Churn Prediction from CRM and Usage Signals for UK SaaS SMEs

A 200-seat UK SaaS business pulled their Q3 renewal list in February. Of the 47 accounts renewing in June, 34 had not logged in to the core feature in the previous 28 days. Nine had raised a support ticket in the past 60 days that closed without resolution. Sixteen had reduced their active seat count by more than 30% in the quarter. All of this data was in their product database and their HubSpot. Nobody had connected the two.

Most UK SaaS SMEs between 20 and 500 seats have the raw material for a working churn model already in their stack — product event logs, support ticket history, seat counts, last-login timestamps. The problem is not missing data. It is the absence of a weighted model that reads those signals and produces a single score per account with enough lead time to act.

The churn signals hiding in your CRM: why most UK SaaS teams have the data and no model to read it

HubSpot stores contact activity, company properties, deal stage history, and support ticket records. Your product database stores session logs, feature events, and seat activity. These two data sources rarely talk to each other in a UK SaaS SME context — there is no pipeline writing product events to CRM properties and no workflow watching those properties for deterioration.

The result: renewal conversations happen with no account health context. A customer success manager opens a record six weeks before renewal and sees last email opened, last meeting booked, deal amount. They do not see: core feature sessions down 60% month-on-month, three of eight seats inactive, one support ticket closed without resolution. Those signals are in different tools and nobody has built the join.

The fix is custom HubSpot company properties, a nightly sync from your product database, and a weighted score property that HubSpot workflows can act on. Built correctly, this takes two to three weeks and produces a churn probability score visible on every company record before the renewal list is pulled.

For a worked example of how we structured a similar scoring model using CRM close data, see our post on predictive lead scoring from CRM close data — the property architecture transfers directly.

Usage-based churn indicators: the five product signals that predict non-renewal 6–8 weeks out

These five signals have the highest predictive value at the 6–8 week pre-renewal mark across the UK SaaS accounts we have worked with:

1. Core feature session frequency delta — percentage change in sessions on the primary value-driving feature over the past 28 days versus the prior 28 days. A drop above 40% is material; above 60% is a near-certain indicator.

2. Active seat ratio — active seats (any login in the past 14 days) divided by contracted seats. Below 0.5 is a strong signal for team-plan accounts. An account at 0.2 has essentially already cancelled.

3. Unresolved support tickets — any ticket that closed without a resolution status in the past 90 days. One unresolved ticket doubles churn probability. Two or more, combined with low session frequency, puts an account above 70% in the cohort data we have seen.

4. API call volume decline (for API-integrated products) — a drop in authenticated API calls of more than 50% over 30 days indicates the integration is being deprioritised or removed from the customer's stack.

5. Time-to-value on last onboarded seat — if the most recently activated seat took more than 14 days to complete the core workflow for the first time, onboarding failure is compounding disengagement across the account. This is a leading indicator, not a lagging one.

None of these require a data science team to compute. They are SQL queries against your product database, materialised into HubSpot properties on a nightly schedule.

Firmographic churn risk: the account characteristics that double churn probability regardless of usage

Some accounts churn because they never fit the ideal customer profile in the first place. These firmographic factors compound usage signals when present:

Single-stakeholder accounts — one named contact on the HubSpot record, no secondary contacts added in the past six months. When that contact leaves, the account follows. Flag records with fewer than two active contacts.

Sub-£3k ACV accounts on annual contracts — the lowest ACV tier consistently shows higher churn because the purchase required less internal approval and cost gets re-evaluated more aggressively at renewal.

Champion departure — detectable via a contact property change or a LinkedIn update via Clay or Apollo enrichment. A new contact at the same account with no product session history is a new sales opportunity, not a renewal.

Discount-sourced accounts — flag accounts with a non-zero discount on the original deal. Price-led acquisition predicts price-led departure.

Build these as binary flag properties on the HubSpot company record. A firmographic risk count of two or more combined with a usage score above 60% signals escalation to a direct call over an automated sequence.

Building a churn score in HubSpot: custom properties, workflow triggers, and the weighted model that flags at-risk accounts

Create these custom company properties in HubSpot:

• core_feature_sessions_28d_delta — Number (populated by nightly sync)
• active_seat_ratio — Number (0–1, populated by nightly sync)
• unresolved_tickets_90d — Number (populated by nightly sync)
• churn_risk_firmographic_flags — Number (count of binary risk flags)
• churn_probability_score — Number (0–100, computed by workflow)
• churn_intervention_tier — Single-line text (Green / Amber / Red)

The weighted score configuration we deploy:

{
  "score_weights": {
    "core_feature_sessions_28d_delta": 35,
    "active_seat_ratio": 30,
    "unresolved_tickets_90d": 20,
    "churn_risk_firmographic_flags": 15
  },
  "normalisation": {
    "core_feature_sessions_28d_delta": "invert_and_scale_0_to_100",
    "active_seat_ratio": "invert_and_scale_0_to_100",
    "unresolved_tickets_90d": "cap_at_3_scale_0_to_100",
    "churn_risk_firmographic_flags": "cap_at_4_scale_0_to_100"
  },
  "thresholds": {
    "green": "0-40",
    "amber": "41-79",
    "red": "80-100"
  },
  "computed_property": "churn_probability_score",
  "update_frequency": "nightly_cron",
  "workflow_trigger": "company_property_updated:churn_probability_score"
}

A HubSpot workflow reads churn_probability_score each time it updates. If a company crosses from Amber to Red, the workflow enrols the account in the intervention branch, sends a Slack alert to the CS rep, and creates a CRM task with the five signal values pre-populated in the description. The rep gets context before they pick up the phone.

