OptimisedMost of the customers you lose this year will tell you they're leaving months before they actually go. They just won't say it in words, and if you're not watching the right signals you'll only find out when the direct debit stops.
Here's the short version. Churn prediction for small businesses doesn't need a data science team or a five figure software bill. It needs a clear definition of what "lost" means, a handful of behavioural signals from systems you already run, a simple scoring rule applied every week, and a person whose job it is to ring the people at the top of the list. Machine learning comes later, if at all, once you've got thousands of customers and a year of clean history. This piece walks through how I'd do it, what it costs in pounds, and which UK rules apply.
What Churn Prediction Means When You're Small
Strip the jargon away and churn prediction is just a ranked list of your customers, with the ones most likely to leave at the top. The practice grew up in big subscription businesses, where the output is usually a probability score between zero and one, but as Dimension Labs explains in its guide to predictive churn, the goal is simply to surface at risk customers early enough for someone to act before the cancellation is final.
For a firm with a few hundred customers, that ranking can live in a spreadsheet. For a few thousand, it can live in your CRM as a custom field. CRM, or customer relationship management software, is the database where you keep contact details and interaction history. Either way, the hard part isn't the maths. It's agreeing what churn looks like in your business and being honest about which signals you can see.
That definition matters more than people expect. A subscription business knows exactly when someone cancels. A trade supplier or an accountancy practice often has to decide that someone who hasn't bought in 90 days has probably gone. Settle that first, write it down, and don't change it mid year or your numbers become meaningless.
The Numbers Worth Knowing Before You Start
You'll want a baseline before you try to predict anything, so measure your current customer churn rate. Take the customers you lost in a period, divide by the customers you had at the start of it, and multiply by 100. If you started January with 200 customers and 14 had gone by the end of the month, that's a 7 percent monthly churn rate. Annualise it properly by converting to retention and compounding, rather than just multiplying by twelve, because compounding makes a real difference at those levels.
Benchmarks are mostly American and I'd treat them as rough context rather than targets. The most quoted figure comes from Recurly's analysis of more than 1,200 subscription sites, which Shno's churn statistics roundup reports as a median SaaS churn rate of 4.79 percent a month. SaaS means software as a service, the pay monthly software model. That's a US dataset with no UK equivalent of that size, so use it as a sanity check and nothing more.
Two findings from that US research translate well to Britain in my experience. First, price point moves churn far more than industry does. SubJolt's summary of Stripe's 2025 churn benchmarks puts the spread at 40 percent annual churn for orders under ten dollars against 15 percent above ten thousand. Second, a large chunk of churn isn't a decision at all. According to Ringly's compilation of 2026 churn statistics, somewhere between 20 and 40 percent of total churn is involuntary, meaning failed cards, expired cards and billing errors rather than customers choosing to leave. If you sell on subscription in the UK and you haven't looked at your failed payment recovery, do that before you build a single predictive model. It's the cheapest churn you'll ever fix.
Closure Risk Is Not the Same as Switching Risk
This is the point most generic churn advice misses, and it's especially important if you sell to other businesses in the UK. Some of your customers won't switch to a competitor. They'll simply cease to exist.
The Office for National Statistics recorded 280,000 business deaths in the UK in 2024, a death rate of 9.8 percent, which was actually the lowest since 2016. That means a UK business selling to other small firms loses accounts to closure alone at close to one in ten a year before a single competitive loss. As Peter Foster of UK Data Services points out in his guide to predicting customer churn for UK businesses, a model that separates closure risk from switching risk stops your retention budget chasing accounts that no offer can save.
Practically, that means keeping a column for "trading health" alongside your engagement signals if you sell B2B. Late payment creeping from 30 days to 60, a change of directors at Companies House, a sudden drop in order size with no complaint attached. None of those mean the customer is unhappy with you. They may mean the customer is in trouble, and that calls for a different conversation, often about payment terms rather than product.
The Signals That Actually Predict Leaving
Every business has its own tells, but I keep coming back to the same short list. Recency of activity is the strongest single predictor in almost every dataset I've touched. If a customer's gap since last order, login or visit is longer than their normal rhythm, they're drifting. A café regular who came in three times a week and hasn't been seen for twelve days is a bigger flag than a monthly customer who's a week late.
