Compliance work is text work: policies, disposition notes, training decks, regulatory summaries, meeting minutes. LLMs are very good at text work. And almost everything a compliance officer touches is confidential. Those three facts collide daily, and pasting client data into a consumer chatbot is the wrong resolution: you’d be moving regulated data to an external service with its own retention behaviour, on the strength of nobody’s approval but your own.
The local-first split
The workable rule is a two-lane split:
- Cloud tools for public-input work. Drafting a training outline, summarising a published MAS consultation paper, explaining an instrument you’re reading, writing code. Input is public or generic; the output is yours to verify.
- Local models for anything with names in it. Client names, transaction details, screening results, internal findings. A quantised model running on your own hardware (LM Studio and similar runtimes make this a lunch-break setup, not a project) processes the text, and the text never leaves the machine.
Local models are weaker than frontier cloud models. That’s fine, because the sensitive tasks are usually the structurally simple ones: reformat this disposition note, extract the entities from this document, first-draft this periodic review from these bullet points. You’re not asking for brilliance; you’re asking for a fast junior who can keep a secret.
Guardrails that make it defensible
Using AI in a regulated firm is itself a process that should survive scrutiny:
- A written rule for what goes where. One page: cloud lanes, local lanes, examples of each. Undocumented judgement calls don’t scale and don’t audit.
- Human review as the control point. The model drafts; a person dispositions. Model output that becomes a compliance record must be reviewed by someone who is accountable for it, the same as work from any junior.
- No silent tooling. If AI-assisted drafting touches regulated workflows, senior management should know the shape of it. Shadow tooling in a compliance function is an irony nobody needs.
- Version awareness. Models drift and runtimes update. Note which model produced material drafts the same way you’d note which template version a policy used.
Common mistake
Framing the choice as “AI or no AI.” The real choice is which lane. Ban everything and staff will use personal accounts quietly: the worst outcome. Give people a fast, sanctioned local lane for sensitive text and the policy enforces itself, because the compliant path is also the convenient one.
General information, not legal or regulatory advice. Instruments are version-sensitive. Verify against mas.gov.sg and sso.agc.gov.sg before relying.