Lee Siang Fong

Blog

AI Agent Approval Checklist for Business Workflows

Plan AI agent permissions and human approvals with a practical checklist covering proposed actions, retries, audit records and recovery before launch.

A human approval checklist for AI agents should define what the agent may do, what a person must review and which exact action the approval authorizes. The application should check that approval before the action happens, rather than relying only on instructions in a prompt.

Start with a narrow workflow. An agent that prepares an invoice draft needs different permissions from one that can send invoices, change bank details or release payments.

Write an inventory of possible actions

List the tools the agent can call and the changes each tool can make. Avoid a vague description such as “access to finance.”

For an illustrative supplier-invoice workflow, the inventory might include reading an invoice, looking up a purchase order, preparing a draft record and sending that record for review. Payment release can remain outside the agent's scope.

If your use case needs only reading and drafting, do not grant write permissions simply because a connector supports them. OWASP's excessive-agency guidance describes the risks of unnecessary functionality, permissions and autonomy.

For the distinction between an assistant that answers and one that acts, see AI agent vs chatbot.

Decide where human approval belongs

Classify actions by their effect on the business workflow. The table below is a planning example; your team should set its own rules for the task and applicable obligations.

Action Suggested starting control
Read an approved internal document Check the user's existing access rights
Prepare a draft Keep it visibly separate from an issued record
Update an operational record Show the proposed change to the authorized reviewer
Send an external message Show recipient and final message before approval
Change permissions or financial details Apply the organization's established control process
Delete important records Provide the required review and recovery process

The objective is a clear decision at the appropriate point. An approval request should tell a busy reviewer what will happen, without hiding essential details inside a long conversation.

Tie approval to the exact proposed action

Show the target system, record, recipient or destination. Include the fields that will change and any document the action relies on.

For an invoice draft, a reviewer might need the supplier, invoice number, amount, currency and matching purchase order. “Approve the next step” is not a sufficiently clear description.

Store approval against a specific version of the proposal. If the amount or recipient changes afterward, the application should require review of the changed action.

A person approving one record has not necessarily approved a batch of records or future actions. Make the scope explicit in the interface and in the stored approval record.

Enforce permissions outside the model

A model can propose an action. The application still needs to decide whether the action is allowed for that user and workflow.

Check permissions when the tool executes. A sentence in a document, a previous model answer or an old approval should not override the current user's access.

Use credentials limited to the required operation. Where possible, separate the ability to prepare a draft from the ability to issue or finalize it.

The AI governance guide explains how to assign responsibility for these operating decisions.

Plan for retries and duplicate requests

Network interruptions are an ordinary part of software operation. A user may click again because the first request appears slow, while the original request is still running.

Design the action so that a retry can identify the earlier attempt. The application should be able to return the existing result or job status instead of performing the external action again.

For the invoice example, use a stable operation identifier and check for an existing result. A disabled browser button helps the user, but the server also needs protection against duplicate requests.

If the external system's outcome is uncertain, show that uncertainty. Do not label a request “failed” and automatically repeat it when it may already have succeeded.

Keep evidence useful to a reviewer

Record the proposed action, the relevant inputs, the approval decision, the executing identity and the external result. Keep enough context to investigate without collecting every piece of unrelated information.

Separate a model's explanation from the system's actual event record. “I sent the invoice” is a claim; the external system's accepted record is evidence of the operation.

Decide who may see these records and how long they are needed. Audit information can itself contain sensitive business details.

Give people a way to pause and recover

Name the person responsible for an operational problem. Show staff where to report a wrong draft, unauthorized action or confusing approval request.

Decide how to pause further actions while an incident is investigated. Check what can be undone and what requires a follow-up action in the external system.

NIST's AI Risk Management Framework is a voluntary reference for organizing risk management. Your recovery steps still need to match the workflow and systems you actually use.

Test the approval boundary before launch

Try a request without approval, an approval for an older version and a repeated click. Test a user who lacks permission and a proposal that changes after the reviewer opens it.

Also include a misleading instruction inside a sample document. The agent should not treat source material as permission to bypass the application's approval process.

Write down the expected result before running each test. Use the failures to improve the workflow, then repeat the same cases.

A useful human approval checklist for AI agents connects the decision on screen to the action in the system. Start narrow, preserve the evidence and make uncertain outcomes visible.

Reusable AI action approval worksheet

Download the editable approval worksheet, or copy the fields below into your existing review system. Use one record for one action or an explicitly bounded batch. This template is a planning aid, not proof that a system is compliant or safe.

Record What the reviewer needs
Proposal Request ID, version, business purpose and accountable owner
Exact action Target system, records, recipient, before-and-after values and exclusions
Evidence Source versions, last checks, calculations and unresolved uncertainty
Limits Allowed identity, amount or volume, conditions and approval expiry
Decision Authorised reviewer, approved version, restrictions and timestamp
Execution Attempt ID, external receipt and confirmed or uncertain outcome
Recovery Pause owner, reversal limits and recovery procedure
Follow-up Expected result, actual result, review date and approved learning

Worked example: approving one warranty reply

Imagine a support assistant drafts a reply for a customer. This is a hypothetical planning example, not a reported deployment.

The proposal identifies the exact customer, final message and warranty-policy version. The reviewer confirms that the evidence supports the promise and approves this message only. Refunds and future messages remain outside that approval.

If the recipient or message changes, the earlier approval no longer covers the proposal. If the sending service times out, the team checks the existing attempt before trying again. The outcome stays uncertain until there is evidence.

After the case closes, the owner records whether the reply resolved the issue and whether the guidance needs correcting. Keeping a result does not establish that AI caused an improvement.

For planning and evaluation, use the enterprise AI roadmap. For the wider controls around the action, see six layers of AI agent governance.

Quick answers

Does every AI action need a human click?

Not necessarily. Define the boundary according to the task, its consequences and your organization's rules. Reading an approved document and issuing an external instruction have different effects.

Is a confirmation message in the prompt enough?

The application should enforce the approval before executing the tool. A prompt alone does not provide a reliable authorization check.

What if the draft changes after approval?

Require review of the changed version before it executes. Approval should refer to the exact action the person inspected.

Can an approval log prove the action was correct?

It records a decision and helps establish accountability. It does not by itself prove that the underlying facts or judgment were correct.

Sources