Frequency trend comes next. Not the absolute number of purchases, but whether the trend over the last three periods is down. A customer who ordered eight, then six, then four times is telling you something even though four is still healthy. Spend works the same way, and if you sell goods, watch basket size shrinking while order count holds steady. People trim before they cut.
Support behaviour is the odd one. A customer who's complained once and been dealt with well is often stickier than one who's never said a word. The pattern I'd flag is an unresolved ticket, a second contact about the same problem, or a complaint that went quiet. Silence after a complaint is worse than the complaint.
Onboarding is where subscription businesses lose the most people, and it doesn't announce itself. Kayako's guide to customer churn rate notes that customers who don't hit their first value milestone within 30 days are disproportionately likely to churn before their first renewal, and that onboarding failure usually looks like someone simply not coming back after week one. So your first predictive rule for new customers is dead simple: did they do the one thing that makes the product useful, and did they do it in the first month?
Then there are the administrative tells. Downgrades, a switch from annual to monthly billing, a data export request, a change of billing contact to someone in finance, a question about the notice period. Any one alone is noise. Two or three together in a fortnight is a customer building a case to leave.
Build a Scorecard Before You Build a Model
I'd start every small business on a points based scorecard and I'd keep most of them there for at least a year. It's transparent, it's cheap, and it forces you to think about your customers rather than about software.
Start by exporting a list of every active customer with five columns: days since last activity, order or usage count for each of the last three months, total spend in each of those months, open or unresolved support tickets, and any admin flags from the list above. If you're on Shopify, its built in customer reports already do a version of this, grouping customers by recency, frequency and monetary value into eleven segments including "at risk" and "almost lost". According to Shopify UK's roundup of churn prediction software, those reports are included on all plans, so check them before paying for anything.
Then assign points. My starting weights, which you should adjust after a couple of months: 3 points if the gap since last activity is more than 1.5 times that customer's own average gap, 2 points if frequency has fallen two months running, 2 points if spend has fallen two months running, 2 points for an unresolved ticket older than seven days, and 1 point for each admin flag. Anyone on 5 or more goes on the call list. Anyone on 3 or 4 gets a personal email. Everyone else gets left alone, because pestering happy customers with retention offers trains them to expect discounts.
Run it weekly. Dimension Labs makes the point that for most B2B subscription businesses a weekly refresh is the practical baseline, quick enough to catch deterioration before a renewal window closes and slow enough that scores don't bounce around. I'd agree, with one exception: if you sell something used daily, run the recency check daily and everything else weekly.
The scorecard will be wrong sometimes. That's fine. What you want after three months is a log of who scored high, whether you contacted them, and whether they stayed. That log is the training data for any future model, and it's the evidence that tells you which of your weights are rubbish.
When Machine Learning Is Actually Worth It
There's a lot of pressure to go straight to AI, and for most small businesses it's the wrong call. Machine learning models find patterns in large numbers of examples. As Fast Data Science explains in its guide to predicting churn with machine learning, the approach becomes valuable when you have very large numbers of customers, typically in a consumer context, and a business with two or three churned customers a year has far too little for any meaningful pattern to show up. Thousands of customers is where it starts to make sense.
You also need history. Most platforms want at least a year of interaction data including actual churn events before their predictions are worth trusting, plus several months of tuning after that. With eight months of clean records, you'll spend money to be told what your scorecard already told you.
When you do get there, the standard advice holds. Logistic regression, which is a statistical method that estimates the probability of a yes or no outcome from a set of inputs, is the right first model because you can explain it to a non technical colleague and because it behaves sensibly when churn is rare. Gradient boosting and random forests, which are more complex methods that combine many small decision rules, usually beat it on accuracy but are harder to interpret. Whatever you use, judge it on precision and recall rather than accuracy. Precision is the share of customers the model flagged who actually left, recall is the share of leavers the model caught. A model that never flags anyone is 95 percent accurate in a business with 5 percent churn, and completely useless.
