An AI usage policy and model inventory for a wealth manager

01

Context

The firm manages discretionary portfolios for around 2,000 households. Over the previous year, advisers and operations staff had started using public AI chat tools to draft client letters, summarize meeting notes and tidy up spreadsheets. Nobody had approved this, nobody knew which tools were in use, and some prompts contained client names, account values and health information shared in financial-planning meetings.

The Chief Compliance Officer wanted to allow AI where it helped, but needed to show the board, and eventually the SEC in an examination, that its use was known, governed and recorded. The firm had no in-house engineering team and no appetite for a policy that existed only on paper.

02

Constraints

Regulatory

As an SEC-registered investment adviser, the firm had to meet the Advisers Act books-and-records rules for business communications, keep client information confidential under Regulation S-P, and follow its own written compliance policies. The policy had to fit into the existing compliance manual and annual review rather than sit beside it.

Data

Client data could not leave the firm's Microsoft 365 tenant for any tool not under a signed agreement. Prompts and responses that formed part of client communications had to be kept for the same period as other client correspondence.

Integration

Everything had to work with the firm's existing Microsoft 365 setup, identity provider and CRM. Staff could not be asked to adopt a new portal for every request.

Timeline

The board wanted a policy in place before the firm's annual compliance review, eight weeks away.

03

What we built

We started with discovery rather than policy. Network and browser logs, expense reports and short interviews with every team showed fourteen AI tools in use, from public chat assistants to features switched on inside existing software. For each one we recorded who used it, for what, and what data went into it.

With that picture, we drafted an AI usage policy with the CCO: what staff may and may not do, which data can never be entered into any AI tool, when a human must review output before it reaches a client, and what must be kept as a record. The policy uses plain examples from the firm's own work rather than abstract principles.

We built the model inventory as a register in the firm's existing SharePoint, so compliance owns it without new software. Every tool and use case has an owner, a risk rating (based on the data involved, client impact and whether output reaches clients unreviewed), the controls that apply, and a review date.

For the uses the firm wanted to keep, we provided an approved route: a company assistant on the firm's own Azure OpenAI deployment behind a private endpoint, signed in with existing staff accounts, with prompts and responses kept under the firm's retention rules. Public AI tools were blocked for client data using the firm's existing data-loss controls.

Finally, we set up the review process: a short request form, a triage checklist, and a monthly review by the risk committee. Low-risk requests can be approved by compliance alone; anything touching client data or client communications goes to the committee.

• AI usage policy, approved by the risk committee and added to the compliance manual
• Model and use-case inventory with owners, risk ratings, controls and review dates
• Discovery report of every AI tool in use and the data each one received
• Approved company assistant on a private endpoint, with prompts and responses retained
• Data-loss rules blocking client data from unapproved AI tools
• Request form, triage checklist and monthly risk-committee review process
• One-hour training session for all staff, with a one-page summary of the rules
• Evidence pack for the annual compliance review

04

Architecture

Expand

Request path, the approved assistant inside the firm's tenant, and data-loss controls. Firm and tenant details have been removed.

Full size✕

05

Results

The policy was approved by the risk committee and adopted before the annual compliance review. Unapproved tools were retired or brought into the inventory, and the company assistant became the standard way to use AI with client information. In the following quarter, compliance reviewed and approved six new use cases through the request process and declined two that would have sent client data to an unvetted vendor.