Compliance AI has to be built narrow to be trusted

I have spent more than a decade building AI systems for large organisations, and the pattern does not change. A general model looks brilliant in a demo. Give it a complex, regulated task with real files and it fails the test, because the task needs to be encoded in a very vertical-specific way: the right datasets, the right tools, and a scope narrow enough that you can tell when it is wrong.

That is why we started with basic file checks on buy-to-let and residential cases, not with a chatbot. A narrow task has a rule list a compliance team can read, an answer a reviewer can test against the file, and a finding that points to the page it came from. A general assistant has none of those. It gives you a fluent paragraph and no way to prove where it came from.

The bigger risk sits in the tools a firm's people already use, not the ones it buys.

I don't think the industry quite realises how big the shadow AI problem is. Individuals want to get work done. Some do not know that uploading a document with personal data to their private version of ChatGPT is not acceptable. Then there is the residency question: do you know whether that data is processed only in the UK and EU, or whether it goes to the US?

The numbers say adoption is already well ahead of understanding. The Bank of England and FCA survey published in November 2024 found 75% of firms already using AI, and 46% reporting only a partial understanding of the AI they use. A third of all use cases are third-party, and in risk and compliance the figure is 64%. Of the top five risks firms named, four are about data, starting with data privacy and protection.

The rules that apply are not new ones. In its 2024 AI Update the FCA describes itself as "a technology-agnostic, principles-based and outcomes-focused regulator". It points firms to SYSC 4.1.1R, which asks for "effective control and safeguard arrangements for information processing systems", and for Enhanced SM&CR firms it says any use of AI in a business area "would fall within the scope of a SMF manager's responsibilities". UK GDPR adds that personal data has to be processed with "appropriate security". A client's bank statement in a personal chatbot account sits badly with all of that.

The fair challenge goes like this. General models improve every few months. Buy an enterprise licence with the right data terms, give it to every adviser and paraplanner, and the shadow AI problem goes away. Why build anything narrow?

An enterprise licence fixes where the data goes. It does not fix what the tool does with the file. A general assistant still has no fixed rule list, so two people asking the same question about the same case can get two different answers. It still produces findings you cannot trace to a rule. And it still leaves the firm judging AI by how convincing it sounds, which is the wrong test for anything a regulator may later ask about.

Better models make narrow systems better too. The model is one part of a compliance system. The rule set, the connections to the case data, the evidence trail and the logging are the parts that make the output something a firm can stand behind. That is the work a general tool leaves out.

Two mistakes come up often.

• Banning without replacing. A policy that says "do not use public AI" while the workload stays the same pushes usage out of sight. People who want to get work done will find a way.
• Judging by the demo. An impressive demo on a clean sample file says little about what happens on case number 4,000, with a scanned payslip and a missing page.

When I am asked how to assess an AI compliance system, I use five tests.

1. Independence. You should not mark your own homework. That is why independent compliance exists.
2. An evidence chain you can reproduce without the vendor in the room.
3. A rule set your firm genuinely understands.
4. Systematic logging that can withstand FCA scrutiny.
5. Training embedded across the full deployment, not left to a launch session.

A firm that cannot get clear answers on all five should keep looking. At Curvestone this is what we build for: narrow checks, each finding with its reasoning and source evidence, and every decision and override logged.

Narrow design also changed the economics. Once each check is defined and testable, it can run on every case rather than a sample. If something is possible, it eventually becomes the norm.

1. Find out what your people already use. Ask, without blame, which AI tools are in use for case work, and on which accounts.
2. Map each use to a data question: where the data is processed and who can see it.
3. Give people a sanctioned route for the tasks they were using public tools for, so the policy is not just a ban.
4. Put any compliance AI through the five tests above before it touches a live file.
5. Name the senior manager accountable for AI in each business area, and record it.

This is evolution, not revolution. The industry, its clients and the FCA need time to adjust, and the firms that build narrow and checkable now will be ready when the volume arrives.

An earlier version of this argument appeared in Mortgage Introducer (October 2026).