The draft looks finished. It has clear headings, a confident answer and three links at the bottom. That's often the point where checking becomes hardest, because a polished document invites a quick skim.
Before sharing AI-assisted work, change the question from “Does this sound right?” to “What would I need to prove before somebody relies on this?” A delivery promise, a calculation and a statement about a person each need a different check.
Here is a practical review method for ordinary business documents. It isn't a statutory checklist or a substitute for a qualified reviewer where the subject needs one.
Decide what could go wrong for the reader
Start with the consequence of an error. An internal list of possible newsletter titles needs less checking than a customer quotation or a disciplinary letter.
Imagine an AI draft that summarises three supplier proposals. If you only want ideas for questions to ask, rough notes might help. If the document will decide who gets a contract, check every claim used in that comparison against the proposals themselves.
Give the reviewer enough time and access to do that work. Someone who can only see the final paragraph can't reliably approve a summary of documents they haven't read.
Separate facts, calculations and judgement
Mark the parts of the draft that can be checked directly. These commonly include:
- Names, dates and contact details.
- Prices, quantities, tax treatment and totals.
- What a contract or policy actually allows.
- Statements attributed to a person or organisation.
- Claims about current law, guidance or software behaviour.
Then mark recommendations and assumptions separately. “The supplier quoted a six-week lead time” is a checkable fact. “This is our best option” needs an explanation of the criteria and the alternatives.
If the draft mixes those together, rewrite it before approval. Readers should be able to tell which parts are established and which depend on a judgement your business is making.
Open the sources and check the claim
A link is a starting point. Open it and find the passage that supports the statement. Check that the source concerns the right product, jurisdiction, date and version.
Microsoft's Copilot guidance treats generated content as a draft that needs review before sharing. Use that principle with other tools too, while checking their own instructions and limitations.
Be particularly careful with a source that supports only part of a sentence. An article explaining that some organisations must follow a rule doesn't establish that every small business must follow it.
If a link doesn't open, find the original publisher through a separate search or remove the unsupported claim. Don't ask the tool to invent a replacement citation. A second answer from the same model doesn't establish that the first one was correct.
Recalculate numbers outside the draft
Use the original figures and an ordinary calculator or a checked spreadsheet formula. Confirm units and rounding as well as arithmetic.
For a fictional quotation, suppose a draft lists 18 items at £47.50 each and a delivery charge of £24. Before discussing any applicable tax, the combined amount is £879. A reviewer should reproduce that result from the quantities and prices, then check what the actual quotation requires next.
Pay attention to changes in meaning. A monthly price and an annual price might both be copied accurately but compared unfairly. A percentage saving needs the right starting value. A deadline needs the right time zone and any conditions that affect it.
Check what the summary left out
Read the original conclusion and any exceptions, qualifications or unresolved points. Then compare them with the summary.
An imagined customer message might say a replacement arrived, but that the customer still wants the installation cost addressed. “Customer received a replacement” is accurate yet incomplete if someone uses it to close the complaint.
The ICO's position on accuracy in generative AI links appropriate accuracy to the purpose of the application. Where personal information is involved, consider whether the output misrepresents someone and how an error will be corrected.
Look for words such as “always”, “confirmed”, “agreed” and “must”. They deserve a closer check whenever the source was tentative or conditional.
Give approval to someone who can change the result
The reviewer needs authority to reject the draft, ask for evidence or use the ordinary process instead. A hurried tick box doesn't improve an answer nobody has understood.
For specialist material, send it to someone competent in that subject. AI can help prepare a question for your accountant or solicitor. It shouldn't replace their judgement simply because the draft resembles professional advice.
Under the EU AI Act, Article 14 concerns human oversight of high-risk systems. It isn't a universal rule requiring the same formal process for every writing assistant. This article's review steps are practical suggestions, with legal duties assessed separately for your use.
Make the final check about the recipient
Confirm that the document contains only information the recipient should receive. Remove hidden comments, unnecessary personal details and internal instructions. Check attachments and working links as well as the main text.
Record approval in the place you already manage the work. For an important document, keep the source version and the corrected final copy so a colleague can understand the decision later.
Use recurring errors in your AI training exercises. If a tool repeatedly turns uncertain delivery dates into promises, teach staff to look for that specific mistake. A short example from the task they actually do is more useful than another reminder to “be careful”.