A wealth manager replaces shadow AI with an approved workspace

The firm

Halden Row Wealth is a fictional UK wealth manager with around 400 staff across advice, client services, investment, marketing, operations and compliance. It serves retail and high-net-worth clients, so the Consumer Duty applies to much of its business.

The challenge

A short staff survey run by Compliance found that people in most departments were already using AI:

• Client services staff were pasting client emails into a public chatbot to draft replies.
• Advisers were using an AI note-taker on some client calls.
• Marketing was drafting newsletter content with a free AI tool.
• Operations staff were using AI to reformat spreadsheets — some of which contained client account details.

None of this was malicious. People were trying to work faster. But client data was leaving the firm, AI-drafted text was reaching clients with uneven review, and there was no record of any of it.

The Head of Compliance considered a ban. The survey responses suggested it would not work: people would move to personal devices, and the firm would lose what visibility it had.

What the firm decided

The Chief Operating Officer, as the senior manager accountable for the AI framework, agreed a plan with Compliance, Risk and IT:

1. Publish an AI acceptable use policy listing approved tools, data rules and what needs approval, and collect acknowledgements from all staff. See writing an AI acceptable use policy.
2. Provide an approved AI workspace with access to capable models from more than one provider, so people had a better option than the public tools.
3. Put controls in the path of every request: models approved by role, a data-classification ceiling for each model, and detection of client identifiers.
4. Register every use case found in the survey, decide which to approve, and stop the rest. See building an AI use case register.
5. Train and attest: a short role-based course, then a quarterly attestation that lets people declare exceptions.

How it works day to day

A client services associate asks the workspace to summarise changes to the complaints procedure for a team briefing, and includes a colleague’s email address. Before anything is sent:

• The model is checked as approved for her role.
• The request is checked against the firm’s AI policy: internal summarising is permitted.
• The email address is detected and replaced with a placeholder before the request leaves the firm.
• No FCA-linked control is triggered, and no approval is needed.

She sees the result of each check under her message, and the decision is recorded.

When a marketer asks the workspace to draft a newsletter promotion, the request is held for Compliance approval before drafting begins, because financial promotions are a high-risk use under the firm’s rules.

When someone tries to paste a client’s sort code and account number, the request is blocked with an explanation, and nothing is sent.

What the attestation showed

In the first quarterly attestation, a handful of people declared an exception to “I have used only approved AI tools for firm work”. One had used a browser translation extension on a client letter. Rather than a disciplinary matter, it became an incident record and a fix: the firm added an approved translation route to the workspace guidance.

What the evidence looks like

Without a separate evidence-gathering exercise, the firm can show:

• Who acknowledged which version of the AI policy, and when.
• Every AI request’s policy checks and outcome, with people pseudonymised in monitoring views.
• Which requests were masked, held or blocked, and why.
• Training completion and knowledge-check results by department.
• Attestation results, exceptions and their follow-up.
• The use case register with approvals and review dates.

What to take from it

• Start by listening. The survey produced a more accurate picture than technical monitoring alone would have.
• Give people a better route. The approved workspace had to be genuinely useful to replace the workaround.
• Make controls automatic. Masking and blocking prevented accidents that policy alone would not.
• Reward honesty. Exceptions declared in attestations became the firm’s best source of information.

See the AI Workspace or read about shadow AI.