Executive AI Writing Delegation Without Losing Authority

TL;DR: Executive AI writing delegation goes wrong at the brief, not at the model. "Professional but warm, confident, concise" describes thousands of leaders, so the draft that comes back belongs to none of them. Authority in writing is carried by measurable habits: sentence length, where you place qualifiers, how you move between points, what you do with punctuation. Hand those to ChatGPT, Claude or Gemini as instructions and the model can reproduce them. Hand it adjectives and it reproduces the average. Better models make the first approach work better and do nothing for the second.

You can delegate the draft. You cannot delegate the voice, and when the two get confused, people in your organisation notice before you do.

The update you sent the leadership team last week was fine. Accurate, polite, on time. Two people replied to your chief of staff instead of you. That is what lost authority looks like in practice: no complaint, just a quiet re-routing, because the email did not read as yours.

Executive AI Writing Delegation Is a Conditioning Problem

The draft comes back generic because the model was never conditioned on you, and no amount of capability substitutes for that. A language model carries the statistical shape of the text it was trained on, plus whatever is in the current conversation. Your writing was in neither, so a request to draft "as you" produces a blend of every leader who resembles the description you typed.

This is the position we build the product on, and it explains most delegation failures we see. Executives assume the problem is model quality, so they upgrade to the newest tier, try Claude instead of ChatGPT, or add a longer paragraph about their style. The draft improves in polish and stays exactly as anonymous. The layer that was missing is still missing.

Think of it as two separate inputs. Capability is what the lab ships: reasoning, instruction-following, vocabulary, restraint. Conditioning is what only you can supply: the evidence of how you actually write. Every release moves the first and leaves the second untouched. If you want the draft to sound like you, the conditioning has to travel with the brief. The Humanity Score test shows the size of the gap between your current setup and your own writing in about a minute, and the rest of this post is about closing it.

What Actually Signals Authority in a Leader's Writing?

Authority is recognised through habits a reader could count, not through a tone they could name. Ask a direct report why a message "sounds like the CEO" and they will say confident, direct, clear. Put two confident, direct, clear emails from two different executives side by side and the same person will tell you which is which in seconds. The adjectives are shared. The habits are not.

Here is what those habits look like once you list them:

Every row is measurable from ten emails you already sent. None of them is carried by "professional yet friendly". That is the core of why tone adjectives are not your voice: AI does not know your tone. It knows words that describe tone, and those words map to a register shared with everyone else who used them.

Authority is also register-specific. The way you write to the board is not the way you write to a first-year analyst, and readers hold you to both. A profile that captures one register and gets applied to the other is the second most common way delegation goes wrong, after the adjective brief.

The Delegation Brief Most Executives Give Cannot Carry a Voice

The standard brief is a paragraph of description plus three pasted emails, and neither half carries enough information to reproduce you. The description is adjectives. The samples are too few for the model to separate habit from accident.

Take the description first. "Direct, warm, decisive, no fluff, confident without being arrogant." Read it as the model does: each word is a pointer to a learned register. "Direct" pulls one style, "warm" pulls another, and the model averages them into something plausible. Plausible is the problem. It reads like a leader. It does not read like you.

Now the samples. Three emails show the model what you sounded like on three days. It has to guess which patterns are yours and which belonged to that Tuesday. Was the two-sentence opening a habit or a rush? Is "which means" your transition or a one-off? With three samples the guess is a coin flip, and the guess drifts as the conversation gets longer. The sample-count question has a real answer, and three is well under it.

The Custom Instructions field in ChatGPT, and the equivalent project instructions in Claude and Gemini, are the right place for a voice. They persist across chats, they apply before the model starts drafting, and they are read every time. What goes in the field is what matters, and a paragraph of adjectives wastes the slot.

Why a Newer Model Will Not Fix It

A stronger model follows instructions more faithfully, so it executes a vague brief more faithfully too, and the result is a more polished version of the same average. Upgrading treats the wrong layer.

