Why does AI output need a different kind of review?
Reviewing a junior colleague's draft and reviewing an AI draft look similar, but the errors differ. A junior tends to make visible mistakes: a clumsy paragraph, a missing section, a question they flagged. AI tends to make confident ones. The prose reads well, the structure is tidy, and a wrong figure or an invented reference sits inside a perfectly fluent sentence. Polished text lowers the reader's guard, which is exactly when errors slip through.
So the reviewer's job shifts. Less time on grammar and layout, much more on whether each statement is true, sourced and right for this client.
Who is responsible for AI drafted work?
The firm and the named professional who approves it, exactly as if they had written it themselves. Professional bodies such as the SRA, ICAEW, ACCA, RICS and the FCA hold firms and individuals to their standards whatever tools were used. "The AI wrote it" is not a defence a client or regulator will accept.
Our rule is simple: AI drafts, people approve. A named person reviews and signs off anything that goes to a client. Write that into your AI policy and into the process itself.
What should a reviewer check?
Work through the draft with these questions. Adapt the list for your sector and keep it beside the reviewer.
Facts and figures
- Does every number match its source? Check the ones that drive conclusions, not only a sample.
- Are totals, dates and periods consistent throughout?
- Are units, currencies and time periods correct?
Names and references
- Is the client, matter, engagement or job named correctly everywhere?
- Do references to documents, clauses, regulations or standards exist and say what the draft claims? Look them up.
- Are people's names, roles and organisations accurate?
Substance and advice
- Is each conclusion supported by the facts in the draft?
- Would you give this advice yourself, in these words?
- Is anything important missing: a risk, a caveat, an assumption the client needs to know?
- Does anything overstate certainty?
Confidentiality
- Does the draft contain anything from another client or another matter?
- Is there personal data that does not need to be there?
- Has any internal note or comment survived into the client version?
Tone and fit
- Does it read as your firm and in your house style?
- Is it right for this client's knowledge and situation?
- Would you be comfortable if the client asked how it was prepared?
How do you build review into the process?
Review works best when it is designed in rather than left to good intentions. The strongest arrangements share these features:
- Sources travel with the draft. Each figure and key statement links back to the record it came from, so checking is quick.
- Changes are highlighted. For recurring documents, the reviewer sees what differs from the last version, not the whole thing afresh.
- Uncertainty is flagged. Where the draft had missing or conflicting inputs, it says so plainly instead of guessing.
- Sending is blocked until approval. The output cannot go to the client until the named reviewer approves it.
- Approval is recorded. Who approved, and when, is kept with the work. Audit trails covers this in more detail.
- Corrections feed back. Common reviewer changes are used to improve the drafting step, so the same mistake does not recur.
How much review is enough?
Match the depth of review to the consequence of an error, not to how the draft was produced. A routine update confirming a meeting date needs a quick read. A report a client will rely on for a decision, or anything that amounts to professional advice, needs the same full review it always did. Decide the levels in advance, by type of document, so reviewers are not making that call under time pressure.
A practical approach is three levels: a light check for routine correspondence, a standard review for reports and updates, and a full technical review for advice, opinions and anything with regulatory weight. Record which level each document type needs.
How do you stop review becoming the new bottleneck?
If every draft lands on one partner's desk, the time saved in drafting is lost in the queue, a pattern set out in the partner bottleneck. Spread review by competence rather than seniority where your regulator and your clients allow it, keep drafts focused, and measure how long items wait for approval. A good draft should take minutes to review, not an hour.
What does this look like on a real process?
Take monthly client reporting. The draft is assembled from the firm's records, with commentary written from what changed since the last period. The named fee earner receives it with the changes highlighted and sources attached, checks the figures that matter, adjusts the commentary where judgement is needed, and approves it. Only then can it be sent. The fee earner's role moves from building the report to checking it.
Every process we rebuild includes a review step like this. To find the process where it would save your team the most time, start with the audit. For the data protection side of using AI on client files, see data protection and AI on client work.