SRA AML compliance AI for UK law firms: what it can do, what it cannot.

What the SRA actually fails firms on

The SRA AML Annual Report 2024/25 inspected 833 firms and found 32.4% non-compliant. The most common specific failures were: (1) missing or inadequate client and matter risk assessments (CMRAs) under MLR 2017 Reg 28(12)–(13) — present in 50% of non-compliant firms; (2) missing firm-wide risk assessments (FWRAs) under MLR 2017 Reg 18; (3) weak ongoing monitoring under Reg 28(11); and (4) inadequate source-of-funds documentation. Critically, 90% of inspected firms already used electronic ID tools — the gap was in the reasoning and documentation layer above the data.

These failures are not caused by lack of effort or expertise. They are caused by the volume of documentation required per matter, the difficulty of consistently applying the regulatory framework across dozens of fee earners, and the absence of an automated system to flag when a matter falls short. These are tractable problems for AI-assisted tooling — if the tooling is designed correctly.

What AI can and cannot do in this context

AI can draft CMRAs citing the specific MLR 2017 regulation and LSAG paragraph; identify gaps in existing documentation; monitor active matters for the triggers the SRA specifically tests for (transaction value changes, counterparty changes, re-screen expiries); and assemble inspection-ready evidence packs on demand. These are appropriate uses where AI functions as a highly efficient drafting assistant, not as a decision-maker.

AI cannot and should not: make the final compliance decision; auto-file SARs or trigger any legal act; assert privilege on behalf of the firm; or substitute for a qualified MLRO's professional judgment. The SRA is explicit in its AI Risk Outlook guidance that firms must not over-rely on AI and must retain human accountability for every compliance record. Any product that does not enforce this — via a mandatory human sign-off step before any output is finalised — is not fit for SRA-regulated use.