When a team runs nine product forms for four kinds of users and operates three content factories, every project can explain why it matters. Without a common denominator there is no basis for trade-offs, and resource allocation degrades into whoever tells the best story.
That is the problem a north-star metric solves. It has to reflect both how many businesses use the product and how deeply, because those are the two input ports of the data flywheel.
One multiplication
Weekly agent tasks = weekly active businesses × tasks per business
Why a multiplicative composite? Counting only active businesses rewards shallow sign-ups. Counting only tasks rewards a handful of heavy users gaming volume. Multiplication forces both legs to move: the number of businesses sets the breadth of transaction data, task depth sets the density of interaction data.
The definition of a task has to be strict: one complete unit of work triggered by a real user, completed by the core engine, and consumed by a person. A conversation counts once. A brief counts when opened, not when generated. An export or an adoption counts.
Three anti-gaming rules
Goodhart's law: once a measure becomes a target, it stops being a good measure. So the rules are fixed up front:
- Machine-triggered doesn't count. A proactive agent generating ten thousand cards overnight moves the north star by zero. Only what people open, follow up on and adopt counts. This one rule turns content-factory evaluation from output to consumption automatically.
- Internal accounts don't count. Marketing, sales and the AI team's own usage are excluded and tracked separately as a dogfooding reference.
- Active businesses are deduplicated by business, not by seat. Aligned with the fact that the paying party is the business.
Every product claims one step
The north star decomposes into three steps: sign-up → activation → deepening. Each product claims the step it mainly acts on and reports that step's conversion. "Contributes to everything" is not allowed in a weekly report.
The side effect was better than expected. It forces every project to answer a simple question: which part of the funnel did you actually change? Projects that can't answer usually can't explain why they exist either.
Baselines talk
A few baseline numbers side by side say more than any one alone. The gap between registered and active tells you the cheapest growth is activation, not acquisition. Cumulative tasks divided by active businesses tells you how many orders of magnitude depth is from the vision. Cost per call times adoption rate tells you what each adopted task costs, which is the real number you need to judge whether the business model holds.
Guardrails get red lines, not targets
Any single metric can be bent. So add five guardrails: task adoption rate, deliverable export rate, business retention, cost per task, and opt-in rate for proactive pushes. Guardrails get no KPI targets, to avoid a second Goodhart, only red lines: if the north star rises while a guardrail breaks, the growth is invalid and you go find out why.
Instrument first
One number for the whole, a decomposition tree for the why. But all of it depends on instrumentation: task-level events, a business-level identifier threaded through every product, a "brief opened" event, cost attributed per task. Without that, the north star is just a pretty formula.