Churn almost never arrives as a surprise. It arrives as a series of small signals that nobody was watching in one place.
The invoice that used to be paid on day 3 is now paid on day 28. Two support tickets in a month where there used to be none, and the tone in the second one is sharper. Logins drop off after the champion changes roles. The renewal is eight months out, so nobody is looking at the account yet.
Each of those lives in a different system. Payments in billing, tickets in the helpdesk, usage in the product, the renewal date in a contract PDF. No single person sees all four. By the time the renewal conversation happens, the decision was made months ago.
This is the problem the Customer Intelligence Agent exists to solve.
 |
What it does, in one line
It runs every morning, scores every customer across five dimensions, and tells you which accounts are quietly deteriorating and which are quietly ready to buy more.
Catching churn while you can still act
The agent produces a churn probability for every customer, built from six signals rather than one:
- Current health score
- Health score trend, which matters more than the absolute number
- Payment behaviour deterioration
- Support sentiment
- Engagement drop
- Renewal proximity, which amplifies everything else as the date approaches
The output is not just a percentage. It is a predicted churn window, a risk tier, a recommended play, and an owner. The plays are specific: executive outreach, discount offer, product review call, technical review, or simply monitor. A customer at 34% risk nine months from renewal needs a different response than the same customer at 34% six weeks out.
Support sentiment deserves a note here, because it is the fastest-moving signal. Every ticket and reply is scanned on ingest. Hostile or legal language triggers an immediate CFO and account manager alert and recalculates churn risk on the spot. You do not wait for the daily run to find out a customer threatened to involve their lawyer. Score deductions then decay over the 30 days after resolution, so one bad week does not permanently mark an account that recovered.
Finding the expansion you already earned
The same scoring picks up the opposite signal. Expansion rarely announces itself either. A customer quietly approaches their seat limit. Usage climbs month over month. Health is strong and payments are clean. Nobody notices until the customer eventually asks, or does not.
The agent classifies opportunity by type, because each needs a different conversation: seat expansion, module upsell, usage upgrade, tier upgrade, or multi-year conversion. It also flags the opportunity window, since expansion timing is as important as expansion detection. A customer at 90% of their seat limit is a conversation this month, not next quarter.
There is a related point worth making, because it is where real money leaks. Expansion signals often surface entitlement gaps: a customer already consuming more than their contract covers. That is not an upsell, that is unbilled revenue. The agent's revenue leakage detection is looking for exactly this.
The engine underneath both
Churn risk and expansion signal are not separate systems. They come from one composite health score, calculated across five weighted dimensions:
Two design choices matter more than the weights.
First, scores are normalised to a cohort percentile rather than judged against a fixed bar. A customer who has always paid on day 20 is not a problem. A customer who always paid on day 3 and now pays on day 20 is. Absolute thresholds miss that. Relative movement catches it.
Second, every score maps to a tier with a defined response: Platinum at 85 and above, Good 70 to 84, Average 50 to 69, Poor 30 to 49, and Critical below 30. At Critical, the CFO is looped in at step one rather than at the end of a playbook. The tier is not a label. It decides who acts and how fast.
The score also does not sit in isolation. When Collections sees an invoice cross 45 days overdue, it pushes an event that rescores that customer's health. When health goes critical, the agent pushes an alert to the CFO Co-pilot. The agents talk to each other, so a payment problem becomes a retention signal automatically.
What the agent does not do
It does not decide. This is deliberate.
The agent scores, predicts, flags and recommends. Whether you offer a discount, write off a bad debt, change the churn treatment, or restructure a contract stays with the CFO, logged with a name and a timestamp. Retention and expansion decisions carry commercial and revenue-recognition consequences, and those belong to a person who is accountable for them.
The agent's job is to make sure the decision reaches that person while there is still time to make it.
Why this is hard to do in a spreadsheet
You could build a version of this manually. Pull the AR aging, export ticket volumes, ask the product team for usage, cross-reference renewal dates. Teams do exactly that, once a quarter, for the twenty accounts someone is already worried about.
The problem is coverage and frequency. Risk does not appear only in the accounts you are already watching, and it does not wait for quarter end. A daily score across every customer is a different instrument to a quarterly review of your top twenty.
That is the whole argument for running it as an agent.
Visit- https://fincelo.app/ai-agents
