Let AI write the suitability report's wording, and keep every fact tied to your own records

Two minutes. That's roughly how long an AI tool needs to draft a suitability report, according to our own guide to AI suitability writing.

The same guide opens with the other side of it. A paraplanner trialling a general-purpose tool got back a first draft built around a fund that doesn't exist, complete with a plausible KIID date. She caught it. On a later case, a draft reached an adviser with two made-up fund ratings in the rationale. The adviser caught that one, and the firm stopped the trial the same afternoon.

Both outcomes came from the same machine doing the same thing. It's very good at shape. It's poor at facts, and it can't tell the difference.

That's the whole argument for how I'd use these tools in suitability work: let the model write the wording, and never let it supply a fact.

The shape is where the time goes

A suitability report follows much the same structure every time. Client objectives, current position, the recommendation, the rationale, the risks, the alternatives considered, costs and charges, ongoing service. Building that from a blank page is slow, repetitive work, and it's the sort of work these models do well.

Our guide puts it at around 90 minutes for a paraplanner to build the structure by hand, against about 15 to review one drafted from the firm's own data. I'd treat those figures as indicative, because every firm's templates differ. The direction is clear enough.

The rationale paragraph is similar. Once the adviser has decided the recommendation, turning the fact-find and the portfolio analysis into readable prose is mostly mechanical. The paraplanner still edits for tone and for the parts that need judgement. But she's editing, which is quicker than writing.

Facts are a different job

Almost everything that goes wrong in AI drafting goes wrong at the point where the model is allowed to fill a gap, whether that's a fund name, a KIID date or a paragraph number from the FCA Handbook, because a general model has read enough of all of these to produce something that looks right, and a reviewer skimming at the end of a long day tends to see what they expect to see.

Looking right is the problem.

So the rule is simple. Facts come from the firm's records: the fact-find, the platform valuation, the risk score, the actual KIID the firm holds. The model can quote them. It can't invent them. Any fact in a draft that wasn't in the input is an error, however plausible it reads.

Target market conclusions under PROD sit on the same side of the line. They come from the firm's own due diligence, and a model's reading of a factsheet is no substitute.

And there's one kind of case I'd keep human-first entirely: clients flagged as vulnerable. Our guide notes that models are unreliable at pitching language for them, sometimes under-explaining risk and sometimes softening the rationale too far. There, the adviser writes and the model tidies.

What the regulator already expects

There's no separate FCA rulebook for AI in suitability reports, and there doesn't need to be. Consumer Duty, the suitability rules in COBS 9 and 9A and the PROD rules apply as they always have. The FCA's 2024 AI Update made the accountability point plainly: using AI doesn't move responsibility for the output. The named paraplanner reviews. The named adviser signs.

In practice, that means a firm should be able to describe in writing what the tool does, what the review catches and what happens when an error is found.

One test before you trust a draft

When you look at any drafting tool, ours included, pick a sentence in the draft at random and ask the tool to show you which record it came from.

If it can, the reviewer is checking a claim against a source. Quick. If it can't, the reviewer is checking the claim against the world, which is slow, and most of the two minutes you saved has gone back into the file.

That single check says more about a tool than any demo. The good tools and the risky ones look much the same until you ask.

How much of your review time goes on checking facts a draft should never have been allowed to write?