Why AI writes like a brochure for advice firms, and what to give it instead

Every adviser I've spoken to about this has tried AI once, got something that read like a brochure, and quietly stopped.

That happens in every industry. In a regulated firm it creates work for somebody else. A paragraph that overstates a benefit has to be caught. So does a missing balancing line, or a figure the model made up. The adviser who tried AI to save an hour has instead made another thing for compliance to read. Do that twice and the firm decides AI isn't for them, which is the wrong conclusion drawn from the right evidence.

The fix isn't a cleverer prompt. It's what you give the model before you ask it anything.

Setting it up means giving it 5 things

A blank chat knows nothing about your firm. Not your tone, not your clients, and not the sentence your compliance team sends back every single time. Asked to write from nothing, it produces the statistical average of everything ever written about pensions, which is exactly the brochure you got.

Here's what goes in, in the order that matters.

1. Your approved wordings. The phrases your firm has already agreed it uses, and the balancing language that goes with them. This is the highest-value thing you can give it, and it's the easiest one to skip. If your firm has a standard way of describing capital risk, the model should have it word for word rather than reaching for its own phrasing.

2. Who you're actually writing to. Not "clients". A 58-year-old who has just been made redundant with a defined benefit transfer decision in front of them is a different reader from a business owner planning an exit in 9 years. Both are your clients. They need different sentences.

3. What you'll never say. The phrases that come back marked up every time. Guarantees. Anything that reads as a recommendation when it should read as information. Superlatives about performance. Write the list down once and it stops happening.

4. Your actual writing. 10 pieces you were happy with, that got through sign-off without a fight. This is the richest source of the lot, and it teaches the model your rhythm far better than any description of your tone.

5. Somewhere for all of it to live. A Claude Project rather than a chat you start again each morning. If the context disappears when you close the tab, you'll be re-explaining your firm every day, and you'll stop within a fortnight.

None of that is a prompting trick. It's an afternoon spent putting your approved wordings somewhere the model can read them.

What that looks like in practice

Take a fictional firm, Aldgrove Financial Planning. Two advisers, a paraplanner, and a compliance function that is one person who is also the managing director.

Before: a client-facing note on market volatility gets drafted, comes back marked up, and goes round again. Most of the comments are the ones that came back last time too.

After the setup: the first draft already carries the firm's approved wordings and the balancing line, and it sounds like the adviser who'll sign it. It still goes to sign-off. What comes back is new comments rather than the same ones again.

That's the honest version of the benefit. It speeds your compliance sign-off and helps drafts get through with fewer rounds. It does not do the sign-off.

Where it stops

This is the part I'd rather say plainly than bury.

A well set up workspace is an assistant to a human decision, never a replacement for one. It doesn't approve anything. It has no view on whether a piece of copy meets your obligations, and any tool that tells you it does is selling you something it can't deliver. Nothing here removes a review step. It makes the thing arriving at that review step better.

Three things I'd keep it away from entirely:

• Anything that constitutes advice to a specific client. Suitability is a human judgement about a real person and their circumstances. That's the job, not the admin around the job.
• Numbers you haven't checked. Models produce plausible figures. Plausible is the dangerous kind of wrong. Every figure gets verified against source, every time.
• Final copy that skips review. If the setup is working, review gets faster. If you're tempted to skip it because the drafts are good now, that's the moment the setup starts costing you.

Where to start

Don't try to build all 5 at once. Start with the approved wordings and the never-say list. Those two carry most of the weight, and you probably already have them written down somewhere, scattered across old emails and a document nobody has opened since the last review.

Get them into one place. Point Claude at them. Write the next client note and see what comes back.

If it still reads like a brochure, the context isn't specific enough yet, and that's a fixable problem rather than a reason to give up.

If you'd rather think it through with something in front of you, the AI Project Starter asks a few questions about what you're trying to do and gives you a one-page brief at the end. It's free, and it doesn't need your wordings to be tidy first.