Skip to content
caadoo
de

Human in the loop: when AI agents need approval

Companies tend to give an AI agent either no freedom at all or far too much, and both go wrong in predictable places. The fix is to judge each action on its own, by whether it can be undone and what a mistake costs. This article gives you a five-step autonomy ladder you can apply to your own processes this week.

By T. Krause7 min read

Picture an agent that reads incoming orders, checks stock and drafts the order confirmation. It works in the demo. Then someone in the meeting asks the only question that matters: are we letting it send that on its own?

The room usually splits in two. One side wants a person to approve every step. They end up with a system nobody uses, because approving takes longer than doing the job by hand. The other side wants to see what the agent can do. They find out on a Friday afternoon what it promised a customer.

Both sides are asking the question about the whole agent. It only becomes useful when you ask it about each action separately. What may the agent do alone at this particular step, and what would convince you to let it do more?

Photo: Unsplash

Sort the actions, not the agent

An agent is a chain of actions. It reads a mailbox, looks up a customer in the CRM, writes a draft, sends an email, posts an invoice. Judge each of these with two questions. Can it be undone? And what does a mistake cost if nobody notices it for a day?

Reading and drafting. Reading the inbox, pulling data together, drafting a reply or a quote. Nothing has left the building yet. A bad draft costs the time it takes to delete it. An agent can do this alone from day one.

Changing things internally. Moving a ticket, filling a CRM field, filing a document. These can be reversed, but errors travel. If the agent assigns an enquiry to the wrong customer, the next three steps build on that mistake. Fine to run alone, as long as every change is logged and can be undone with one click.

Going outside. An email to a customer, a delivery date, a reply to a complaint. Once it's out, you can't pull it back. A promised Thursday delivery is still a promise when software wrote it. Put a person in front of these, at least at the start. Quotes are a typical case, and how to automate quoting shows where that sign-off sits.

Moving money and deleting. Payments, credit notes, deleted records, cancelled contracts. Either irreversible or reversible only with effort and an apology. A person approves these every time.

There is also a limit set from outside. When an agent makes decisions about people, Article 22 of the GDPR applies across the EU. It gives people the right not to be subject to a decision based solely on automated processing that produces legal effects for them or affects them in a similarly significant way. An automatic rejection of job applicants can fall under it. There are exceptions, and your data protection officer should be the one checking them, not the agent.

The autonomy ladder

The ladder has five steps. Every action an agent takes sits on exactly one step, and each action gets placed separately. The same agent can read emails on its own and still need permission before it sends one.

Step 1, the agent reads. It gathers information, summarises it and shows it. It writes nothing and changes nothing. Every new action you're unsure about starts here.

Step 2, the agent drafts. It produces a proposal that a person accepts, edits or throws away. The person carries out the action themselves, in their own tool.

Step 3, approval before. The agent prepares the action completely. A person reviews it and approves it with one click. Without the click, nothing happens.

Step 4, review after. The agent carries out the action alone. Once a day a person goes through the list and can reverse any of it. This only works for actions that can really be reversed.

Step 5, alone within fixed limits. The agent acts without review, but only inside rules you have written down. An amount limit, certain customer groups, certain templates. Anything outside those rules drops back to step 3.

Here is how to use it. Take one process, say order intake at a 60-person wholesaler. Write each action the agent performs on its own line, and next to it the step and the name of the person who approves. Two rules help. New actions start one step lower than your gut tells you. And nothing irreversible ever goes on step 4, because a review the next morning does nothing about an email that has already been sent.

Approving has to be faster than doing it yourself

Step 3 usually fails at the approval, not in the agent. If approving takes five minutes and the task took seven without the agent, people route around it, and you've paid for a tool that sits on the shelf.

One screen. One view shows what the agent wants to do, the finished draft and the source it worked from. That means the customer email, the order, the line in the price list. Below that, three buttons: approve, edit, reject. The person approving should never have to switch to the ERP to look something up.

The reason sits next to the draft. The agent explains its proposal in one sentence. "Customer asks for a delivery date, 40 units in stock, standard lead time per price list three working days." The reviewer checks a line of reasoning instead of guessing how the agent got there.

Where people already work. The approval request arrives in Teams, in the inbox or in the ticket system. Nobody opens a new portal with its own login after the third week.

Rejecting costs one sentence. Rejecting or editing opens a short field for the reason. It's optional, but it's there. That one sentence teaches you more about your agent than any dashboard.

Widen autonomy on evidence only

After a few weeks someone will want the agent to do more on its own. That's the right instinct, as long as the decision rests on data and not on the mood in the team.

Edits and rejections are your test set. Store every edit and every rejection alongside the original draft. After a month you have a collection of real cases where the agent got it wrong. Before you switch models or change the agent's instructions, run those cases again. If the agent repeats an old mistake, the change doesn't go live.

Move up by rule. Decide in advance when an action earns the next step. For example, when nobody has changed anything of substance over a set period and a set number of cases. The right numbers depend on the risk. A CRM field needs less evidence than an email to a customer.

Watch for rubber-stamping. Once approving becomes routine, people click without reading. The EU AI Act names this pattern in Article 14 for high-risk systems. It calls it automation bias, the tendency to rely too much on what a system produces. You can see the signs in your data. Long drafts approved in a few seconds, and weeks without a single edit. That means the agent is ready for the next step, or nobody is looking any more. Find out which before you move it up. A short weekly spot check often does the job: pick one approval and ask the person why they let that draft through.

Moving down is normal. After a model change, a new product line or a price change, an action drops back a step until the edits show it's ready again. The ladder goes both ways.

An agent that may do nothing on its own saves you no time. One that may do everything will cost you a customer sooner or later. Between the two sits a table listing every action, its step and the name of the person who approves it. Writing that table takes an afternoon, and it is the one part of your agent you should never leave to software.

Topics

  • AI agents
  • Automation
  • Approval workflows
  • Management
  • Human in the loop
All articles

Keep reading

  • 8 min read

    AI assistant for company documents: fix files first

    Most companies planning an AI assistant for their own documents spend their time choosing a model. The real work is in the file share, with its outdated versions, documents nobody owns and access rights nobody has checked in years. Here is the order to do it in, from one bounded question area to a test with 30 real questions before anyone else gets access.

  • 7 min read

    Shadow AI in the workplace: find it, don't ban it

    Many managing directors either assume nobody in their company uses AI or plan to ban it. Meanwhile staff are pasting customer data into private accounts, often on their own phones. This piece shows how to find out what's really happening and bring it under control with one sanctioned tool, short rules and, in Germany, the works council.

  • 8 min read

    AI literacy under Article 4: what companies must do

    Most mid-sized companies file the EU AI Act under problems for software vendors. Article 4 applies to any company whose staff use ChatGPT or Copilot at work. This piece covers what the obligation asks for, what it doesn't, and how to meet it without hiring a consultancy.

End of call

You can hang up now.

Or book 15 minutes. I'll ask you the questions I'd ask on the phone, and tell you at the end whether it fits. If it doesn't, we've both lost a quarter of an hour instead of a month.

Or call. Cold, if you like. My AI assistant picks up and I call you back. +49 5136 9414999

Analytics

With your consent I use Google Analytics to see which pages get read. Without it, no analytics code loads. Your answer is stored locally in your browser.

Privacy notice