Every analytics tool groups users by what they bought. Free, Plus, Team, Enterprise. This tells you nothing about what they actually do with the product, and the people who do the same thing never share a tier. The fix is the same fix that makes [product analytics for AI agents](/blog/what-is-locus) different from SaaS analytics in the first place.
Why is plan-tier segmentation the wrong unit for AI agent analytics?
Plan tier. Country. Signup source. These are facts about the user, not about the use. They are easy to store in a row, easy to filter on, and easy to chart. They are also almost useless for product decisions on an AI agent. Two Plus users can behave nothing alike. One is a writer drafting emails. The other is an analyst running SQL through natural language. Grouping them by plan reveals nothing about either.
This is why the data your team already has is unreadable: even when you have the conversations, the cohorts your dashboards are built around are demographic, and the patterns that matter are behavioural. The shape of the data is wrong.
What is a behavioural cohort for an AI agent?
A behavioural cohort is a set of users defined by what they do — the shape of their conversations, the tools they call, the use cases they return to — rather than by what plan they pay for or which country they signed up from. These cohorts emerge from the conversations themselves. They cut across every plan tier, every country, every signup source. And they are the cohorts whose drift actually tells you something about the product.
The six groups Locus surfaces on a typical multi-purpose agent:
- Writers — prose drafting, often paired with creative brainstorming.
- Code-first — mostly code, with a learning tail.
- Researchers — summaries, explanations, citations. High question-rate, low tool use.
- Analysts — spreadsheets, SQL, numeric work.
- Generalists — broadly spread across four or more use cases. Often the curious first-month user.
- Advice-seekers — career, decisions, life moves.
These six groups account for around 92% of the users on a typical 1,000-user agent in production. The remaining 8% sit in an ‘Other' tail that is itself a useful signal — when the tail grows, a new use case is starting.
Why is behavioural segmentation hard for AI agents?
Behavioural segmentation is hard for the same reason reading conversations is hard: free text. A SQL query can slice by plan tier in a second. It cannot slice by what the user was trying to do — because that is not a column anywhere. The group has to be inferred from the conversation. That requires the same content read that traditional product analytics tools (Mixpanel, Amplitude, PostHog) cannot perform on plain text.
It is also hard because the groups drift. A user can start as a Generalist in their first month and converge into a Code-first user by month three. Demographic cohorts do not change. Behavioural ones do. That movement is itself the leading indicator most teams want — but only if there is a system reading every conversation to detect it.
What changes for an AI PM when cohorts become behavioural?
Three things. First, retention math becomes legible — you can see which group is leaving and which group is sticky, instead of seeing a single retention curve that averages over six different products inside one. Second, the roadmap becomes specific — a feature that helps Researchers may do nothing for Writers, and you can finally tell the two effects apart. Third, drift detection becomes possible — when a group's behaviour shifts, you see it in the cohort weeks before it shows up in the aggregate.
Most teams discover at least one cohort they did not know they had. Often it is a use case that grew in a corner of the product without ever being named — the team treated it as noise, and the cohort had been growing 5x quarter-over-quarter unnoticed. That is the kind of thing reading by sample systematically misses.
Locus reads the free text and produces these groups automatically. See six of them in a live sample, or book a call and we will produce them on your own runs.
Frequently asked questions.
How is product analytics different for AI agents than for SaaS?
SaaS analytics counts events — clicks, sessions, page-views — and segments by demographic facts. AI agent analytics has no events to count, because the product is the conversation. The unit of behaviour is the conversation, segmented by intent and grouped by behaviour. Plan-tier segmentation is even less useful in AI products than in SaaS because two paying users can have nothing in common about how they use the product.
How do you separate behavioural cohorts from plan-tier cohorts in practice?
By reading every conversation, classifying it by intent, and clustering users whose intent mixes look alike. The clustering is unsupervised — the team does not have to define the groups. The groups emerge from the data. Plan tier becomes a filter on top, not the basis of segmentation.
What is the difference between behavioural cohorts and personas?
Personas are aspirational descriptions written by the product team. Behavioural cohorts are descriptive groups derived from observed behaviour. Personas describe who you wish your users were. Behavioural cohorts describe who they actually are. The two often disagree, which is when the cohort read becomes most valuable.
How many users do I need before behavioural cohorts are stable?
Around two thousand conversations a month produces stable groups. Below that the noise dominates and reading by hand is more reliable. Above that, six-to-eight groups typically emerge clean and account for around 90% of users.
Can I use behavioural cohorts for retention reporting?
Yes — and for most AI agents this is the only retention reporting that means anything. A single retention curve on an AI product averages over multiple use cases that have nothing to do with each other. Per-cohort retention reveals which group is sticky, which is leaking, and which is silently growing.