The 90% Trap
The dangerous failure is not the one an agent makes — it is the work that looks completed. Why I check almost nothing my agents produce, where that stops being safe, and why a lost conversation never shows up in any report.
My agent brought a client all the way to payment. The conversation went smoothly, the person agreed, all that was left was to send the link. The agent sent it. And the client vanished.
He never saw the link. The platform had stripped it out of the message — that is how its spam protection works. The client was staring at nothing, unable to work out what was being asked of him. And the agent, believing the task was done, sent it again.
Nothing broke. The agent did exactly what it had been told to do. There was no error — and the payment was nearly lost.
That is the ninety percent trap. The danger is not the mistake. The danger is work that looks completed.
I am not the first to fall into it. In Stack Overflow's 2025 developer survey, sixty-six percent named this as their single biggest frustration with AI: solutions that are almost right, but not quite.
And yet I check very little
Here I have to say something that contradicts most of what gets written about working with machines.
I barely check my agents' output. By my own estimate, verification takes me less than two percent of the time the same work would have taken by hand.
The reason is simple and worth understanding. Checking the process is expensive. Re-reading how an agent assembled a result, step by step, takes longer than assembling it yourself. Checking the result is cheap: skim the text, open the page, look at the document. Seconds.
So I look at outcomes and I do not look into process. And the checking itself I often hand to a separate agent — verification is work too, and work can be delegated.
Most of what gets produced I do not check at all.
So where is the trap
It sits wherever the result cannot be seen.
You will open a document and scan it — you will spot the line that is wrong and the one that does not belong. You will read the text. You will open the page. In every one of those cases the agent's mistake is lying in front of you and costs one second of attention.
A conversation lies nowhere. The person simply did not reply. He did not write "I didn't like the way I was spoken to"; he wrote nothing at all. There is nothing to look at, and no report will ever show you that loss.
That is exactly why the link problem survived longer than it should have. From the system's point of view, everything was fine.
I had a second case of the same kind, in another part of my operation: the system that keeps my website stocked one day decided there was nothing left to publish, and removed everything. It did not break and it did not complain — it reported success. I found out only when I opened the site and looked.
What I do about it
In the early weeks I read the conversations live. Not as quality control, but because this was the one place where a mistake could not be seen after the fact.
When the agent said the wrong thing, I took the conversation over, carried the person through myself — and only afterwards went and fixed the agent.
That order is the whole point. Save the client by hand first, repair the system second. The temptation runs the other way: you spot the fault, you dive into the settings, and meanwhile the person leaves. A conversation cannot be paused.
What this means if you run a team
Your attention is a resource, and it should not be spread evenly.
Anything that leaves a trace can be checked quickly and selectively: documents, emails, reports, listings, descriptions. Mistakes there are visible and cheap to catch.
Anything that happens inside a live conversation deserves different treatment. Not because the machine is worse at it, but because a failed conversation leaves no trace. The customer does not complain. He leaves.
And when you count the time you saved, remember what that count is made of. Savings are measured in what you can see. Losses inside conversations appear in no report at all.
What not to do
Do not treat a clean report as proof of a result. A system reports what it did, not what arrived.
Do not fix the agent while the client is still in the conversation. The person first, the settings after.
And do not build your checks where it is convenient to look. Build them where the mistake cannot be seen.