Checking the output before you use it

Hardly anyone has time to check every word of every AI output. The good news is you do not need to. What you need is an order: check the parts that hurt most if wrong first, then move down to lighter parts until you run out of time or risk.

The order we use has four layers.

Layer one, things that bind: money amounts, quantities, dates, deadlines and commitments. "Price includes tax", "we deliver by Friday at the latest", "the discount applies to every branch". Sentences like these can become obligations once sent. One wrong figure in a quotation costs more than ten awkward sentences.

Layer two, names and references: people's names, job titles, company names, document numbers, rule titles, links, quotes. Getting a client's name wrong makes the reader stop trusting the rest. An invented reference makes you look as if you did not check.

Layer three, facts about the world: "this rule took effect in ...", "the market grew by so many percent", "most customers prefer ...". This is where hallucination hides most often: it sounds reasonable with nothing behind it.

Layer four, flow and tone: does the order make sense, is it polite, does it suit the reader. Important, but mistakes here rarely cost as much as the three layers above, and they usually show as soon as you read.

One thing is easy to miss in every layer: what is not there. A model can drop an important condition from a contract it summarises, or forget an exception in a policy. When checking a summary, also ask "what should be here that does not show up?".

For important figures, re-reading is not enough. How to recalculate a figure by another method is covered in the lesson "Check the numbers and measure the result" in the Produce real work module. Here, just remember that figures always belong to layer one.

AI can help with checking, as long as its role is clear: it helps find what needs checking; it does not decide what is true.

The way to do it is to ask the model for a list of claims. A claim is any sentence that can be true or false: a figure, a name, a date, a cause and effect, a commitment. For each claim, ask it to note its type, where it sits in the text, whether it came from the material you gave or not, and how much harm it would do if wrong. A list like this helps because people easily miss a claim tucked into a long sentence, while a model is diligent enough at breaking text into small pieces.

Two things make the list more honest. First, run it in a new conversation, not the one that wrote the draft. A model tends to defend its own writing. Second, do not ask the model to declare claims "true" or "verified". Ask it only to flag them and give a risk level. The matching against sources is still yours, starting with high-risk claims.

This is also a small evaluation: you test the output in a structured way, not on the feeling that "it looks right".