An artificial intelligence (AI) assistant can draft a supplier message, recommend a response to a disputed invoice and send a payment instruction. Those actions should not inherit the same authority just because they happen in one conversation.
Human approval for AI belongs where a proposed action needs judgement or authority the system has not been given. The decision depends on the consequences, how far the action can be undone, the uncertainty in the evidence and who is entitled to decide.
Our earlier articles separated capability from authority and explained which permissions an agent should have. This instalment asks what a person must actually review before an action proceeds—and how to stop approval becoming a routine click.
Human approval for AI should match the action
Not every generated sentence needs a separate approval. An assistant saving a private draft can often operate within an agreed permission boundary. A message disclosing confidential information needs a different decision, even if it contains only one sentence.
Use four considerations together:
| Consideration | Practical question | Example |
|---|---|---|
| Consequence | What happens if this action is wrong? | A mistaken internal label differs from an incorrect external commitment |
| Reversibility | Can we restore the previous situation, including external effects? | Deleting a draft differs from recovering information already emailed |
| Uncertainty | Is the evidence complete, consistent and suitable for this decision? | A familiar invoice may still contain changed payment details |
| Authority | Who is allowed to make this decision? | Someone who can prepare a payment may not be allowed to authorise it |
These are decision prompts, not a scoring formula. High confidence does not grant authority. A small amount does not make a disclosure harmless. A reversible-looking record change may already have triggered another system.
Human oversight of AI agents should therefore be attached to a defined action and operating context. “Ask me before doing anything important” is too vague to be the only rule.
Separate drafting, recommending, approving and executing
Return to the hypothetical invoice assistant used earlier in the series. Its job is to assemble invoices, matching records and discrepancies for a finance operator. The following is an illustrative starting design, not a universal finance policy or a claim about a particular product.
| Stage | Useful assistant contribution | Approval boundary |
|---|---|---|
| Draft | Extract fields and prepare a clarification message in a restricted review area | No separate click for each draft; authorised access and retention rules still apply |
| Recommend | Explain a mismatch and propose a next step with source evidence | The proposal remains distinguishable from a confirmed decision |
| Approve | Present the exact message or record change for a decision | A person with the relevant authority accepts, edits, rejects or escalates it |
| Execute | Carry out an approved action, if the application has that capability | The application checks the approved version, current permissions and result; payment execution remains outside this assistant's scope |
The Open Worldwide Application Security Project (OWASP) recommends human approval before high-impact actions and enforcement in downstream systems in its excessive agency guidance. That supports a useful distinction: the model can propose an action while the application decides whether the required authorisation exists.
A person approving the invoice assistant's message is not granting it general access to change supplier records. Approval is narrower than a new permanent permission.
Show the reviewer the decision, not just a reassuring summary
Imagine the assistant receives a revised invoice with a changed bank account. It prepares a polite message asking the supplier to confirm the new details.
An approval card saying “Send routine clarification?” hides the most important fact. The reviewer needs to see the changed account, the source invoice, the proposed recipient and the actual message. They also need to know that replying to the same correspondence does not independently verify the requested change.
For this example, the supplier-record change follows the business's established verification process. The assistant can prepare evidence and flag the exception. It cannot make the new details authoritative because the email looks convincing.
A useful AI approval workflow puts five things together: the proposed action, what differs from the current record, supporting sources, unresolved questions and the consequence of approving. Attachments and destinations matter as much as the text.
AI workflow review should make checking easier without hiding the underlying material. The reviewer may inspect the relevant source passage first, then open the full document where context matters. A generated explanation is something to assess, not independent evidence that the recommendation is correct.
A human in the loop can still become a rubber stamp
People cannot provide meaningful review if the queue is too large, the evidence is difficult to find or the application makes acceptance the easiest way to clear work. Adding an approval button does not resolve those conditions.
Microsoft's overreliance framework recommends helping users understand limitations, recognise when to verify and check outputs with less effort. It also cautions that explanations and apparent sources can increase trust even when they are unreliable. The design needs evaluation with actual users.
