Checking AI Output: A Verification Routine for Professional Work
Check AI output in proportion to what is at stake, and always against sources the AI did not produce. A draft email to a colleague needs a careful read. Anything that goes to a client, a regulator, a court or a patient needs more:
- every fact, figure, quotation and citation checked against a source you have opened yourself;
- every calculation redone;
- the output compared with the original for anything it left out.
Generative AI fails in a particular way: its mistakes look like correct work. That is why checking has to be a routine, not a mood.
This article covers verification for professionals who use AI in their own work. For what happens when a customer-facing chatbot gets something wrong, see when your chatbot gives a wrong answer.
The short answer
- Match the depth of checking to the stakes: internal drafts, decision support, external communication, regulated professional work.
- Make the output checkable: ask for quotations with their location, a list of sources, and calculations shown.
- Open every source yourself. A citation is verified only when you have read the cited passage.
- Recompute every number that matters, outside the AI tool.
- Check what is missing, not only what is there.
- Record who checked what, so the work can be defended later.
Why plausible is not the same as correct
The International AI Safety Report 2026 states that “current AI systems sometimes exhibit failures such as fabricating information, producing flawed code, and giving misleading advice.” The OWASP Top 10 for LLM applications lists misinformation among the main risks of building on these models.
The courts have put it more sharply. In June 2025, the High Court of England and Wales dealt with two cases in which lawyers had put fictitious authorities before the court. In one, eighteen of the forty-five citations did not exist. The judgment warned that AI tools’ “coherent and plausible responses may turn out to be entirely incorrect”. It described a “professional duty therefore to check the accuracy of such research by reference to authoritative sources”. The case was about lawyers, but the principle applies to any profession that signs its work.
Match the check to the stakes
| Level | Examples | Minimum check |
|---|---|---|
| Internal draft | Notes for yourself, first drafts, brainstorming | Read it through; nothing leaves your desk unread |
| Decision support | Summaries for a manager, analysis behind a decision | Key facts and figures checked against the source; AI assistance noted |
| External communication | Client emails, proposals, website content | All facts, names, prices, dates and commitments checked; tone read as the recipient |
| Regulated professional work | Advice, filings, tax returns, clinical notes, contracts | Every claim traced to a primary source you opened; calculations redone; a qualified person signs off |
The errors AI tends to make, and how to catch them
| Error | What it looks like | How to catch it |
|---|---|---|
| Invented sources | A case, regulation, article number or report that does not exist | Open every cited source in an official or original location |
| Misquoted sources | A real source cited for something it does not say | Read the cited passage, not just the title |
| Wrong numbers | Arithmetic slips, wrong units, totals that don’t add up | Recompute in a spreadsheet; check units, currency and dates |
| Omissions | A summary that drops the inconvenient clause | Compare against the source’s structure: every section, every obligation |
| Outdated information | Rules or prices that have since changed | Check the date of every source and the current version of any law |
| Mixed-up entities | The right figure for the wrong customer, party or period | Check names, references and periods against the record |
| Overconfident commitments | Promises, guarantees or deadlines nobody approved | Read as the recipient; delete what you would not sign |
| Translation drift | A meaning that shifted between languages | Have a bilingual reviewer compare the key sentences |
A six-step verification routine
1. Ask for checkable output. Before you check anything, change what you ask for. Request quotations with page or clause numbers, a list of sources, and calculations shown step by step. Tell the model to write “not stated” where the source is silent. Prompt engineering for business covers how.
2. Open every source. A citation is verified only when you have found the source in an authoritative place — the official gazette, the regulator’s website, the original report, your own records — and read the passage it is cited for.
3. Recompute the numbers. Redo every figure that matters outside the AI tool: totals, percentages, tax, conversions, dates and deadlines. Check that units and currency are consistent throughout.
4. Check completeness. Compare the output with the structure of the source. If a contract has twelve obligations and the summary lists nine, find the other three. Asking the model “what in the document is not covered by this summary?” can help, but it is no substitute for looking.
