aiDex for Executive Assistants: Draft in Their Voice, Not the Model's

Most jobs want AI to sound competent. Yours does not. An executive assistant writes as someone else, in that person's register, to people who know exactly how that person writes. The default output of a strong model is fluent, polite, slightly generic business English, and that is precisely the wrong answer: it reads like a stranger borrowed the account.

The second half of the job is worse for a single chatbot. Deciding what reaches your executive is a judgment call with a lopsided cost. Forwarding one thing too many wastes thirty seconds. Missing the one request that mattered costs a great deal more. A single model gives you one confident opinion on that call and no way to tell how confident it should have been.

A panel fixes both, and it fixes them the same way: by turning one opinion into three, so that agreement means safe and disagreement means look closer. That is the core of a multi-model workflow, applied to a job where the stakes belong to someone else.

Why does one AI model get your executive's voice wrong?

Because it has never seen your executive write. Voice is not a talent problem for a language model, it is a constraint problem. Left unconstrained, every model falls back to its house register, and the house register of every major model is roughly the same polished corporate neutral. Ask three of them to decline a meeting and you will get three variations of "Unfortunately, I won't be able to attend, but I'd love to reconnect soon."

The fix is examples plus explicit rules. Anthropic's own prompting guidance recommends including three to five relevant, varied examples and wrapping them in tags so the model can tell samples from instructions (multishot prompting). That is a voice card, and it is the single highest-leverage asset an assistant can build.

How do I build a voice card once and reuse it forever?

Open aiDex, start in Solo, and paste five to eight messages your executive actually sent: one warm, one curt, one declining something, one to a board member, one internal. Then ask for the rules as a checklist, not as prose. You want output you can argue with, like this:

• Greeting: first name only, never "Hi team"
• Sign-off: initials, never "Best regards"
• Sentence length: short, often fragments
• Never apologizes for timing, never writes "just checking in"
• Says "let's do it" instead of "let's proceed"
• Declines in two sentences and does not explain

Then edit it yourself. You know things the samples do not show, such as who gets the warm register and who gets three lines. That edited card is the rubric for everything below, and it takes one sitting to build.

Can AI draft emails in my executive's voice?

Yes, to roughly the point where it stops being useful on its own, which is why the panel matters more than the model. Run the draft through Pipeline: Draft, Critique, Revise, Polish, which is the four-stage document pipeline in its shortest form. The trick is what you hand the Critique stage. Do not ask "is this good writing?" Ask it to score the draft against the voice card, line by line, and quote every line that breaks a rule.

Then send the polished version to Judge with two questions: does this match the card, and would sending this cost my executive anything with this specific recipient? Three models, two questions, pre-written criteria. A draft all three approve, you send. A draft they split on is the one you read word by word before it leaves.

How do I triage an inbox without missing the urgent thing?

Write the escalation rules down first, then let Compare do the sorting. Paste the same batch of subject lines and first paragraphs into a comparison and give every model the identical rule set: what counts as Escalate, what you Handle yourself, what gets Archived. Each model sorts the batch independently.

The rule that makes this safe is simple, and it comes from the cost asymmetry, not from the models:

Disagreement is not a failure of the panel. It is the panel telling you the item does not fit a clean rule, which is exactly the item a human should look at. A single model would have hidden that ambiguity behind one confident label.

How do I build a meeting or travel briefing packet?

Use Team and let the documents do the work, the same way a multi-model meeting notes pass turns a transcript into decisions. Drop the deck, the contract and the last set of notes into the chat so every model in the room reads the same source material, then give each seat a job: one summarizes what is in the documents, one plays the skeptic and lists the questions your executive will be asked that the packet does not answer, one writes the final one-pager. A lightweight moderator model keeps the order straight so you are not refereeing.

The skeptic seat is the one worth protecting. Summaries are easy and every model produces a decent one. The gap list is the thing that saves a meeting, and it only appears when a model is explicitly told to hunt for gaps instead of being helpful.

What should never go into a cloud model?

Compensation, medical and legal matters, unreleased board material, and personal travel that carries home addresses or family names. Those stay local. Run the Dex against an Ollama model on your own machine for that category of work, and accept the honest tradeoff: local models are noticeably weaker at polish, so use them for structure, triage and first passes, and write the final sentences yourself.

Everything else is a choice about cost and control rather than safety. Use your own provider keys or the ones we manage, and pick the models you want. If your organization already holds enterprise agreements with a provider, your own keys keep the traffic under that agreement.

What should the panel never decide?

It never sends. It never decides permanently who gets access to your executive, because access is political and the models do not know the politics. And it never sets the tone of a genuinely hard message, such as a decline that will sting, an apology, or anything touching someone's job, without you reading every word first.

That last boundary is not caution for its own sake. The reason an assistant is trusted with the account is judgment about people, and judgment is the part of the job that does not delegate. The panel buys back the drafting time so you have more of it to spend on the calls that matter.

How long does this take to set up?

One sitting. The voice card and the escalation rules are the only real work, and both are reusable assets rather than prompts you retype. After that, a draft is a paste, a triage pass is a paste, and a briefing packet is three documents and one instruction. Set up Teams if more than one person supports the same executive, so you are both drafting against the same card instead of two private versions of it.