Before you send anything an AI tool helped you write, run a short check: confirm it answers the request, verify every fact and number against your own source, remove details you did not supply, check the tone for the reader, and make sure no confidential data slipped in. A short email takes a few minutes.

This article gives you that routine as an eight-step checklist, explains why each step matters, and shows how to make it a habit. It applies to emails, reports, summaries and presentations drafted with ChatGPT, Gemini, Claude, Copilot or similar tools. It does not cover reviewing AI-generated code or images.

Why do you need to check AI output at all?

Because AI assistants generate text that sounds right, and sounding right is not the same as being right. A fluent paragraph can still contain a wrong date, a made-up reference or a figure that was never in your notes.

The tool does not carry responsibility for what you send. You do. Your manager, client or professor will judge the email by your name at the bottom, not by the tool that drafted it.

Checking is also a core skill, not an optional extra. IBAIC’s Certified AI Practitioner (L1) blueprint gives Evaluation & Assurance a 15% weighting, as a domain in its own right.

What is the routine, step by step?

Run these eight checks in order. The early ones catch the biggest problems, so if you are short of time, do not skip them.

  1. Does it answer the actual request? Re-read the original email, brief or question. Then read the draft. Check that it answers what was asked, not a nearby, easier question.
  2. Is every fact and figure right? Mark every number, date, name, price, percentage and quote. Check each one against your own source: your notes, the spreadsheet, the original document. Do not ask the AI to confirm its own facts.
  3. Did it add anything you did not give it? Look for details that were not in your notes: a deadline, a reason, a statistic, a “recent study”. Delete them or verify them from a reliable source.
  4. Do the references and links exist? If the draft cites a report, law, website or person, open the source yourself. Check it exists and actually says what the draft claims.
  5. Is the maths right? Recalculate any totals, percentages and comparisons yourself or in a spreadsheet. Do not trust arithmetic just because it is neatly formatted.
  6. Is the tone right for this reader? Read the draft once as the person receiving it. Is it too formal, too casual, too long? Does it promise anything you cannot deliver?
  7. Is there anything that should not be shared? Check for customer details, internal figures or confidential information that should not go to this reader. A draft can repeat details you pasted in only as background.
  8. Can you defend every sentence? If someone asked “why did you say this?”, could you answer for each line? If not, rewrite or remove that line.

When all eight pass, send it.

How long should this take?

For a short email, a few minutes. For a report of a few pages, set aside longer, because steps two to five scale with the number of facts.

A useful rule: the more people who will read it, and the more it matters, the slower you go. An internal note to one colleague needs a quick pass. A client proposal or anything with figures needs the full routine, slowly.

Which mistakes slip through most easily?

The ones that look plausible. A wrong total in a neat table, a confident date, or a polished sentence that quietly changes your meaning are harder to spot than an obvious error.

Watch especially for:

  • Invented specifics. Exact numbers, names, dates and sources that were not in your input.
  • Softened or strengthened claims. “We may be able to” turning into “we will”, or the other way round.
  • Missing caveats. Your notes said “subject to approval” and the draft dropped it.
  • Wrong spelling conventions. American spellings in a document meant for an Indian reader, or the reverse.
  • Over-long answers. A two-line question answered with five paragraphs.

How do you make checking a habit?

Attach it to a moment you never skip. For most people that moment is just before pressing send.

Some practical ways:

  • Keep the checklist visible. A sticky note on your monitor or a pinned note on your phone with the eight steps.
  • Ask for the right input up front. Prompts that tell the tool to “use only these facts” and “list any assumptions” make step three much easier. See A simple prompt structure for reports and emails.
  • Highlight as you read. Mark every fact in one colour as you check it. Anything unmarked at the end has not been checked.
  • Let it sit. For important documents, step away for a few minutes before the final read.

Try it: Take one AI-assisted email you sent recently. Run the eight steps on it now. Note which step, if any, would have changed what you sent.

Should you say that AI helped you write it?

Follow your organisation’s policy, your college’s rules, or the client’s contract. If there is no policy, ask your manager or professor rather than guessing.

Whatever the policy, the responsibility for accuracy stays with you. Disclosure does not replace checking.

What should teams agree on?

If several people in a team use AI for drafting, agree on a shared minimum check, and on which documents need a second human reviewer. This avoids each person inventing their own standard.

A simple team agreement covers:

  1. Which documents can be AI-drafted at all.
  2. The minimum checks before anything goes outside the team (the eight steps above are a reasonable start).
  3. Which documents always need a second reviewer, such as anything with prices, legal terms or customer data.
  4. What data may and may not be pasted into which tools. For this, see Keeping company data safe when your team uses AI tools.

What to do next

Copy the eight steps somewhere you will see them, and run them on the next AI-assisted draft you send. If you want your team to build these habits together, see the path for teams on the learn page.