5. Read as the recipient. Would a client, a regulator or a patient read anything here differently from how you intend? Check names, promises, tone, and anything confidential that should not be there.
6. Record the check. Note who checked the output, what was checked against what, and which tool produced the draft. For important work, keep the prompt and the output with the file.
Techniques that help, and their limits
- Self-review by the model. Asking a model to critique its own answer catches some errors, but it shares the same blind spots. Treat it as a first filter.
- A second model. Cross-checking with a different model can surface disagreements worth investigating. Agreement between two models is not proof.
- Sampling for volume work. When AI extracts data from hundreds of invoices, check a random sample plus every item above a value threshold, and widen the sample if errors appear.
- Validation rules. In systems, let code check what code can: totals match line items, VAT is calculated correctly, dates fall within the period. People then check what rules cannot.
- Self-reported confidence. A model saying it is “95% sure” is not a measure of accuracy. Rely on sources, not stated certainty.
Beware automation bias
The World Health Organization’s January 2024 guidance on large multi-modal models in health warns that they can encourage “‘automation bias’ by health care professionals and patients, whereby errors are overlooked”. The risk is not limited to medicine. The better an AI tool usually is, the easier it becomes to stop checking it.
Three habits counter it:
- Form your own view first on the key question, then read the AI’s answer.
- Rotate reviewers on important work, so checking doesn’t become a formality.
- Test the checkers. Occasionally put a deliberately flawed draft into the review queue and see whether it is caught.
What verification looks like by profession
| Profession | The check that matters most |
|---|---|
| Accountants | Figures reconcile to the ledger; tax treatment checked against the law and Tax Authority guidance. See AI for accountants. |
| Lawyers | Every authority opened in an official source and read in full. See AI for lawyers. |
| Clinics | The note matches what happened in the consultation; drug names and doses checked. See AI for doctors and clinics. |
| HR | Nothing in a shortlist or assessment rests on a protected characteristic. See AI in HR and recruitment. |
| Engineers | Standards, specification clauses and quantities checked against the source documents. See AI for engineering and construction firms. |
Building verification into systems
When the same check is done hundreds of times, it belongs in the system: validation rules, approval steps, and a log of who approved what. From prompts to workflows explains when a repeated prompt should become a tool, and who can see what: access control and audit trails covers the logging.
Frequently asked questions
How can I tell if AI has made something up? Often you can’t tell by reading, because invented content looks plausible. Open every cited source in an authoritative location and read the passage it is cited for. If you can’t find it, treat it as invented.
Is it enough to ask the AI to check its own work? No. Self-review catches some errors but shares the model’s blind spots. Use it as a first filter, then check facts, figures and sources yourself.
How much time should checking take? It depends on the stakes. Internal drafts need a read-through; regulated professional work needs every claim traced to a source. If checking takes longer than doing the work yourself, the task may not suit AI, or the prompt needs to ask for more checkable output.
Which AI outputs need the most careful checking? Anything with citations, figures, legal or medical content, or commitments to a third party, and anything a professional will sign or a client will rely on.
Should we tell clients when AI was used? Follow your profession’s rules and your client agreements. Some courts already expect it: the DIFC Courts’ guidance on AI in proceedings says parties “should declare at the earliest possible opportunity”, and should not rely on AI-generated content without first verifying its accuracy.
Can verification be automated? Partly. Validation rules, reconciliations and citation lookups can be automated in a system. A qualified person should still review anything that carries professional or legal responsibility.
This article reflects the court guidance and sources available on 28 October 2026, the date of publication. It is not legal advice.
Sources: International AI Safety Report 2026 — executive summary · Ayinde v Haringey and Al-Haroun v Qatar National Bank — High Court judgment, 6 June 2025 · WHO — Guidance on large multi-modal models in health · DIFC Courts — Practical Guidance Note No. 2 of 2023 · OWASP — Top 10 for LLM Applications
Want verification built into your AI tools rather than left to memory? ZAWAT builds AI and automation with validation rules, approval steps and audit trails designed in. Book a call.