AI Writing for Consultants: Client Comms That Still Sound Like You

TL;DR: AI writing for consultants breaks at the first client email, and it breaks in a specific way. The draft is correct, polite and reads like it was written by anyone. That happens because "professional but warm" is a tone label, and a tone label is not a voice. Your voice lives in sentence length, qualifier placement, transitions and punctuation, and your clients have learned all four from months of your emails. AI does not know your tone; it knows words that describe tone. The fix is to read your own habits off five past client emails, write them as rules, and put the rules where the model reads them on every request. A prompt describes how you hope to sound. A profile built from your own writing is evidence of how you actually write, and each new model executes that evidence better than the last.

You paste the week's progress into ChatGPT, ask for a client update, and get one in six seconds. Correct structure. Right level of politeness. Reads like anyone's.

Your client has read forty of your emails. He knows you open with what changed, not with "I hope this finds you well". He knows you put the risk before the ask. This draft does neither, and he reads that before he reads the update.

Why AI Writing for Consultants Fails on the First Client Email

The consulting relationship runs on a small number of documents that the same person reads over and over: a status update every week, a proposal cover note at the start, a difficult-news email two or three times per engagement. The client learns the writer from these, in a single register, at a predictable cadence, and the learning is not conscious. Nobody analyses a consultant's style. They simply know it, the way you know a colleague's footsteps in a corridor.

That is what makes the consultant the hardest case for AI drafting. An internal memo is read by twenty people who each see one of your documents a quarter. A client update is read by two people who see one of your documents a week. The audience is small, attentive and trained on you. When a draft arrives with the same content and a different gait, the reaction is not "this was written by AI". The reaction is "something is different with this engagement", and for a consultant that is the worst thing a client can feel between invoices.

The pattern in consultants who try AI drafting and stop is always the same: the draft comes back correct, and they rewrite it anyway, because correct was never the problem. It was not theirs. The consultant profile page shows what the alternative looks like.

What a Client Actually Recognises in Your Writing

A client recognises placement and rhythm, not vocabulary. Ask a client what their consultant's emails sound like and you get an adjective: direct, careful, thorough. Ask what they noticed in the AI draft and they cannot say, because what they noticed is a set of positions and counts they have never put into words. Here is what a consultant finds when comparing a real sent email against an AI draft of the same update:

Every row is a placement or rhythm choice. None of them is a grammar error. A client cannot tell you which row bothered him, and he does not need to. He knows the email is not you.

This is the difference between a tone and a voice, and it is why adjectives are not your voice. A voice lives in sentence length, qualifier placement, transitions and punctuation. A tone is a word someone else uses to describe the effect. The model only ever receives the word.

Why Doesn't "Professional but Warm" Fix It?

Because every common fix is a better description, and a description carries almost no information about you. AI does not know your tone; it knows words that describe tone. Three fixes get tried in roughly this order, and each fails for the same reason.

A longer adjective list. "Direct, structured, warm but not casual, no jargon, client-friendly." Each adjective narrows the register a little and describes you not at all. Two consultants can both be direct and structured and differ on every countable feature: one opens with the change, the other with the decision needed; one hedges with "on current scope", the other with "assuming no change to the brief"; one writes a six-word recommendation, the other a twenty-word one with a fallback. The adjective list fits both. The model picks the middle.

One example email pasted in. This is closer, and it fails in a subtler way. The model imitates the surface of the example on that request and reverts on the next one. Worse, one email is one occasion. If the example was a good-news update, the model learns to open with the win even when the news is a slipped milestone and you would have opened with the plan. Style in the moment is not a profile. A profile is what stays constant across forty emails of different weather, and a single sample cannot show that.