If you'd rather not build this yourself, Fast Data Science is listed on the UK government's Digital Marketplace under the G-Cloud framework and builds churn models for clients. Expect a consultancy engagement rather than a subscription, and don't commission it until your scorecard has run long enough to give them something to learn from.
What the Tools Cost in the UK
Here's where I'll be blunt, because the tools market for churn is full of expensive products aimed at companies with a customer success department. Most small UK businesses need one of four things: subscription analytics if you sell on subscription, a CRM with a custom score field if you sell B2B, your ecommerce platform's own reports if you sell goods online, and a spreadsheet in all three cases.
For subscription analytics, ProfitWell Metrics, now owned by Paddle, remains free for core metrics including churn and cohorts, and it connects to Stripe as well as Paddle. Joe Wilkinson's comparison of ProfitWell, ChartMogul and Baremetrics describes it as the best free option for Stripe connected SaaS at an early stage, and I'd agree. ChartMogul is free up to roughly ten thousand dollars of monthly recurring revenue, with its paid Scale plan starting at around a hundred dollars a month, and it handles multiple billing systems better than the other two. Baremetrics is the polished, Stripe native option and its published entry price has moved around over the past year, with independent checks reporting between 75 and 129 dollars a month for the Launch plan, so verify on its own pricing page before you commit. All three bill in US dollars, so the pound figures are approximate: call it free, roughly £80 a month, and roughly £60 to £100 a month respectively at the time of writing. Remember to add VAT on top of the sterling equivalent when you budget.
None of those three predicts churn for you. They show historical churn and cohorts, which is exactly what you need to build and test a scorecard, but the scoring is still yours to do. Products that promise machine learning prediction at a flat monthly price exist, but I'd want their accuracy claims tested on my own data before paying.
For B2B firms, HubSpot is the CRM I see most often in UK small businesses and it publishes proper GBP pricing. Its free CRM covers two users, and the Starter tier is £18 per seat per month at list price, with a new customer promotion running at £7 per seat on annual terms when SpotDev checked the UK HubSpot prices in July 2026. Sales Hub Professional is £85 per seat per month, and Professional and Enterprise carry a mandatory one off onboarding fee, £1,310 at Professional according to Expertsure's UK pricing breakdown. All of those figures exclude 20 percent VAT. The point for churn work is that the free and Starter tiers give you custom properties, so you can store a churn score against each contact and filter on it. You only need Professional if you want automated workflows to trigger tasks off the score, and for firms with under a thousand customers a weekly manual filter is enough.
For UK ecommerce, start with what's already in the platform. Shopify's RFM customer reports are the obvious example, and similar segmentation exists in most email marketing tools. RFM stands for recency, frequency and monetary value, and it's the scorecard idea with a longer history.
The tools I'd steer a small business away from are the enterprise customer success platforms. They're built for teams of account managers, priced accordingly, and their AI features are only as good as the year of clean data they'll ask you to load. If someone quotes you more than a few hundred pounds a month and you've got fewer than a couple of thousand customers, spend the money on the person who makes the calls instead.
The UK Rules That Apply to Churn Scoring
Two sets of rules matter here and both are British, not American, so ignore anything you've read about US state privacy laws.
The first is data protection. Scoring customers by behaviour is profiling under the UK GDPR, and until this year the rules on solely automated decisions were quite tight. That changed on 5 February 2026, when section 80 of the Data (Use and Access) Act 2025 came into force and replaced Article 22 of the UK GDPR with new Articles 22A to 22D. The effect, as Bird & Bird summarise it, is that the general prohibition has been replaced by a permissive, safeguard led approach, and only solely automated significant decisions based on special category data remain restricted. Special category data means things like health, ethnicity or religious belief. For everything else you can rely on a lawful basis such as legitimate interests, provided you tell people how and when automated decisions are made, let them contest a decision, and give them a route to human intervention.