Connecting product analytics to CRM: Mixpanel, Amplitude, and Segment events written as HubSpot properties

Segment's HubSpot destination writes group and identify calls directly to HubSpot company and contact properties. If your product already fires Segment events, push computed metrics — 28-day session count, feature adoption rate — to HubSpot company properties without custom sync code. Configuration takes under an hour.

Amplitude and Mixpanel both support scheduled data exports that feed an n8n workflow. The pattern: export a daily CSV of per-company aggregates, transform the column names to match HubSpot API property names, then PATCH the Companies API in batches.

For teams without any analytics platform, a nightly n8n workflow with four nodes covers it: Postgres query for the five signal values per account, function node computing deltas and ratios, HubSpot lookup matching on company domain, and HubSpot update node PATCHing the properties. The HubSpot CRM Companies API handles 100 requests per 10 seconds on a Professional plan — sufficient for 500 company records in under two minutes.

For a practical example of connecting structured data to downstream automation, see the voice AI document analysis case study — the same integration pattern applied to a document pipeline.

Totango's customer health score model argues that scoring belongs inside a dedicated customer success platform because CSPs offer richer workflow primitives. The counterargument for UK SaaS SMEs under 500 seats: CS teams live in HubSpot, and a score that exists in a separate tool is a score that does not get checked daily. Keep the score where the team already works.

Intervention routing: what to trigger when an account crosses the churn threshold (CS alert, voice call, win-back sequence)

Red tier (80%+): direct CS call within 24 hours, no automated email sequence. At 80%+ churn probability, email sequences get ignored — engagement rates below 20% are typical for accounts that have not logged in for 30 or more days. A direct call from the assigned rep, with signal values visible before dialling, changes the conversation from a generic renewal push to specific problem-solving.

Amber tier (60–79%): automated sequence plus CS notification. Day 1: email acknowledging low usage and offering a product walkthrough. Day 4: in-app prompt tied to the underused feature. Day 8: personal email from the rep's own inbox using HubSpot personalisation tokens that reference the account's specific signal values — the rep's name on a generic template is worse than no email at all.

For Red-tier accounts below £5k ACV, a voice agent renewal call is better ROI than CS headcount. The agent reads churn_probability_score, active_seat_ratio, and unresolved_tickets_90d from HubSpot as call variables, references them in conversation, and routes to a human only when the account expresses willingness to renew on different terms.

At the intervention trigger point, pause any upsell or expansion sequences on the same account. An expansion campaign on an account 80% likely to churn contradicts the retention message and in some cases accelerates departure.

Renewal probability by cohort: the segmentation that separates accounts worth saving from those already decided

Not all at-risk accounts warrant the same investment. Segment by ACV and churn score before allocating CS time:

The "already decided" cohort is commonly ignored. Accounts at 90%+ churn probability with zero logins in the past 60 days are not going to renew. Routing them into the same intervention queue as recoverable accounts wastes CS capacity and inflates your failure rate. Track them separately, review the signals, and use the pattern to refine the model.

For accounts that churn despite intervention, the customer win-back automation framework covers the 90-day re-engagement cadence in detail.

What changed in 2025–2026: HubSpot Breeze churn prediction and AI-native product analytics platforms

HubSpot's Breeze AI layer, generally available from mid-2025, added predictive contact scoring that factors in engagement signals across email, meetings, and support tickets without custom property configuration. The limitation: it scores contacts, not companies, and it does not read product event data from outside HubSpot unless you write it in via the API. For team-plan SaaS accounts where the champion is one contact out of 40 seats, it misses most of the signal.

Amplitude's retention forecasting features, expanded in late 2025, generate per-cohort churn probability directly within Amplitude without a separate ML pipeline. Forecasts update daily and export to HubSpot via webhook. The limitation: Amplitude's signal weighting is not adjustable, so the score reflects their assumptions, not yours.

The direction: CRMs and analytics platforms are both building churn prediction natively. The value of a bespoke ML model is declining for straightforward cases. The value of the data integration — connecting product signals to CRM properties where CS teams work — is not declining. The same integration layer supports AI sales forecasting from CRM data — worth building alongside the churn model because both draw from the same nightly property sync.

Good / Bad / Ugly: three churn prediction approaches and their 90-day retention impact

The "Bad" failure mode is nearly universal in UK SaaS SMEs under 150 seats: renewal conversations happen when a CS manager creates a task triggered by a date, not when a usage signal crosses a threshold. By then, the account's decision is often already made and the conversation is about terms, not retention.

The "Ugly" failure mode hits SMEs that bought a customer success platform before solving the CRM integration problem. Gainsight has strong scoring capabilities but requires dedicated administration. If the CS function is two people who live in HubSpot, a Gainsight implementation produces a score nobody checks daily. Buy the integration before the platform.

Build the model in the tool your team already has open. Add external data to it. Do not add a new tool and expect behaviour to change.