AI for mortgage broker compliance notes

There is a version of this idea that is sensible and a version that will get a firm into trouble, and they look similar from the outside. Both involve software producing text that ends up in a client file. The difference is whether the software is helping a person write down their reasoning, or manufacturing reasoning that nobody had.

The distinction that governs everything

A compliance note records why something was done. Its value comes entirely from the fact that it reflects an actual thought process in an actual adviser's head.

If an adviser considered three products, discounted two for identifiable reasons, and needs help expressing that clearly and consistently, drafting assistance is legitimate and useful. The reasoning existed; the tool improved the expression of it.

If a system is handed the case facts and asked to generate a rationale, something different has happened. The text will be fluent, it will be internally consistent, and it will describe reasoning that may not correspond to anything the adviser actually did. That is not a record. It is a plausible construction, and if it is ever tested against the adviser's recollection in a complaint, the gap will be obvious and damaging.

Everything below assumes the first version.

Where it genuinely helps

Consistency of structure. Compliance notes across a firm are usually wildly variable in quality, because they reflect how much time the adviser had that afternoon. A tool applying the same skeleton to every note raises the floor considerably, and the floor is what file checks find.

Completeness prompts. A well-constructed template can ask whether each required element has been addressed: circumstances, objectives, term rationale, product type rationale, fees disclosed, alternatives considered, protection discussed or declined. Advisers forget elements. A checklist that runs every time does not.

Plain English. Notes stuffed with abbreviations and internal shorthand are hard for a checker to follow and impossible for a client to follow if they ever see them. Rewriting into clear sentences is a task these tools do well.

Length discipline. Both directions. Some advisers write four words, some write two pages of narrative that buries the point. Bringing both towards a usable middle is a real gain.

Summarising a long file into an index. For a complaint response or a periodic review, an ordered chronology of what happened and when, drawn from documents you already hold, saves considerable time. Verify the dates.

Where it must not go

The suitability rationale itself, as discussed. The adviser supplies the reasons; the tool may tidy the words.

Any assessment of vulnerability, and any decision about what to do in response to one. Consumer Duty expects firms to identify characteristics of vulnerability and respond appropriately, and that response is a judgement about a person.

Any statement of a lender's criteria or a product's features. Check those against source. A note that misstates a product feature is worse than a thin note, because it is wrong in a way that looks authoritative.

Complaint responses. You can draft with help, but the assessment of whether the firm did something wrong is a decision under DISP, made by someone with the authority to make it.

Anything backdated. Producing notes for old files where notes are missing is not a remediation exercise, it is the creation of records that imply contemporaneity they do not have. If a file is thin, the honest fix is a note written today, dated today, saying what is known today and how it was reconstructed.

Supervision of the output

If drafted notes go into files, the supervision arrangements have to account for it. That means a few concrete things.

• Your file checking sample should specifically include AI-assisted notes, and the checker should know which ones they are.
• The check should test whether the note matches the case, not merely whether it is well formed. A generated note that reads perfectly and describes a different client's circumstances is the failure mode to hunt for.
• Someone should be looking across notes for repetition. If eleven files this quarter contain the same rationale sentence for eleven different clients, that is a signal that advisers are approving rather than writing, and a checker who reads files one at a time will never spot it.
• Training and competence records should reflect that advisers have been briefed on what they may and may not delegate.

What to write down before you start

Your compliance function will want a short policy, and so will your network if you are an appointed representative. It does not need to be long.

State which document types may be drafted with assistance and which may not. Name who reviews. Say how the review is evidenced, which in practice means a timestamp and a named adviser stored against the final version. Say which tool is approved and what data may go into it. Say what happens to the drafts that were rejected, because a discarded draft that contained client data still needs handling.

Then tell your network before you deploy it, not after. Appointed representatives in particular should assume their principal has a view, and finding out during a supervisory visit is the expensive way to discover it.

A test worth running

Take five recent files. Ask the adviser to write the compliance note the usual way. Separately, produce a drafted version from the same inputs. Put both in front of whoever does your file checking, without saying which is which.

You will learn two things. Whether the drafted version is actually better, which is not guaranteed. And whether your checker can tell the difference, which tells you something about how much scrutiny the drafted ones will really get once this becomes routine.

If the checker prefers the drafted note and cannot tell it apart, you have a useful tool and a supervision problem to solve at the same time. Solve the second one before scaling the first.