Use an AI Assistant on Client Work Without Sounding Like an AI

You have accepted the tool. The problem now is not whether to use it — it is that the client can tell. The even pacing, the em-dash tics, the sentences that all weigh the same: whatever you feed a capable model, what comes back sounds like a capable model, and your client reads a dozen of those a day.

The fix is not to stop using it. The fix is to use it on the parts where its voice is an asset, and keep the parts where your voice is the product.

The line you should never cross

The rule is simple and it is the whole system: the model drafts, you decide. You are the one who chooses what the client sees, what the document commits to, and what the numbers mean. A model that drafts from your notes is a tool. A model that drafts your decisions is a replacement — and the client can tell, because the document that comes back is the one you would have written for anyone.

This is also the legal and professional line. The value you sell is judgment, not typing. Typing was never the scarce part.

What the model is actually good at

Three jobs benefit from the machine voice, because nobody reads them for voice:

• Turning messy notes into a first draft. Your call notes, your meeting transcript, the half-sentence list you made at the end of the day. The model converts chaos into structure; you then cut everything it got wrong.
• Writing the boring middle. Status updates, follow-up sequences, the routine email that exists in a thousand similar forms. These are form documents; the form is fine.
• Reframing a paragraph you are stuck on. Give it your draft and one instruction — "make this less defensive" — and use the output as a mirror to see what your own instinct was saying.

Notice what all three have in common: they produce raw material, not final copy.

What you should never hand over

These are the parts where your voice is the deliverable:

• The actual advice. What to do, why now, what it will cost. The model never knows the client's real constraints; only you have seen the meeting.
• The pricing. Number, rationale, and the moment you say it. A model can write you a script; it cannot hold the pause after the number.
• The bad news. The slip, the overrun, the changed date. The client is buying honesty, and honesty is not a template.
• Anything that commits you. Scopes, guarantees, deadlines, definitions of done. If the model wrote it and you did not verify every word, it is a promise you did not make.

The test for any sentence: if the client read it aloud to a room, would they be describing you, or describing a tool? If the latter, rewrite it by hand.

The three-pass workflow that keeps your voice

Most people using AI do it in one pass: paste, copy, send. The workflow that keeps the voice is three passes, and it takes five extra minutes.

Pass one — the model drafts. Give it your notes and the structure. Do not polish the input; the point is to see what a competent stranger makes of them. What surprises you here is usually the model seeing a gap you left — that is the value.

Pass two — you cut. Delete every sentence you did not need to say, every point you would have made differently, and anything you do not believe. Cutting is where the voice comes back. Your draft is your draft again the moment it no longer matches the model's cadence.

Pass three — you add the specifics. Insert the one number only you know, the client's own phrase from the last call, the constraint you have been carrying. One specific changes the entire register of a document. Readers cannot articulate why, but they can tell the difference between a document that knows them and one that is merely correct.

The tells, and what to do about each

If your writing reads like AI, the cause is usually one of five, and each has a fix:

Run the check once and it takes three minutes. Run it every time and it stops being a check — it becomes how you write with a tool.

The honest use

There is a version of this conversation happening in every consulting practice right now: how much do I admit to the client? The answer that survives contact with the work is the one that makes it a non-issue: you are not evading the question, you are running a different process from the start.

When it is worth saying, say it plainly: "I draft with an assistant, and everything that commits us to anything is mine." Clients do not punish honesty about tools; they punish the feeling that the work was out-sourced. The first sentence is a process detail. The second is a promise, and it is the one that matters.

This note describes a workflow. No model benchmarks and no accuracy rates are claimed; the tell-and-fix table is an experienced-based checklist, not research.

More field notes on running a one-person consultancy: consultant-stack.netlify.app