For human-in-the-loop AI workflows, give the reviewer time, relevant knowledge and permission to stop the case. “Return for more information” must be a usable outcome. A person who cannot assess the exception should be able to route it to somebody who can.
Avoid asking the same person to approve every harmless preparation step and then expecting exceptional attention at the consequential one. Keep routine work inside a clear delegated boundary; make changed destinations, missing evidence and policy exceptions stand out.
Measure whether reviewers catch the important mistakes, not only how quickly they clear the queue. A high acceptance rate can mean good recommendations or superficial review. It does not distinguish them by itself.
Approval should survive only while the approved action remains the same
Suppose the operator approves a supplier message with one attachment. Before sending, the assistant rewrites the message, adds another document or changes the recipient. The earlier approval should not silently cover that new action.
OWASP's transaction-authorisation guidance stresses presenting significant transaction details, enforcing authorisation on the server and checking it before execution. It also addresses changes to transaction data and the validity of authorisation. Applying those principles to an AI workflow means binding approval to what the person actually reviewed.
For the invoice example, require a fresh decision when the recipient, content or attachments change. If relevant evidence or the approver's authority changes while work is waiting, the application should recheck whether execution is still permitted.
The same principle applies to contract review. Approval of a service schedule should not authorise a different schedule merely because both belong to the same customer agreement.
The AI workflow approval process needs to preserve who approved what, which version they saw and which action actually occurred. A chat message saying “approved earlier” is not sufficient evidence.
Decide when to stop and who can restart
AI agent permissions set the outer boundary. Within it, uncertainty can still require a pause. Missing source documents, conflicting amounts, an unexpected recipient or a proposed action outside the reviewer's authority are reasons to hold the case.
Do not treat silence as consent. If the approver is away or the approval expires, the case should remain pending or go to an authorised delegate under the business's policy. A deadline approaching is a reason to escalate, not a reason for the assistant to grant itself permission.
Some actions should remain unavailable to the assistant even after an ordinary reviewer clicks approve. In our example, supplier-master changes and payment execution follow separate controls. A review screen should not offer a shortcut around them.
AI workflow governance includes naming who owns those rules and who may change them. That responsibility belongs outside the assistant's ability to rewrite its own instructions. In a small business, one person may hold several roles, but the authority being exercised should still be clear.
When can action proceed without another click?
A business can deliberately delegate some bounded actions in advance. For example, an assistant might label a document in a restricted staging queue or save a draft comparison without asking every time.
That decision should follow from demonstrated performance, limited consequences, controlled scope and a practical way to detect and correct errors. It should not emerge accidentally because somebody switched off an inconvenient approval prompt.
Review after the event can be useful for low-impact work. It cannot prevent an email already sent or a commitment already communicated. Decide which actions need a gate before execution and which can be monitored under an explicit policy.
There is no universal confidence percentage or transaction amount that makes autonomy appropriate. A model's confidence statement is not a measured error rate for your process. Expand delegation using evidence from the actual task, including exceptions, and revisit it when tools, inputs or responsibilities change.
Test a useful approval and a stale one
Ask the supplier or builder for a short demonstration with synthetic records. First, the assistant should prepare an ordinary clarification message. The reviewer sees its recipient, content, attachment and sources, approves it, and the application records the actual sending result.
Next, change the recipient or attachment after approval. The send should stop until the changed action receives the required decision. Also try the changed-bank-details case: the assistant should surface the exception while remaining unable to update the supplier record or release payment.
These are proposed tests, not results from a product evaluation. Have the builder check the enforced action path, not only whether the model politely refuses.
AI workflow monitoring should distinguish waiting, rejected, authorised, failed and completed actions. If sending times out, establish whether it actually occurred before retrying. Approval to send once should not become accidental permission to send repeatedly.
The useful test of AI workflow safety is whether the person can understand and stop the consequential action while useful preparation continues. Ask: can the reviewer explain exactly what they approved, and can the system show that this is what it did?
The next instalment will turn these boundaries into a small, measurable workflow pilot. For the wider reasoning behind them, the AI control problem explains why a well-worded goal still needs enforced limits.
