed afterwards. Urgency should change the route, never skip the record.
Put a verifiable approval between your agents and your money
DueReason asks the right people in order, collects their evidence and gives the executing system a signed, single-use grant bound to the exact payload, so a changed amount or account is refused.
Read the API docsSources
- Regulation (EU) 2016/679, General Data Protection Regulation (EUR-Lex)
- Article 29 Working Party, Guidelines on automated individual decision-making and profiling, WP251rev.01 (European Commission)
This guide explains the regulation as we read it; it is not legal advice. Check the official text and your national law, which may add obligations.