- Home
- Insights
- New ways of working
- Which steps need human approval when an AI agent does the work?
Which steps need human approval when an AI agent does the work?
Putting an approval on every step does not make a process safe; it only makes it slow. This guide covers the steps where approval is indispensable, three levels of approval and how a good approval step is designed.
Why does approving every step not work?
The first reaction is usually the same: whatever the agent does, a person should look at it. On paper that seems safe; in practice it has two results.
First, the process does not get faster. Work the agent prepared in minutes waits in the approval queue. Second, and more importantly, an approver who faces dozens of requests a day stops looking at the content. Approval stops being a check and becomes a button. The request that really matters passes at the same speed as all the others.
The aim is not more approvals but approval in the right place.
How do you identify a step that needs approval?
Four questions are asked of every step in the process. If the answer to even one is “yes”, that step is tied to human approval.
| Question | Example step | Why approval is needed |
|---|---|---|
| Can it not be undone? | Deleting a record, sending an email. | Once the mistake is noticed, there is no way to correct it. |
| Does it spend money? | Raising an ad budget, placing an order. | The cost of the mistake is incurred immediately. |
| Does it leave the company? | Publishing, writing to a customer. | What is said in the company’s name cannot be taken back. |
| Does it affect a person’s situation? | Screening an application, rejecting a request. | The decision lands on a person, who may ask for the reasons. |
What are the three levels of approval?
Approval is not one thing. One of three levels is chosen according to the risk of the step.
- Approval first. The agent prepares; nothing happens until a person approves. Steps with a “yes” to any of the four questions sit here.
- Review afterwards. The agent acts; a person examines a sample at set intervals. For steps that can be undone and have limited impact.
- Record only. The agent acts and logs it; the log is looked at only when a problem appears. For steps such as reading data or preparing an internal draft.
Which step sits at which level is written in the agent’s brief.
How is a good approval step designed?
If the approver is shown nothing but “do you approve?”, approval becomes a formality. A good approval request shows four things:
- What will be done. The action itself, in one sentence.
- On what basis. The data that led the agent to this suggestion.
- What happens if it is not done. The consequence of waiting or refusing.
- How it is undone. The steps if there is a way back; a clear statement if there is none.
Two more rules are added: who gives the approval is known by name, and if no answer comes, the agent does not act.
What does this look like in a marketing team?
Take an agent that monitors the advertising accounts. Different steps of the same agent sit at different levels:
- Pulling and summarising weekly performance data: record only.
- Adding clearly irrelevant search terms as negative keywords: review afterwards.
- Moving budget between campaigns: approval first, because it spends money.
- Putting a new ad text live: approval first, because it leaves the company.
When is approval relaxed?
When evidence has built up. While a step is at “approval first”, how often the approver changes the agent’s suggestion is recorded. If suggestions have been approved unchanged for a long period and the step can be undone, it can move to review afterwards. The threshold is set by the company according to its own appetite for risk, and written down.
The reverse also holds: a step that produces an error under review afterwards goes back to “approval first” until the cause is understood. The person who makes these decisions and keeps the record is the agent’s owner; the whole role is described in the AI Agent Manager article.
Our recommendation
Decide the approval arrangement when the process is put on paper, not when the agent is being set up.
- Write the steps of the process one under another and ask the four questions of each.
- Note the level next to every step: approval first, review afterwards or record only.
- For every step at “approval first”, write the name of the approver and of their stand-in.
- Make it a rule that the agent waits when no answer comes.
- Look at the records after three months: which approvals passed without ever being changed, which review turned up an error? Update the levels accordingly.
Frequently asked questions
Is it not safer to require human approval on every step?
No. As the number of approvals rises, the approver starts approving without reading. A few approvals that are actually read are safer than a token approval on every step.
Who should give the approval?
The person who is accountable for that step’s outcome and has the knowledge to judge it. Approval authority goes to a named person, not a role, and who covers in their absence is written down in advance.
What does the agent do if the approver does not respond?
The default is to wait. When the time limit passes, the agent does not act and reports the situation to its owner. A setup that treats silence as approval removes the approval step in practice.
Can human approval be removed completely over time?
On some steps yes, on others no. For reversible, low-impact steps, approval can become after-the-fact review once enough evidence builds up. On steps that cannot be undone or that directly affect a person, approval stays.
Read next
All insightsWho is accountable for AI agents? The AI Agent Manager role
Agents no longer stop at answering; they act on a company’s behalf. This piece describes the role that owns those actions: what it takes on, what it is not, and where a company starts.
Read the article AI Agent ManagerWhen does the role arrive? Analysis · 7 min readWhen do you need an AI Agent Manager? Authority decides, not the number of agents
Companies usually think about this role when the number of agents grows. What decides it is not the number but the authority the agent is given. This piece covers the signs that the role is needed, and who takes it on at what scale.
Read the article