In practice, a churn score that decides who gets a phone call is nowhere near a significant decision, and it's not solely automated if a human looks at the list and chooses whom to ring. Where I'd be careful is if you start using the score to do things that affect the customer, like withholding a renewal offer, refusing a refund, or moving someone to worse terms. The ICO's draft guidance on automated decision making and profiling, published in March 2026 with the consultation closing on 29 May, is clear that human involvement has to be active rather than a token gesture. Final guidance is still to come as I write this, so keep a one page record of what data feeds your score, what it's used for, and who reviews it, and you'll be in decent shape whatever the final wording says. The ICO's advice for small organisations section is written in plain English and is where I'd start.
The second set of rules is about contacting the people you've flagged. If your retention outreach is a marketing message by email or text, the Privacy and Electronic Communications Regulations apply. You need consent, or you need to meet the soft opt in conditions for existing customers: you collected the contact details during a sale or sale negotiation, you're marketing similar products or services, and you gave a clear chance to opt out both when you collected the details and in every message since. The ICO's guide to electronic mail marketing sets this out and notes the soft opt in doesn't apply to prospective customers or bought in lists. Two useful details: these rules apply to individuals, including sole traders, rather than to corporate entities, and a genuine service message, such as telling someone their card has failed, isn't marketing at all. A "we noticed you haven't been in, is everything alright" email sits in a grey area, so I'd keep it free of offers and discounts unless you've got the marketing permission sorted.
What to Do Once You Know Who's Leaving
Prediction without a response is just a sad spreadsheet. The response should differ by why someone's on the list, which is why I keep the reason codes visible next to the score.
For failed payments, automate it. A polite email the day the card fails, a second three days later, a text on day seven, and a retry schedule. Paddle Retain, Baremetrics Recover and Stripe's own tools all do this, and it's the one place where software beats a person, because the volume is high and the message is boring.
For onboarding stalls, the intervention is help, not a discount. Ring them or send a short personal email asking what they were hoping to do and offering to walk them through it. Most people who stall in month one aren't unhappy, they're busy, and a fifteen minute call beats a 20 percent off code.
For customers whose usage or spend is declining, ask before you offer. Discounts win back a slice of customers who were leaving on price and teach everyone else to wait for the next one. A call that starts with "what's changed for you" tells you whether it's price, a competitor, a change in their business or something you broke.
For B2B accounts showing closure signals, the conversation is about terms and timing rather than product. A smaller regular order, a pause rather than a cancellation. You're keeping a relationship alive through a rough patch, and customers who survive it remember who was decent to them.
Where I've Seen This Go Wrong
The most common failure is building something clever and not acting on it. I've seen a beautifully weighted model in a business where nobody had time to make the calls, and churn didn't move a point. Decide who owns the call list before you decide how to build it.
The second is measuring the wrong thing. If you only track whether flagged customers stayed, you'll congratulate yourself on saving people who were never leaving. Track the churn rate of the flagged group against the unflagged group, and what happened to customers you didn't contact. That's the only way to know whether the scorecard finds real risk or just describes your quiet customers.
The third is over trusting an accuracy claim. Vendors quote figures like 85 or 90 percent, and accuracy is easy to inflate when churn is rare. Ask for precision and recall on a business your size, and what happened in the first six months before the model settled.
And the fourth is forgetting that some churn is healthy. A customer who was never a good fit, or who cost more to serve than they paid, is not a failure. The best small businesses I've worked with have a short list of customers they'd be quietly relieved to lose, and the scorecard leaves them alone.
A Workflow You Can Run This Week
Monday morning, pull the export: every active customer, days since last activity, three months of counts and spend, open tickets, admin flags. Apply the points, sort descending, and split into call list, email list, leave alone.
Tuesday, make the calls. Five to ten of them, no script beyond the opening question, and write two lines in the CRM about what you heard. Wednesday, send the personal emails to the middle band, service tone, no offer. Thursday, check your failed payment queue and make sure the automated sequence is actually firing.
Friday, spend ten minutes on the log. Who did you contact last month, what did they say, are they still here. After three months you'll know which signals earned their points, and after a year you'll have the history that makes any future model worth the money. That's churn prediction for small businesses in its honest form: a repeatable weekly habit that gets sharper the longer you keep it up, and a list of at risk customers that turns into conversations instead of cancellations.