Addaly is in open beta. Things will change, and AI answers can be wrong — check anything that matters.

Checking your own work, and talking to a sceptical boss

AI at Work · lesson 8 of 8 · 10 min

Two things left, and they are the same thing seen from inside and outside.

You are the one who signs it

When AI-assisted work goes wrong, nobody blames the AI. In 2023 two New York lawyers were fined after filing a brief citing six cases that did not exist. The court did not care where the citations came from. They had signed the filing.

That is the rule everywhere. Your name on it means you vouch for it. So build a check that scales with what a mistake costs.

Low stakes — an internal note, a first draft, a summary for yourself. Read it. If it is obviously fine, ship it.

Medium stakes — goes to a client, a manager, a parent. Read it as though a junior colleague wrote it and you are responsible. Verify every number, name and date. Delete anything you cannot defend.

High stakes — money, health, legal effect, someone's job, anything published. Verify every factual claim against a source you trust and can point to. Have a second human read it. Assume you will be asked, in a room, why you wrote each line.

The four checks that catch most of it

  1. 1Every specific. Numbers, dates, names, prices, section numbers, citations. These are where fabrication concentrates, because they are exactly what a model has to produce precisely rather than plausibly.
  2. 2Everything you did not supply. Go back to lesson one. Trace each fact: did I give it this, or did it come from somewhere? The second kind gets checked.
  3. 3The confident sentence you cannot source. If you cannot say where a claim came from, it is not yet a claim. It is a draft note to yourself.
  4. 4What is missing. The hardest and most valuable. Would an expert reading this say "but you have not mentioned…"? Omission is the failure mode that never announces itself.

One discipline is worth all the others: never send something you do not understand. If a paragraph is convincing but you could not explain it aloud, you have just outsourced your judgment to a text generator. Ask it to explain, or cut it.

Talking to a sceptical boss or team

Now the outside view. Your colleague who says "it just makes things up" is not being unreasonable. They have probably seen it happen. The worst response is enthusiasm.

Agree with the accurate half of their objection. "You are right that it invents citations. That is why I check every one and why we would not use it for anything filed." You have now established that you are the careful person in the conversation. Every subsequent thing you say lands differently.

Bring a result, not a demonstration. Not "look how impressive this is." Instead: "The tender response usually takes me two days. It took four hours. Here is the draft — tell me if the quality dropped." Let them judge the output. That is a question they can actually answer, and it respects their expertise.

Name the boundary yourself, first. "I use it for first drafts and for summarising long documents. I do not use it for anything with client data in it, and I do not use it for the final numbers." A colleague who hears you draw a line stops worrying that you have no lines.

Do not oversell. If you promise transformation and deliver a slightly faster Tuesday, you have burned credibility you will need later. Undersell it. "It saves me a few hours a week on drafting" is both true and believable.

Ask about the fear underneath. Sometimes "it is unreliable" means "will this be used to cut my team?" You may not be able to answer that. Pretending it was not the question is worse than saying so.

If you are the one deciding for a team

Write down three things and circulate them: what people may use it for, what data must never go near it, and who to ask when unsure. One page. Most organisations have no policy, so people either avoid the tools entirely or use their personal accounts on real client data. The second is much worse, and silence produces it.

Where to start on Monday

Pick one recurring task from your lesson-one list. Something textual, low stakes, and annoying. Do it with AI for two weeks. Notice honestly whether it was faster and whether the quality held. Then pick a second one.

That is the whole method. Not transformation. One task at a time, checked, with a clear line about what never goes in the box.

Before you move on

A civil servant uses AI to draft a policy briefing. She verifies every statistic, every date, and every cited regulation against official sources. All check out. Which weakness in her review process remains?

Pick the one you would defend. Nobody sees your answer.

No ads. No data sale. No public scores on people. Ever.

© 2026 Addaly