Your AI doesn't sound like you — here's the fix

A lot of AI writing has the same problem.

It is polished enough to be useful, but bland enough to be suspicious. It reads like something produced by a capable person who has never met you, never seen your work, and has a slightly unhealthy relationship with LinkedIn.

That is not really a writing problem. It is an evidence problem.

Most people ask AI to write in their style, but they do not give it any meaningful evidence of what that style is. They give it a vague instruction — “make this sound professional”, “make it more senior”, “write it like me” — and then act surprised when the output sounds like a generic business update wearing a lanyard.

A new colleague would have the same problem. If they had never seen how you write, they would not know whether you are direct or cautious, whether you prefer structured paragraphs or bullets, how much context you usually include, how you make recommendations, or how you disagree without making a meal of it.

AI is no different.

The useful shift is to stop asking it to guess and start giving it evidence.

Take a small set of writing you are happy with: a project update, a board note, a difficult email, a LinkedIn post, a briefing paper. Strip out anything sensitive. Then ask the AI to analyse the pattern, not copy the content.

You are looking for things like:

That becomes a reusable voice profile.

The prompt changes from this:

> Write this in my style.

To this:

> Use the voice profile below. Draft this as a clear leadership note. Keep the judgement and structure. Avoid corporate filler. If anything important is missing, ask before finalising.

That is a very different instruction. It gives the AI evidence, constraints and a standard to work against.

It still needs review. Obviously. We are not outsourcing judgement to a stochastic intern with excellent grammar.

But the first draft gets closer, and the editing task changes. You are no longer fighting generic tone from the first line. You are refining something that at least understands the shape of the answer.

This matters because competent writing is becoming cheap. Tidy paragraphs are no longer a differentiator. The value is in the judgement behind them: what you include, what you leave out, where you challenge the brief, and when you say “this is not the right problem”.

If AI output sounds bland, do not just blame the model.

Give it better evidence.

If you missed the previous issue on why synthesis is where AI starts to get useful, it is here: https://theaidirective.co.uk/p/the-one-thing-ai-does-better-than-any-human

And the earlier piece on why most AI projects fail is here: https://theaidirective.co.uk/p/why-most-ai-projects-fail-and-what-to-do-instead

---

The simple version is this.

Choose three to five examples of your own writing that you would be happy to send again. They do not need to be public pieces. In fact, internal writing is often better because it shows how you actually think when there is a decision to make.

Use approved tools. Remove sensitive detail. Do not paste confidential material into a public model and then call it productivity. That is not innovation; it is an incident report with better branding.

Then use a prompt like this:

> Analyse the writing samples below. Extract a reusable voice profile I can give to an AI assistant. Cover tone, structure, sentence style, level of directness, how I make recommendations, how I handle uncertainty, and what the assistant should avoid. Do not imitate the private content. Extract the style pattern only.

Ask for two outputs:

1. a plain-English profile you can review; and

2. a compact instruction block you can paste into future prompts.

Then test it on something low risk: a meeting follow-up, a project update, or a first draft of a LinkedIn post.

This is not magic. It is just context discipline.

Unfortunately, context discipline is where a lot of AI adoption either starts working or quietly turns into theatre.

---

I have made a YouTube video on the same problem:

Don't Lose Your Voice to AI

The point is not to pretend AI was not involved. That is the wrong objective.

The point is to make sure AI-assisted work still carries the judgement that made the writing worth reading in the first place.

---

1. If your organisation has Copilot, do not start with a huge prompt library. Start with five useful patterns: project updates, decision briefs, risk summaries, meeting follow-ups and document comparison.

2. Personalisation is not automatically a privacy problem. The risk is putting sensitive material into the wrong tool. The opportunity is extracting reusable patterns from safe examples and using them inside approved environments.

3. The most useful AI work is often boring: naming files properly, agreeing what good looks like, writing reusable instructions, and checking source material. Annoying, yes. Also the difference between a toy and a workflow.

---

“Should everyone in a company have their own AI voice profile?”

— Operations leader

Eventually, maybe. But I would not start there.

Start with people who write high-leverage material: leadership updates, customer communications, board papers, policy notes, sales proposals and operational briefings.

For many teams, a shared team profile is more useful than 50 individual ones. How does this department write? What must always be included? What language should be avoided? What does a good decision note look like?

Individual profiles can come later.

Otherwise you get personalised chaos. Very authentic. Not necessarily useful.

---

> AI will average you into the internet unless you give it evidence, constraints and a reason not to.

---

If your organisation already has Microsoft 365, you probably have enough safe material to improve this quickly: approved templates, public copy, anonymised examples, old project updates and agreed briefing formats.

The trick is not asking AI to be creative.

It is teaching it what good looks like in your context.

Next week: turning that profile into a reusable AI agent.

See you Tuesday.

— Toby

@TobySmithFirth on YouTube

---

These are tools I use personally. Affiliate links marked — I earn a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend tools I actually use.