Rubber stamping AI: the costly habit of approving fast

Have you ever clicked "accept" without reading? Everyone has. Now picture doing that fifteen times a day, with an AI proposing a change to a price, to a customer record, to a message that goes straight to the buyer. That is where automation starts to get expensive.

A recent study covered roughly 40,000 runs of AI agents where a human had to approve every command before it executed. The result: one in three risky actions went through unnoticed. Not because people were careless. Because approving became routine, and routine switches off the brain.

This matters even if you have never heard the phrase "AI agent". If you use AI to answer customers, build quotes, update a spreadsheet or send messages, you are already rubber stamping something.

Why we stop looking

Approval turns into theater when three things line up:

• The AI is right 90% of the time. You learn it is right and drop your guard.
• Every item looks the same on screen. A comma change and a price change look identical.
• There is a queue. Twenty pending approvals and five spare minutes do not produce careful reading.

The problem is not that AI makes mistakes. It is that it makes few enough for you to trust it, and bad enough to hurt when it happens.

Good automation is not the kind that asks for fewer approvals. It is the kind that makes each approval worth it.

What actually goes wrong

A simple case: a shop used AI to answer quote requests on WhatsApp. It worked well. One day a customer asked about an item outside the price list and the AI invented a number 40% below cost. The clerk approved it, because the last fifty answers had been fine. The shop honored the price to avoid a fight. A full day of margin gone in one message.

Another one: a company let AI update order status in the system. It started marking anything with tracking movement as "delivered". Nobody checked for two weeks. When customers complained, the history was already a mess and the team spent days rebuilding it.

In neither case was the AI stupid. It did exactly what the context allowed. What was missing was a brake.

Sort the work by risk

The answer is not reviewing everything. Reviewing everything is the same as reviewing nothing, because it wears you out and you go back to autopilot. The answer is splitting by risk.

Make this list for your own process. It takes twenty minutes:

• Reversible and cheap: answering a common question, booking a slot, tagging a record in the CRM. Let it run alone.
• Reversible and expensive: sending a proposal, changing a deadline, reassigning an owner. Let it run, but with a notice and an easy way to undo.
• Irreversible: price, discount, cancellation, anything that leaves the building with your name on it, anything touching money. Real approval, every time.

Notice that most tasks land in the first group. That is why you can automate a lot and still sleep well.

Make the mistake visible

Rubber stamping happens because everything looks the same. So break the sameness:

• Highlight what changed, not the whole text. If the AI altered a number, the number should be shouting on screen.
• Set a ceiling. Discounts up to 10% go through, above that it stops and calls a human. One simple number solves half the problem.
• Group low risk approvals into a daily digest and let high risk ones arrive one at a time. A mixed queue is an ignored queue.
• Keep a log. Not to audit anyone, but so you can answer "what did it do yesterday?" in thirty seconds.

None of this is sophisticated. It is the same logic as a shop assistant who can discount up to a limit and then calls the manager. We knew how to do this long before AI existed.

Trust is measured, not felt

Ask whoever uses the automation: "out of ten things the AI proposes, how many would you change?" If they cannot answer, nobody is looking.

The honest way to measure is sampling. Once a week, take ten actions the AI did on its own and review them carefully, including the approved ones. Within a month you will know whether to loosen the reins or tighten them. Without that, your confidence is just a feeling.

There is a good side effect too: people who review start spotting patterns. "It always gets it wrong when the customer writes in all caps." That kind of observation is worth more than any new setting.

Start small on purpose

If you are setting this up now, resist the urge to switch everything on at once. Pick one process, the most boring and repetitive one you have. Let the AI run it for two weeks with real review. Then expand.

The gain from automating does not disappear if you go slowly. What disappears is the chance of finding a big problem only after it has already turned into a loss.

AI is a great worker for repetitive tasks and a terrible judge of exceptions. Your job is to design the process so it only touches an exception with a person standing next to it.

If you want to automate without turning approval into theater, let's design it together.