The customer who leaves quietly
The most useful line to come out of Preseem's Proactive ISP Virtual Summit in May 2026 wasn't a product announcement. It was an observation from a session on churn prediction: most customers who cancel their broadband never called support first. They just leave. For an ISP that relies on tickets and complaints as its early-warning system, that means the signal arrives after the decision has already been made.
The summit, held on 5-6 May 2026, was built around practical uses of AI in ISP operations. Its churn session featured speakers from QueSee AI and from Wisper, a US fixed-wireless provider, and it sat alongside Preseem's announcement that it is combining network quality-of-experience data with QueSee AI's customer-interaction analytics. The stated goal: give regional ISPs real-time context to cut both churn and unnecessary truck rolls.
What a churn model actually looks at
According to the summit recap, Wisper's model pulls together router performance, radio telemetry, support-ticket history, billing patterns, and what competitors are available in each customer's area. It produces a ranked list of at-risk subscribers, and Wisper's top-100 list was reported to be about five times more accurate than random outreach. A rural-ISP panel described scoring on more than 50 data points, including whether Starlink is present in a customer's ZIP code.
Sonar, a billing platform in the US market, describes the same idea from the support side. Its retention add-on, built with QueSee, scores every customer interaction for sentiment and competitor mentions, and watches for patterns like repeated billing frustration or tickets that keep reopening. Sonar says 350 Broadband found that only 9% of its calls needed action once AI sorted them, freeing staff to focus on subscribers who looked genuinely at risk.
A fair caution: nearly all the performance figures here come from vendors and operators presenting their own results, not independent audits. The summit's own illustrative maths is an example, not a promise -- an ISP with 10,000 subscribers and 1% monthly churn that saves 5% of controllable cancellations recovers roughly $50,000 a year. The direction is credible; treat the precise numbers as best-case.
Why this matters outside the US
None of these operators are in Kenya, and a model fed by FCC availability maps won't transfer. But the underlying logic travels well. In markets where prepaid PPPoE and hotspot plans dominate, the churn signal often isn't a complaint at all -- it's behaviour. A subscriber who used to renew on the 1st and now waits until the 9th. Someone whose usage has been falling for three weeks. A customer who hit repeated session drops and never reported it.
- Payment timing: late or irregular renewals are an early, cheap-to-measure sign that a client is shopping around.
- Usage trend: a steady decline in data consumed often precedes cancellation.
- Connection quality: repeated disconnects or poor signal that the client never raised.
- Competitor pressure: new fibre or wireless coverage on the same street.
The data layer comes first
One of the summit's more practical points came from Preseem's co-founders: AI agents can only reason reliably if the operator has a clean, normalised data layer, and most ISPs run equipment from five or more vendors. That's the unglamorous truth behind every AI-retention pitch. A model is only as good as the billing records, session logs and payment history feeding it.
This is where an ISP billing and RADIUS platform quietly does the heavy lifting. XpressRADIUS keeps PPPoE and hotspot sessions, M-Pesa and Kopo Kopo payments, plan renewals and usage counters in one system, with automated disconnect and reconnect tied to payment status. That means the signals a churn model needs -- when someone paid, how much they used, whether their sessions kept dropping -- are already structured and queryable rather than scattered across spreadsheets and chat threads. Even without a machine-learning model, a simple weekly list of "clients whose renewal slipped" or "clients whose usage halved" is a retention tool most small ISPs aren't using yet.
Start with the list, not the algorithm
You don't need a data-science team to act on this. Pull last month's cancellations, look at what their payment and usage behaviour looked like in the 30 days before they left, and see how many shared a pattern. Then call those same-pattern clients this month. The ISPs that win retention in the next few years will be those who reach out before the silent customer has made up their mind.
If you'd like to see how much of that data your billing system could already be giving you, start a free XpressRADIUS trial and explore real-time PPPoE and hotspot billing built for growing ISPs.
Ready to see every payment, session and renewal in one place?
Start a free trial