The register all three produce is measurable. Our Average AI baseline, built from 320 samples across five models and four languages, puts English AI output at a formality score of 58 and a consistency score of 53. Formality in our scale weighs hedge frequency per thousand words, function-word density and semicolon use. Consistency measures how evenly sentence length is distributed. A score of 53 means moderate, uniform variation: no five-word verdict, no thirty-word sentence carrying three conditions. How we built that baseline has the method. The short version is that the model writes the middle of the distribution, on every request, for every consultant.

Three Client Documents, Three Registers

The same consultant writes the status update, the proposal and the bad-news email differently, and a single tone label averages all three. A profile that says "direct and warm" makes the proposal too blunt and the bad-news email too cheerful, because it has one setting for three jobs. Read your own sent folder and you will find something like this:

• Weekly status update. Shortest sentences of the three. Bullets for what shipped, one paragraph of judgement, a dated next step. Almost no hedging, because the facts are the facts.
• Proposal cover note. Longest sentences. More qualifiers, placed early ("based on the two workshops so far"). The recommendation is still short, but it arrives later, after the framing.
• Difficult news. The plan before the problem, or the problem before the plan, but always the same order for you. One explicit acknowledgement, never two. A concrete date in the last line.

These are three registers of one voice. The words change, the shape does not. We wrote about the mechanics of holding one voice across contexts in writing style and context switching, and client work is the sharpest case, because the client sees all three documents and expects the same person behind them.

A Rule Block for Client Comms

A rule block is a set of countable instructions read from your own emails, written so that every line can be checked against the output. It replaces the adjective list. Here is one built from a real consultant's sent folder, with the specifics changed:

CLIENT COMMS RULES

Status updates:
- Sentence one states what changed since the last update. No greeting line.
- Shipped items as bullets, max 6 words each. Judgement as prose, one paragraph.
- Risk before ask. If a decision is needed, the risk it resolves comes first.
- One hedge per timeline, phrased "assuming [condition] by [date]".
- Recommendation sentence under 12 words.
- Close with one next step and a date. Never "let me know if you have questions".

Proposal cover notes:
- Sentences may run to 25 words. Qualifiers go in the first half of the sentence.
- Framing paragraph first, recommendation second, price and scope last.
- One question to the reader, in the final paragraph.

Difficult news:
- Plan first, problem second. One acknowledgement sentence, never two.
- No exclamation marks. No "unfortunately". Name the owner of the fix.
- Final line is a date.

All documents:
- No semicolons. No "I hope this finds you well". No "please find attached".

Every line can be checked against the output. Every line came from something you already sent. That is the difference between a prompt and a profile: the prompt is a hope, the profile is a record.

How Do You Set Up AI Writing for Consultants in 20 Minutes?

Spend twenty minutes once on the profile, then two minutes per email. The samples already exist in your sent folder, and five of them are enough.

Step 1: pull three status updates, one proposal cover note and one difficult-news email (5 minutes). Choose updates from different engagements, ideally one good week and one bad week, so the profile captures what stays constant when the news changes. Redact client names and figures if you are pasting anywhere outside tools your clients have approved; the profile measures how you write, and the content is irrelevant to it. How to select writing samples covers the general rules; for client work the specific rule is that an email you sent as written counts more than one you rewrote four times with a colleague.

Step 2: count, do not describe (10 minutes). Go down the table in the second section and write a number or a phrase next to each row. Not "I am direct" but "change stated in sentence one in 3 of 3 updates". Not "I hedge carefully" but "one hedge per timeline, always 'assuming sign-off by'". You will be surprised by at least one count. Most consultants are surprised by how short their recommendation sentence is, and by how often they close with a date.

Step 3: write the counts as rules (5 minutes). Use the block above as a template. Then place it where the model reads it on every request: ChatGPT custom instructions or a Project, a Claude Project, a Gemini Gem. ChatGPT Projects for email voice walks through the placement. Placing it once is what makes it a profile rather than a prompt; the rules apply to next Friday's update without you remembering to paste them.

Step 4: draft with the rules, then run the five-minute check below. Give the model the week's facts, the risk and the decision you need. When the draft comes back, put it next to your most recent real email to the same client.

