majordome.The Monday before
A client meeting produces two things: a conversation and a list of commitments made out loud. Send a document, call back at a given date, check a policy, come back with a quote. In this agency, those commitments lived in a notebook, on sticky notes, and in the agent’s memory.
Around it, a locked-down environment: the business desktop on Citrix, email under constraint, the contacts in the phone. A task application had been tried and then abandoned, for a reason you rarely hear in a specification: “when you tick it off, it disappears”. Creating tasks was not enough for the agency. It wanted to keep the record of what had been done.
What it was costing
The notebook holds up as long as the day is quiet. Past a certain pace, it becomes a gamble. And in this job, a reminder that does not go out is not an administrative oversight: it is a relationship being damaged, with someone who trusted you with their home or their health.
On top of that constraint comes a second one, non-negotiable: the portfolio data is policyholder data. It cannot go wandering.
How we worked
July 2026. A work plan in seven strands, then the foundation: the connection to the agency’s Google workspace and the import of more than a thousand contacts, so that the register starts from something real.
Before touching an AI, the rule on the data. A pseudonymiser runs on the workstation and replaces the identities in the portfolio exports, tested on a real export. Identities are replaced before any AI processing.
Then the register, and the chain. The agent records the meeting with a small recorder. The recording becomes a summary, the summary becomes proposed tasks, and every task waits for the agent’s approval before it exists. The chain has been in production since 5 August. Deployment goes through an API now, no longer through copy and paste.
Then operations: a grid limit fixed at the end of August, a column mapping fixed at the beginning of September, seventeen versions of the register. Going live opens a phase, it does not close one.
What is running today
A register that holds the follow-up in a shared structure, with the contacts as the anchor. A chain that turns every recorded meeting into a summary and a set of tasks, without re-entry. A pseudonymisation tool that prepares the processing without exposing the policyholders.
The same engine is at three different stages: this agency in production, a second in pilot, and a third being delivered.
Technical: register application, Google Sheets, integration of a Plaud recorder, local pseudonymisation, deployment through an API.
What changed, and what remains
The notebook is no longer the weak link. The agent comes out of the meeting with the written summary and the commitments laid out in front of them, to approve or to set aside. They keep control over what exists and what does not.
The signal I hold on to: the agency’s staff member sent me, of her own accord, a running list of requests she had prepared with an AI. When the team starts putting needs into words on the basis of the system, the system has been adopted. What remains: measuring the use over time, before talking about commercial gain.
If this sounds familiar
If you work in a regulated profession and that is the reason you have not tried anything yet, that constraint is an input to the diagnostic, not an obstacle you discover at the end.