← All Insights

Practical Guide

How to check an AI draft before sending it to a client

To check an AI draft before sending it to a client, do five passes: source facts, numbers, confidential material, commitments and final ownership. The writing can sound ready while one unsupported sentence changes what your firm appears to promise. A quick review works better when each pass asks a different question.

Picture a fictional Hong Kong account team preparing a renewal proposal on Thursday. The brief says the current service ends in December, the proposed scope is still under review, and a pricing table sits in an approved spreadsheet. An AI draft calls the scope "agreed", lifts last year's fee and thanks the client for a meeting that has not happened. None of those errors needs a complicated model failure. The draft simply filled gaps with plausible prose.

How do you check an AI draft before sending it to a client?

1. Trace each factual claim to the source. Highlight dates, names, product descriptions and statements about what the client said. Find each one in the brief, email thread or approved document. If you cannot point to a source, delete the claim or ask the person who owns it. Do not ask the same model to certify its own answer. For a broader method, see my guide to checking AI for invented facts.

2. Recalculate the numbers. Compare every fee, percentage, currency and period with the controlled spreadsheet. Check whether a figure is a quote, an estimate or a prior-year number. In the fictional proposal, copying last year's fee could turn an unfinished commercial discussion into an accidental offer. A correct number with the wrong label can be just as damaging as an invented one.

The dangerous line is often the one nobody notices because it sounds like something you would say.

3. Check what went into the tool and what remains in the draft. Was the source material cleared for the tool used? Does the output contain a client's personal details, another customer's example or internal commentary that should not leave the firm? The Hong Kong Privacy Commissioner's checklist for employee use of generative AI recommends defining permitted tools, input information and use of outputs in an internal policy. It is a control to apply before drafting, as well as a check before sending. My work-data guide walks through that boundary.

4. Circle every promise. Words such as "will", "guaranteed", "included" and "agreed" deserve a second look. Compare the draft with the signed scope, approved terms and actual delivery plan. In the example, "agreed scope" must become "proposed scope" until the client and your team have agreed it. This is not a grammar edit. It changes the commercial meaning of the document.

5. Name the final owner. The person sending the proposal checks whether it answers the client's real question, uses the right tone and carries the right approvals. A colleague with authority over pricing or scope may need to sign off those parts. Put the owner in the workflow, rather than treating "human review" as a box that somebody else will tick.

A review prompt that helps without pretending to approve

Use AI to make the review easier, but give it the source and a narrow job. For example: "Compare this draft with the approved brief. List every factual claim, number and commitment. For each, quote the supporting source passage or mark it unsupported. Do not rewrite the proposal." Run this only in a tool your firm permits for that material. The free prompt library can help you shape the instruction on fictional or cleared inputs.

Then do your own five passes. A model can miss an omission, misread a spreadsheet or accept a bad premise in the brief. It also cannot know the nuance of a client relationship from the documents you chose to provide. The PCPD's AI model framework treats human oversight as a key risk control where AI systems affect people. The same habit is useful here even when the document is a routine proposal: a named person remains accountable for what leaves the firm.

The goal is a faster first draft with a stronger final check. I help professionals put that review step inside their actual client workflow, so time saved on writing does not turn into time spent repairing a mistake after Send. If your team produces the same kind of client document every week, let's examine the hand-off together.

Make the final check part of the workflow.

If your team sends the same kind of client document each week, we can define the sources, checks and owner together.

Request a Private Consultation