Forecast AI usage with customer cohorts, not one average

A forecast built from "total customers times average usage" is tempting because it fits on one line. It's also fragile. New customers poke around, established ones settle into routines, and a few successful accounts wire up automations that run far more often than any human clicks.

Grouping customers makes all of that visible - without turning your forecast into a complicated simulation.

Group people by what you can actually see them do

Start with a handful: trial, just-paid, established, and heavy. The labels matter less than having groups that genuinely behave differently. For each, measure actions per account, how much text per action, which models they use, and how often things fail.

Don't go so far that every account is its own group. Three to five is usually enough to show the economics and still make sense to the people making pricing and capacity calls.

Let customers move between groups

A customer who signed up in June shouldn't be treated like one who's used the product for a year. Estimate how many move from trial to paid, how many become established, how many grow heavy, and how many leave. Then apply the usage pattern for whichever stage they're in that month.

This puts the growth costs in the right place. A promotion might create lots of cheap trials now and a smaller wave of heavier paid usage later. A single average either overstates the first month or understates the later ones.

Treat automation as its own thing

API access and scheduled jobs change the whole shape of usage. A person takes natural breaks; an automation runs overnight and at weekends. Split the automated accounts from the hands-on ones, even when they're on the same plan.

Watch requests per day and how many happen at once, not just the monthly total. The bill can look fine while bursts of traffic quietly cause capacity or rate-limit headaches.

Forecast a range, not a single ceremonial number

Build a low, expected, high and worst-case scenario by changing a few things that matter: conversion, retention, how active each account is, and what share are heavy users. Keep the message-size and model assumptions in view too - a product change that adds richer context can push cost up even when customer behaviour stays flat.

A range is only useful if it leads to a decision. Note which spending or capacity threshold each case crosses, and what you'd do about it.

Check it against reality every month

Compare your forecast to actual usage, group by group. If the total was off, work out whether it was the customer count, their behaviour, the model mix or the cost per action - and fix that assumption, rather than smearing a vague correction across the whole spreadsheet.

Over time, the forecast turns into a short story of how people adopt your product. It shows whether they're finding more value, whether your free plan is pulling in the wrong kind of usage, and whether your pricing still matches the cost it creates. That's far more useful than being able to say last month's average was technically correct.

Related reading and tools