The problem
An advisory firm's most useful data for AI is exactly the data it can't share: CRM notes, client email, and tax documents. I wanted a repeatable boundary for what an AI-assisted workflow could receive, so I built one for my firm.
How it works
The third-party sources I currently use in these workflows, including Wealthbox CRM, Outlook email, and uploaded documents, are routed through a local proxy. The proxy uses a central masking layer built on Microsoft's Presidio engine plus custom rules: key-based masking for structured fields, denylists built from CRM contact records, and regex fallbacks for emails, phones, SSNs, account numbers, and street addresses.
Masking is contextual. Client and household names become pseudonym tokens so a governed workflow can preserve attribution without including the original names. Dates are preserved when chronology matters; birthdates are masked because they're identifiers. Responses are minimized by default. Raw audit payloads stay local and are available only when explicitly requested for human review.
I use this workflow at my firm to turn masked CRM notes and email into governed source records for tasks such as meeting prep, follow-ups, and document review.
My role
I designed and implemented the proxy, masking package, connector integrations, and audit model. The judgment calls about what must be masked and what planning context must survive came from being the CFP® accountable for the workflow.
Walkthrough
The code is in a private repo and the system touches real client data, so the working artifacts stay off this page. On a call I'll walk through the architecture, show a masked-vs-raw example from synthetic data, and share the design decisions that would transfer to your firm.
Discuss similar work
If your firm is working through data boundaries for AI-assisted workflows, it's the same class of problem I've been solving at mine.