A 1:1 Tool Can Become Surveillance Software

EMpath started as a 48-hour solo build for a familiar management problem. Notes are scattered, action items disappear, and by review time both the manager and engineer are reconstructing months of work from memory. AI can help connect those sessions. It can also make the product much worse if it erases the boundary between private preparation and shared feedback.

That boundary became the main architecture decision. A manager's raw notes stay manager-owned. The engineer has a separate self-profile rather than a second view over the manager's profile. When a 1:1 is completed, the engineer can receive a generated second-person summary and explicit action items, but not the private notes that produced them. This is not a cosmetic permission check. It shapes the tables, API routes, role guards, and the prompt context each job is allowed to read.

I considered a shared record because it looks transparent and is easier to explain. The problem is that it changes the behavior of the meeting. If every rough observation is automatically exposed, the private workspace becomes a surveillance feed and managers either stop recording useful context or write for the audience. Merging the manager's growth profile with the engineer's self-assessment creates a different problem: nobody can tell who owns a statement or who should be able to correct it.

The implemented path is narrower. Invitations and onboarding establish the role boundary. Manager notes and the manager-owned growth model sit behind that boundary. After the session, three independent AI jobs handle the derived work instead of one long request pretending that summary, actions, and growth processing have the same failure mode. Seven schema migrations during the build are not a vanity number. They show how much the data contract moved while those ownership rules became explicit.

There is more to do before I would call the privacy model mature. Retention and deletion need user-facing controls. Role changes need an audit trail. Background processing needs durable, idempotent delivery at higher volume. Generated summaries also need an editing and correction path without revealing the private source to the wrong person.

The useful outcome is not that AI writes nicer 1:1 notes. It is that the model works inside a boundary the product can explain: private input stays private, shared output is deliberately generated for sharing, and deterministic authorization owns the difference.