How to check AI output before it reaches a client: a checklist

I'd check AI-assisted work in the same way you'd check a capable junior's first draft, but with extra suspicion in four places: the numbers, the sources, anything that drifts from the brief, and confident statements nobody can back up. The responsibility doesn't move to the tool. If the work goes to a client with your firm's name on it, any error is yours.

Why does checking matter more now?

Because AI's errors look right. It rarely makes typos; it makes plausible mistakes. In June 2025 the High Court in London ruled on two cases where court documents cited cases that didn't exist. In one of them, 18 of the 45 cases cited were fictitious. The citations had come from the client's own AI research, and the solicitor hadn't checked them. Dame Victoria Sharp, President of the King's Bench Division, said lawyers have "a professional duty… to check the accuracy of such research by reference to authoritative sources". The court also put managing partners on notice about their own responsibilities.

And it isn't only lawyers. In 2025 Deloitte refunded part of an A$440,000 fee to the Australian government, after a report it wrote was found to include a fabricated quote from a court judgment and references to academic papers that don't exist.

Specialist tools help, but they don't remove the need to check. When Stanford researchers tested commercial legal research tools in 2024, the tools gave false or unsupported answers between about 17% and 34% of the time. The tools have been updated since, but the lesson is the same. Fewer errors isn't the same as no errors.

Why do people stop checking?

Partly it's confidence. A 2025 study by Microsoft Research and Carnegie Mellon University, surveying 319 knowledge workers, found that the more confidence people had in generative AI's ability at a task, the less critical thinking they reported applying to its output.

Partly it's the uneven edge of what AI can do. In a 2023 Harvard Business School and BCG experiment, 758 consultants were randomly assigned to work with or without GPT-4. Those using it did much better on tasks within its abilities. But on a task just outside them, they were 19 percentage points less likely to reach the right answer than consultants working without it, and nothing in the output warned them which kind of task they were on.

Unchecked output also costs colleagues time. Researchers at Stanford and BetterUp came up with the term "workslop" for AI-generated work that looks finished but isn't. In their 2025 survey of US desk workers, about four in ten said they'd received some in the past month, and each instance took nearly two hours to deal with.

What should be on the checklist?

This is what I'd run through before anything AI-assisted leaves the firm.

1. Numbers. Trace every figure back to its source, and recalculate totals and percentages yourself.
2. Sources and quotes. Open every citation, link and quotation, and check it exists and says what the draft claims it says.
3. The brief. Read the draft against what the client actually asked for. Look for what's missing, what's been added, and any questions answered that nobody asked.
4. Names, dates and details. Client names, people, products, dates and jurisdictions are where plausible errors tend to hide.
5. Unsupported confidence. Mark any sentence that states something firmly without evidence, and either support it or soften it.
6. Judgement calls. Recommendations and conclusions need a person's reasoning behind them, not just fluent wording.
7. Client data. Confirm the work was done in a tool the firm has approved for client information.
8. A record. Note who checked it and which tool was used, so you can answer if a client asks.

Who is responsible for the check?

The person who sends the work, and the person who supervises them. UK professional guidance is consistent on this. The Law Society tells solicitors to "carefully factcheck" what generative AI produces and to "authenticate the outputs". The Bar Council's guidance says "the ultimate responsibility for all legal work remains with the barrister".

I think the same principle applies in any firm that sells its expertise. Clients are paying for your judgement, so checking is part of the work, not an extra step at the end.

How do you know your people can check well?

Checking is a skill, and in my view it varies a lot more than most firms expect. The most reliable way to see it is to test it on realistic work, for example a draft with errors you already know about, and see what people catch. Asking people how good they are won't tell you much, as confidence in AI tends to go with less checking, not more.

If you'd like to see where you stand yourself, the free Osava challenge takes about five minutes and includes checking AI's work. For firms, checking AI output is one of the things Osava measures, alongside cost, time and quality. You can read more in how to measure the ROI of AI.