Using AI on client work without putting client data at risk

In most finance and consulting firms, AI adoption follows the same pattern. A few people start using it quietly on their own accounts. Someone in risk notices. A blanket ban goes out. The quiet use carries on, now with less oversight than before.

The ban doesn't remove the risk. It moves it somewhere you can't see it.

A better approach is to decide what can go where, write it down in one page, and give people a sanctioned way to do the work. Here is the framework we use with teams.

Sort the work into three buckets

Green: no client or personal information. Drafting a generic proposal section, explaining a regulation, writing a formula, summarising a public annual report, improving the wording of an internal email. Any approved AI tool is fine.

Amber: client information, but nothing personal or price-sensitive. A client's process, their industry context, anonymised figures. Use only your firm's business or enterprise AI plan, and strip names and identifiers where the task doesn't need them.

Red: personal information, or material non-public information. ID numbers, account details, health information, unannounced deals or results. This stays out of general AI tools unless your firm has a specific, approved setup for it, with a named owner.

Most day-to-day work turns out to be green or amber. That is the point: once people know the rules, the safe majority of work can move quickly.

Choose the tool before the use case

Consumer AI apps and business plans are not the same product. Before rolling anything out, someone should be able to answer four questions about the tool your team uses:

1. Is our data used to train the provider's models? Many business plans say no by default. Check the actual terms, not the marketing page.
2. Where is it stored, and for how long?
3. Who in our firm can see what others have shared?
4. Can we switch off features we don't want, such as connecting to email or file storage?

If you can't answer these, you are not ready for amber work.

Make the safe way the easy way

Rules on paper get ignored when they slow people down. Three things make them stick:

Build shared skills and knowledge bases. If the approved tool already knows your report templates, your house style and your methodology, nobody needs to paste a client file into a random app to get a good result.

Teach masking as a habit. "Client A, a mid-sized retailer" is almost always enough context for the AI to do the job. Replacing names takes seconds once it becomes routine.

Keep a person accountable for every output. AI drafts, a person checks and signs. That is how you already treat work from a junior analyst. Treat AI the same way.

POPIA still applies

Under the Protection of Personal Information Act, your firm remains responsible for personal information it processes, including when a tool or service provider does part of the processing. Using AI doesn't create a new exemption, and it doesn't create a new ban either. It means the usual principles (process only what you need, keep it secure, know who your operators are) apply to one more set of tools.

If you are unsure where a specific use case falls, that is a question for your compliance team or legal adviser. The framework above is designed to make those conversations shorter, not to replace them.

Where to start

Write your one-page green, amber, red guide. Pick one approved tool. Then train the team on real, green and amber work, so the first thing they learn is how to do it properly.

Safe use is the first module of every AI training programme we run, because the firms that get it right early are the ones that get to use AI on the work that matters.