This is worth stating plainly because the upgrade is the first thing most executives try, and it looks like it works for a day. The new model writes cleaner sentences. It hedges less. It stops opening with "I hope this email finds you well." Those are capability gains, and they are real. What has not changed is that the model still has no measurement of you, so it is still drawing from the professional default, just a better-executed one.

The reverse is the interesting part. If the brief contains measurements instead of adjectives, a more capable model reproduces them more reliably than the last one. "Sentences of 10 to 16 words" was a rough target for an older model and is a tight constraint for a current one. "Never open with a pleasantry" used to be ignored half the time. Now it holds. Capability and conditioning are separate layers, and better capability makes a real profile more valuable, not less. We wrote about that mechanism in why better models still do not know your voice, and the delegation case is the sharpest version of it, because the executive is paying for the gap in credibility rather than in editing time.

A Voice Lives in Four Measurable Places

A leader's voice can be written down as four sets of numbers and rules, and once it is, any capable model can hit them. The four places are sentence length, qualifier placement, transitions and punctuation.

Sentence length. Not the average, the shape. Most executives have a typical range and a deliberate short sentence they use for emphasis. Count words per sentence across ten emails and both show up immediately.

Qualifier placement. Everyone hedges. Authority comes from where the hedge sits. "I think we should ship Friday" and "We should ship Friday, I think" carry the same information and land differently. Most leaders hedge before a recommendation and never after a decision, and never notice they do it.

Transitions. The words that move a reader from one point to the next. A voice usually has three or four it uses and a long list it never touches. The never-list is more distinctive than the use-list.

Punctuation. Colons or full stops. One comma per sentence or three. Whether a question mark ever appears in an email to a subordinate. These are habits a reader feels without naming.

Write those four down as instructions and you have something the model can obey and you can check. That is the whole difference between a Style Profile and a tone paragraph. The first is evidence of how you write. The second is a hope about how you sound.

How Do You Delegate a Draft Without Handing Over Your Voice?

Separate the two things you are delegating: the substance, which you supply as bullets, and the form, which you supply as a measured profile. The model's job is to combine them. Your job is the check at the end.

In practice the protocol has four steps, and the order matters:

1. Brief the substance in bullets. Who it is to, what they already know, what you need from them, by when. Thirty seconds. Do not write sentences; the model will keep your sentences and you lose the time saving.
2. Load the profile, not a description. Your instruction block from the four places above sits in the persistent instructions field so it applies to every draft. If you are drafting for a different register (board versus team), switch to that register's block.
3. Draft. Ask for the email, not for "a few options". Options invite the model to vary style, and variation is what you are trying to remove.
4. Check by counting, then send. Words per sentence in range? Ask in the first line? Hedges where you put them? Sixty seconds, and it gets faster because the draft converges on the numbers.

Notice what is missing: the paragraph where you describe yourself. It never helped. A board memo written with AI follows the same protocol with a longer brief and a stricter register block, and the check at the end is the same count.

What to Keep Writing Yourself

Delegate the drafts where the form is routine and the substance is settled, and keep the ones where the act of writing is itself the message. That line is where executive authority actually lives, and it has nothing to do with the model.

Keep for yourself:

• The note after a layoff or a death on the team
• A personal apology
• The first message to a board member or investor you have not met
• Any decision you have not yet made, because drafting it will make it for you

Delegate freely:

• Status updates and recaps
• Meeting follow-ups with owners and dates
• First drafts of memos where you supply the bullets
• Replies where your position is already known

The mistake is not delegating too much. It is delegating the wrong category and then compensating by delegating nothing. A profile that reproduces your form lets you hand off the routine tier with confidence, which is what frees the time for the tier only you can write.

The Chief of Staff Test

A draft is ready to send as yours when the person who knows your writing best cannot tell it was delegated, and the honest test is to ask them. Most executives already have this reviewer: a chief of staff, an executive assistant, a long-tenured deputy who has read a thousand of your messages.

