Who defines
“active user”?

I / One definition

Ask two people what “active user” means. You get two answers.

The CEO owns the definition. Every agent reads it.

• Our CEO once asked how many paying users we had, at 11 p.m., before an investor meeting. The answer looked great. The real number was lower: a trial user is Pro for a while, and a nightly job moves them back to Free.
• The fix was not a better query. It was one place where the words are defined: a metrics glossary, written as a book, that the CEO owns.
• Every agent reads that book before it answers. The semantic layer is a book, so a person can read it, change it and see its history.
• A department can keep its own definitions. Product can count “active” differently from finance, as long as every analysis says which definition it used.

II / The analyst in Slack

An analyst agent that lives where the questions are asked.

When a number looks wrong, the CEO tags a person in the same thread.

• Anyone tags the agent with a question, such as “What can you tell me about the last paying user?” It answers in about a minute, from the database, the product events and the web.
• Every morning it posts a brief to the CEO: the sign-up funnel, daily active users, and a note on what moved and why.
• It runs in Slack, not in each person's own chat app. One team controls the prompts and the data access, and every answer is in the open, where someone can question it.
• Each agent keeps its own memory on top of the shared glossary: who asked for which report, and how they like it. Jev, a small, fast model, reads the memories on every message and picks the ones that matter.
• A memory can become a proposed change to the glossary. That is how a definition gets better without a ticket.

III / Dashboards

Build the dashboard you need. Let the rest expire.

Not a snapshot. A dashboard that runs its query when you open it.

• Our product dashboard is built in-house, on our own data, instead of a BI tool. Each release is a line on the usage chart, so you can see what a release changed.
• For one-off analysis, a Claude artifact connected to a data MCP server keeps the query, not the result. Whoever opens it gets fresh numbers and can comment on it.
• That covers the work-in-progress analysis that never deserved a permanent dashboard. Access still needs a decision: who may run which query.

IV / Maintenance

Agents that keep the queries fast.

We did not know to time every query. A book did.

• Our heaviest pages were slow. An agent audited every query against database books, found the waste, and cut the slowest query from about 1,078 ms to 1.2 ms on average, roughly 900 times faster.
• The audit found a known unknown: we did not record how long each query took. Now every query's timing is recorded, so the next slow one shows up in the numbers.
• Treat data work like refactoring. Tune queries, tables and precomputed views on a schedule, not only when something breaks.
• To choose the metrics in the first place, Lean Analytics helped most: find the few numbers that matter at your stage.

V / Adoption, and what still breaks

The agent comes last. The analysts come first.

A lot of what an analyst knows is only in their head.

• An analyst agent imposed from the top fails. The analysts hold the context, so they build it with you, or it answers the wrong question with confidence.
• Start in one team. A whole enterprise cannot work this way today, but a small team inside it can, and it shows everyone else what is possible.
• What still breaks: a definition drifts when nobody owns it, memories go stale, and anyone with read access and their own chat app can still get an unchecked number at 11 p.m.

VI / For teams

One glossary, many owners.

The CEO owns the book. Each department owns its chapter.

• Share a book with the whole team, or with one department. Every teammate’s agent reads the same definitions, and a book shared with Finance stays out of Sales’ agents.
• Each book has an owner, and each chapter can have its own: in our glossary, “Product Metrics” can have a different owner from “Churn” or “Active User”. The owner is the person to ask, and maintainers can edit it directly.
• Everyone else’s agent reads the chapter but cannot rewrite it. When it tries, it is told who looks after that chapter.
• When an agent cannot find an answer in a team book, it files a request. The chapter’s owner is notified, writes the answer in and marks it covered, and the person who asked is told.
• Every change to a book is saved as a version, and the owner can restore an earlier one.
• The team page shows how often agents read each team book this week, and the most-read page. Owners and admins also see an eight-week trend per book.
• Invite people by email as members or admins. Every member gets everything in Pro.

See the Team plan