e27 published a piece on 11 November 2026 about an enterprise knowledge management workflow. AI had been used to turn years of internal documentation into a clean wiki article. The article explained the engineering setup behind a legacy product. The goal was simple. Give new engineers one reliable place to understand the product before they started working on it.
The page wouldn’t load for me, so I am working from the feed summary only. I won’t guess at what went wrong after that. The title tells you where the piece is heading: the costly mistakes are the ones that don’t look like mistakes.
A clean article gets trusted
Think about how a bad draft gets caught. It has typos, odd phrasing, a section that trails off. Someone reads it and feels uneasy, so they check it.
An AI-written wiki article has none of those signs. It is tidy, well organised and sure of itself. Nobody feels uneasy, so nobody checks. The polish that makes the output useful is the same thing that makes people skip the review.
That matters more for a wiki than for most documents. A wiki is meant to be the place people go so they don’t have to ask anyone. A new engineer reading it has no way to spot a wrong line. That is the reason they are reading it. They will build on it.
Is there a real need here?
The need is real. Years of scattered documentation about a legacy product is a genuine problem, and new people do need one place to start. Using AI to pull it together is a sensible idea. I would do the same.
What changed is the speed and the cost of the first draft. Work that once took weeks of someone’s time now takes an afternoon.
What didn’t change is who is responsible for the content. The old documentation was written by people who knew the product, and it was still often wrong or out of date. Summarising it with AI doesn’t make it more accurate. It makes it more readable, and readable text gets believed.
My take: you direct, the AI produces, you check and decide
I hold to a simple rule. Trust nothing unverified. You direct, the AI produces, you check and decide. You are responsible for every output that leaves your hands.
A wiki is where that rule gets tested, because the output doesn’t leave your hands once. It stays up and gets read again and again. One wrong sentence gets read by every new hire.
So the check needs a name. Not “the team will review it”. A named person who signs off before the page goes live, and who knows the product well enough to catch a wrong claim.
Take a clinic that has AI turn its old procedures into one handbook for new front-desk staff. Or a service business that has AI write up how its pricing and cancellation rules work. Or a marketing team that has AI summarise its brand guidelines. In each case the clean result will be believed. If the cancellation rule is wrong, staff will tell customers the wrong thing, confidently, for months.
The check matters most where being wrong is costly. A typo in a welcome page costs nothing. A wrong refund limit, a wrong safety step or a wrong system dependency costs real money. Spend your review time there first.
Where this shows up in my own work
I build AeroChat, and the same rule applies to how I think about it. When a business sets up customer service workflows, the limits on what can be promised, like refunds and discounts, are set by the business and held by the platform. They are not left to anyone’s reading of a document. Rules belong where being wrong is costly, and judgement goes where it is needed.
What a business should do
- Pick a named owner for each AI-written source of truth. One person, written on the page, who signs off before it goes live.
- Choose someone who knows the subject. A person who can read “the legacy system connects to X” and know whether it does.
- Check the costly claims first. Money, safety, legal terms, system dependencies, dates and limits. Read each one against the original source or a person who knows.
- Keep the sources. If the AI summarised ten documents, keep links to them so the reviewer can check a claim in a minute.
- Put a review date on the page. Wikis go stale, and AI-written ones go stale while still looking fresh.
Do this before the wiki is announced to anyone.
Source: e27: The most expensive AI mistakes don’t look like mistakes