The Five-Minute Check Before You Send

Four comparisons tell you whether the rules are in place. Run them on the draft and on your most recent real email to the same client:

1. Circle the first sentence in each. Does it state what changed, or does it greet?
2. Find the risk and the ask in each. Same order?
3. Count the hedges on the main timeline in each. Same count, same phrase?
4. Count the words in the recommendation sentence in each. Within three words of each other?

If the draft fails two of the four, the rules are not in place, or the model is not reading them. If it passes, your client reads an update with this week's facts and your gait, which is the only outcome that matters. ChatGPT for business leaders covers the wider set of documents where the same profile applies: the steering committee note, the end-of-engagement summary, the two-line email that unblocks a decision.

Why a Better Model Widens the Gap Instead of Closing It

The expectation is that the next GPT or the next Claude will simply know how a good consultant writes. It will know that better than the last one did, and it still will not know how you write. Model capability and user-data conditioning are separate layers. The lab improves the first with every release and cannot supply the second, because the second comes from your emails and the lab has never seen them. Given a vague adjective, a stronger model produces a smoother average. That is why better models still don't know your voice, and why the distance between a generic draft and yours grows rather than shrinks as models improve.

The same separation works in your favour once the profile exists. "Risk before ask, one hedge per timeline, recommendation under twelve words" is a nuanced instruction. Each new generation follows nuanced instructions more faithfully, so each one executes your profile more accurately than the last. OpenAI's own custom instructions documentation describes the mechanism as persistent preferences the model applies to every response, which is precisely the slot a rule block belongs in. The model is the layer that improves on its own. Your profile is the layer only you can supply, and it is the one your client is listening for.

MyWritingTwin does the counting. A Writing DNA analysis measures your sentence rhythm, hedge placement, formality and punctuation from the client emails you already have, against the 320-sample Average AI baseline, and turns the measurements into a Style Profile you carry to ChatGPT, Claude, Gemini or whatever your clients' policies allow. The Pro tier builds separate Communication Matrix sections for the status update, the proposal and the difficult-news email, because the same consultant writes those three differently and a single profile should not average them.

Frequently Asked Questions

Can consultants use AI for client emails without the client noticing?

Yes, if the model is working from your own habits rather than from a tone label. A client who has read six months of your emails has learned your shape: where the recommendation sits, how you hedge a timeline, how you close. AI reproduces that shape only when you hand it the measurements. Given "professional but warm" it produces the average of every consultant it has read, and a client who knows you will notice inside two paragraphs.

Why does my AI client update read like a template?

Because the instruction was a template. "Professional, clear, client-appropriate" describes thousands of consultants, and the model resolves it to one learned register that is identical for all of them. AI does not know your tone; it knows words that describe tone. A voice lives in sentence length, qualifier placement, transitions and punctuation, and none of those appear in an adjective.

Will the next ChatGPT or Claude release write better client comms on its own?

It will write a smoother average. Model capability and user-data conditioning are separate layers: each release improves how faithfully the model follows instructions, and no release contains your last forty client emails, because the lab has never seen them. The same improvement works for you once you have a profile, since a stronger model executes "recommendation in sentence one, one hedge per timeline" more precisely than the last one did.

Is it safe to paste client correspondence into ChatGPT or Claude to build a profile?

Build the profile from the least sensitive emails that still carry your voice: a status update from a closed engagement, a proposal cover note, an internal summary you wrote yourself. Redact client names and figures; the profile measures how you write, so the content is irrelevant to it. Once the profile exists, it holds rules about your style and nothing from the source emails, so it can travel to any tool your clients' policies allow.

---

If you would rather see what a client-comms profile looks like before deciding anything, the consultant profile page walks through one. The free Writing DNA Snapshot measures a few of your samples against the Average AI baseline and shows you, on one chart, where the model would have flattened you.