AI suitability report writing for UK advisers — what's safe, what's not

Last September a paraplanner at a mid-sized firm I spoke with — let's call her Rachel — was running a trial of a general-purpose AI tool to speed up suitability drafting. The first draft she got back referenced "Schroder Strategic Balanced Income — KIID dated 14 March 2024" as the cornerstone fund of the recommendation. The fund doesn't exist. Schroders has a Strategic Income range, but no Strategic Balanced Income product, and certainly not with that KIID date. The model had confabulated something plausible. Rachel spotted it before sign-off. Her senior didn't, on the next case, and a draft went to an adviser with two made-up FE Analytics ratings embedded in the rationale. The adviser caught it. The firm pulled the trial that afternoon.

None of this is a reason to write off AI in suitability drafting. It's a reason to be precise about what AI in suitability drafting actually means.

What the FCA expects

The starting point is straightforward. The FCA hasn't published bespoke "AI in suitability reports" rules — it doesn't need to. The existing framework already covers it. Under Consumer Duty, firms must produce good outcomes consistently and evidence the process. Under COBS 9 and 9A, suitability rests on a documented understanding of the client. PROD requires the firm to know which products match which target markets. None of that changes when AI is involved.

What does change is the operational risk profile. The FCA's 2024 AI Updateand Dear CEO letter to firms using AI made the firm's responsibilities explicit: the use of AI doesn't shift accountability, the human reviewer remains responsible for the output, and the firm must be able to evidence the controls around how AI is being used. Translated into practice, that means three things for a paraplanner-led suitability drafting workflow:

• Provenance.Every fact the draft asserts must be traceable to a primary source the firm holds — the fact-find, the platform record, the risk score, the KIID, the IM document.
• Human review.The named paraplanner reviews; the named adviser signs off. Neither role can be delegated to the model.
• Auditable controls.The firm can describe to the FCA, in writing, what the AI does and doesn't do, what the review process catches, and how errors are handled when they're found.

Where AI helps

Inside those constraints, AI does save time. Three places specifically.

Structural drafting.The suitability report is, structurally, the same shape every time — client objectives, current position, recommendation, rationale, risks, alternatives considered, costs and charges, ongoing service. AI is good at pouring real data into a known structure. A paraplanner who used to spend 90 minutes building the structure can spend 15 minutes reviewing it.

First-pass rationale prose.Once the recommendation's decided, writing the rationale paragraph from the fact-find inputs and the portfolio analysis is mechanical. AI handles the mechanical bit. The paraplanner edits for the firm's house tone, the case-specific commentary, the bits that need judgment.

Consistency checks.Where a firm produces 80 suitability reports a quarter, AI can flag inconsistencies — same client circumstance, different rationale phrasing — that a human review would miss across a large sample. Compliance officers find this most useful for the quarterly file-sample review.

Where AI hurts

The risks are specific. Each comes with a specific control.

Hallucinated facts. Models trained on general internet content invent fund names, KIID dates, FCA paragraph references, fund manager names, performance numbers. Rachel's Schroder example is the typical case. Control:never let the model write facts. Facts come from the firm's records — the fact-find, the platform's live Morningstar feed, the actual KIID PDF — fed into the draft as inputs the model can quote but not invent. If the model writes a fact that wasn't in its input, the draft is wrong.

Misapplied PROD rules. AI confidently asserts which target market a fund is appropriate for. It guesses, and the guess often sounds reasonable. Control:never let the model conclude on PROD. The firm's target-market mapping comes from the firm's own due diligence, not from the model's reading of a fund factsheet.

The regulated-advice line. A paragraph drafted by AI can read like advice — "this fund will provide the income you need in retirement" — without the qualifications a UK-regulated adviser would always add. Control:the firm's template constrains the model to recommendation language with the right qualifiers built in. The adviser reviews specifically for advice-language drift.

Vulnerability flag handling. Models are not reliable at calibrating language for clients flagged as vulnerable under the FCA's four-driver framework. They under-explain risks, or over-soften the rationale, or use jargon a vulnerability protocol would replace. Control:vulnerability-flagged cases get human-first drafting, not AI-first; the adviser writes, the model copy-edits.

The disclosure question.Should the firm disclose AI use to clients? The FCA hasn't mandated it as at May 2026; the Information Commissioner's Office has views on transparency where AI affects an individual's decisions. Best practice — and where WA points clients — is to disclose AI assistance in your editorial-policy-equivalent client communication policy, and on the firm's website. The named adviser remains responsible. Disclosure is honesty, not abdication.

How Wealth Analytica draws the line

This is the part where the article gets concrete about what WA does and doesn't do today, because the brief said be honest about current capability vs roadmap.

WA today produces structured drafts of suitability reports off the firm's live fact-find data, risk score and portfolio analysis. The drafts inherit the firm's house template and tone. They include only facts that are in the underlying client record — no model-invented holdings, no model-invented fund names, no model-invented FCA references. The paraplanner reviews the draft, edits the case-specific commentary, the rationale paragraph and the alternatives-considered section. The adviser signs off. Both signatures are recorded in the audit trail with timestamp and named user.

WA today does not auto-publish suitability reports. WA today does not generate fund recommendations — those come from the firm's CIP. WA today does not advise the client. The firm is the FCA-authorised entity. WA is the technology supporting them.

On the roadmap: better vulnerability-aware language calibration, configurable case-type templates, deeper integration with the firm's quarterly file-sample workflow. We'll publish those as they ship, in the format the FE wedge article in our insightsuses — factual, not aspirational.

What to look for in any AI suitability tool

If you're evaluating tools — WA or otherwise — the check-list is short.

• Provenance:can the tool show you, line by line, which input record produced which sentence?
• No model-invented facts:does the tool ever assert a fund name, holding, performance figure, FCA reference, or KIID date that wasn't in the input?
• Adviser-in-the-loop by design:can the workflow even produce a published report without a named human signing off?
• Vulnerability-flag awareness:does the tool change its behaviour when a client record carries vulnerability indicators?
• Audit trail:when something goes wrong on a 2027 sample-file review, can the firm show what the tool did, what the human reviewed, and what was changed?

Any tool that can't answer those five is not ready for UK-regulated suitability work, however slick the demo.