Give them three drafts, one you wrote and two the model produced under your profile, and ask them to pick yours. If they pick correctly with confidence, the profile is missing something, and they can usually tell you what. "You would never say 'reach out'." "You always name the date." Each of those is a rule you add to the block.

If they hesitate, or pick wrong, you have a profile that carries your authority. Run the test again after a model upgrade. In our experience the drafts get harder to spot, not easier, because the model hits the numbers more precisely, which is the practical face of the capability-versus-conditioning argument.

A 20-Minute Setup You Can Do Today

You can build a first version of the instruction block by hand from ten emails, without any tool, and it will beat any adjective paragraph you have tried. Here is the procedure.

1. Collect ten emails you wrote and were satisfied with, all in one register (pick the one you delegate most, usually internal updates).
2. Count words per sentence across all ten. Write down the typical range and the shortest sentence you used on purpose.
3. Find every hedge ("I think", "probably", "it may be worth") and note where it sits relative to the recommendation.
4. List your transitions, then list the ones you never used. Both lists go in the block.
5. Note the opening and the closing you use most. First-line ask or not. Named owner and date or not.
6. Write it as rules, not adjectives: "Typical sentence 11 to 15 words. Decision in the first sentence. One hedge per email, before the recommendation. Transitions: so, which means, a colon. Never however, never reach out, never hope this finds you well. Close with one owner and one date."
7. Paste it into the custom instructions field of the assistant you already pay for.

That is a rough profile and it works, because it gives the model numbers instead of words about numbers. A Writing Twin built from a larger sample set does the same measurement with more precision and across more registers, and pastes into ChatGPT, Claude or Gemini the same way. Either route puts the conditioning layer back where the model cannot supply it. If you want to see how far your current setup sits from your real writing first, the Humanity Score test takes about a minute.

FAQ

Can an executive delegate email drafting to AI without sounding generic?

Yes, if the model is given measurements of how you write rather than a description of how you would like to sound. A brief such as "confident, warm, concise" points the model at a register shared by thousands of leaders. Instructions such as "open with the decision, one hedge per email, sentences of 10 to 16 words, no however" point it at you. The draft still needs your read, but it starts close instead of starting average.

Why does "write this in my voice" fail with ChatGPT or Claude?

Because nothing about your voice is in the model. ChatGPT, Claude and Gemini know the statistical shape of the text they were trained on plus whatever sits in the current conversation. "My voice" resolves to nothing unless the conversation contains it. The model then falls back on a professional default, which is why every executive who types that phrase gets a draft that reads like the same competent assistant.

Will a newer, more capable model solve executive AI writing delegation on its own?

No. Model capability and conditioning on your writing are separate layers. Each release improves the first, and none of them can supply the second, because your emails were never in the training data and no update adds them. A stronger model does follow precise instructions more faithfully, so a measured profile of your writing gets more valuable with each release while a tone label stays exactly as vague.

What should an executive still write personally instead of delegating to AI?

Anything where the act of writing is the signal: a note after a layoff, a personal apology, the first message to a board member you have not met, and any decision you have not fully made yet. Delegate the drafts where the substance is settled and the form is routine: status updates, meeting follow-ups, recaps, first drafts of memos where you supply the bullets. Keep the judgement calls, hand off the typing.

How do I check whether an AI draft actually sounds like me?

Take an email you already sent, give the assistant only the brief you would have given a colleague, and compare the two drafts by counting rather than by feel. Words per sentence, where the ask sits, how many hedges and where they fall, which transitions appear, how many commas per sentence. If the numbers differ, the draft is not in your voice, however polished it reads. Repeat across three registers before you trust it.

You can measure the gap between your AI's drafts and your own writing in about a minute with the Humanity Score test, or build a Writing Twin from your own emails and paste it into ChatGPT, Claude or Gemini today.