24 agents designed. 3 in daily use.
An operating model for a firm whose best people spent hours finding information it already owned.
Where the time went
The work started with a normal week. Partners and staff walked through how their days actually went, role by role. The pattern was the same everywhere: experienced people spending hours finding things the firm already had. An email thread, a deadline in the case system, last month's numbers. The work was not hard. Finding the inputs was.
So the design question was never "where can we use AI?" The question was which of those searches an agent could do instead, and who that agent should answer to.
One job, one human
The operating model has 24 agents. Each one has a single job, a defined set of data it may read, and a named person it reports to. Three are in daily use:
- An executive assistant in Microsoft Teams that turns overnight email into a short list of what needs the founding partner, with replies drafted for review and never sent on their own.
- A CFO agent that answers financial questions from a read-only snapshot. Only people cleared to see the firm's finances can ask, and the agent declines anything outside that job.
- A marketing agent that supports the firm's marketing operations and reporting.
The other 21 are designed and waiting until their workflows are ready. Three agents people rely on are worth more than twenty-four nobody trusts.
Inside the firm's walls
- Everything stays in the firm's tenant. The app, the data and the models run inside the firm's own Microsoft and Azure environment.
- Default-deny access. A new user sees nothing until someone grants access, and the check happens on the server.
- Search trimmed to the person asking. One index, filtered at query time by what the signed-in person may see.
- A four-zone data architecture that sets out where each kind of data may live and what may cross between zones. The firm's IT provider reviewed it and cleared it.
- Costs metered from day one. Email triage runs at about $0.0003 per message.
Reporting also moved off a third-party vendor and onto an in-house data layer in the firm's own Azure, so the agents read the firm's own numbers directly.
What this means for your firm
Every firm's systems are different, so no two builds are the same. The method is: map where the time goes, build one agent at a time inside your tenant, and keep each one running as your practice changes.