E-commerce · AI
AI for retention: how to raise LTV without hiring a data team
Right now your database holds customers who are about to leave — and you almost certainly don’t know it yet. Retention is the cheapest way to grow revenue: a 5% increase adds 25–95% to profit, and winning a new customer costs 5–25 times more than keeping an existing one (Bain, Harvard Business Review). In 2026 you don’t need a data team for it: a CRM export, a no-code model and a list of fifty names are enough to start.

What a customer who stays is worth
The maths isn’t new, but it works. According to Gorgias, repeat buyers are 21% of the customer base but bring 44% of revenue and 46% of orders. The top 1% of customers spend five times more than the rest. According to Bain, by months 31–36 of the relationship a customer spends 67% more per order than in the first six months.
The conclusion is simple: marketing’s basic job is not to find a new customer every month but to understand who is already there, what they are worth and why they haven’t left yet. That is done not with “emails to everyone” but with segmentation by value.
And the reason this became this year’s business is the market itself. According to Data Insight (“E-commerce in Russia 2026”), online orders grew by 24% in 2025, the lowest growth in eight years: the market is entering its mature phase. When new customers add less in absolute terms, the main revenue lever shifts from acquisition to retention.
Why a data team is no longer a must
A year ago there was one answer: hire a data scientist, build a pipeline, wait three to six months. Today there is a market of no-code predictive retention tools: export from the CRM to CSV, a model, and 30 minutes later a list of “who is about to leave and why”. Tenur, LowChurn, ChurnGuard, Pecan AI — they all work in a similar way: XGBoost, CatBoost or ensembles of classic algorithms trained on the company’s own history, a risk score for each customer and an explanation of which signals made the model flag them.
Sales promises of “85–95% accuracy” are worth checking on your own data: a nice number is more often marketing than statistics. But the entry threshold really has dropped. The minimum is an export of orders for 12 months. That is enough to start.
Checklist: building an LTV system
1. Collect the data — without a data warehouse.
Orders (1C, Bitrix24, your CRM, YooMoney or card acquiring), categories, support requests, email history. The minimum set for a first run: customer, order date, amount, category. No infrastructure is needed — just a clean export. One reminder: the export contains personal data, and before it goes to a third-party no-code tool, that is covered by a data processing agreement.
2. Segment by value, not by activity.
The formula is the classic one: average order value × purchase frequency × length of the relationship. For each customer, use their own numbers — and sort the base by the result.
To keep it concrete: an order of 5,000 rubles, two purchases a year, a two-year relationship — that is 20,000 rubles. Extending the relationship by one more year adds 10,000 rubles of revenue at no acquisition cost. The same customer brought in by advertising costs at least the price of one campaign — hence the 5–25× range from the lead: retention is always cheaper than acquisition; the only question is for which of your customers it works.
Three segments are enough: high value, medium and low. From there, the actions for each segment differ — which is exactly why “an email to everyone” doesn’t work.
3. RFM as the foundation.
RFM (recency, frequency, monetary value) is the classic of customer segmentation, and it is still the foundation: academic papers from 2025 show that combining RFM with machine learning predicts customer value noticeably more accurately than RFM alone. In no-code tools it works under the hood: you get a list of “who is on the edge”, and the logic inside starts from RFM features.
4. A model with an explanation.
What matters is not the model but the reason. “Customer #10482, risk 78%: three months without a purchase, the last two orders on discount, one open support request” — a marketer can work with a list like that. Without the reason it’s just a number people scroll past.
5. Personal offers by segment.
According to McKinsey, 78% of consumers are more likely to buy again from brands that personalize their communication, and personalization adds 10–15% to revenue on average (up to 25% for some companies).
In practice it looks like this:
- high value — not a discount but more: early access to a new collection, priority on remaining stock, a personal manager, “you’ve run out of what you bought in March — we’ve put it in your cart”;
- medium value — a cross-sell based on the last order and add-ons: “people who bought this usually add…”;
- at-risk segment — not an offer but a question: “what went wrong?” One honest email with a direct link to reply often keeps a customer more cheaply than any discount.
A discount for the low-value segment costs more than its LTV — more on that mistake below.
6. AI agents in communication.
Personalization at scale is a question of channel. Here AI agents do what no manager can: an individual offer by email, in a messenger or on the site — in the brand’s voice, for thousands of customers overnight. An example from our own work: a Boson Agents support bot trained on 50,000 conversations resolves 85% of requests on its own. Support is not a cost centre but a retention channel: a customer who got an answer within a minute doesn’t want to leave. An agent inside support is the cheapest personalization you already have.
7. Judge by profit, not by AUC.
A recent paper, e-Profits (arXiv, 2025), shows that if the same models are ranked by AUC/F1 and by business value — customer LTV, the cost of the intervention — the winners differ. A model that “accurately” predicts churn in a cheap segment is worth nothing in a retention campaign: saving those customers costs more than they bring. The metrics: revenue recovered, LTV in the segment with the intervention against a control group.
Five mistakes
“Let’s email everyone.” A discount for every segment eats the margin: a high-value customer would have taken it anyway, and for the cheap segment it costs more than its LTV.
An average instead of segments. One average LTV across the base tells you nothing: it mixes loyal customers, chance buyers and those about to leave. Segments are the whole point of the exercise.
The model before the data. A churn forecast is trained on history: with less than 12 months of exports, collect data first and plug in a tool once you have the material. Until then, work from a manual list (step 4 of the week).
No control group. Without 10–20% of the base left without an offer, you can measure nothing. McKinsey asks for the same principle in personalization: incrementality tests.
Personalization without tone control. Recent work on generative communication in B2B records “tone drift”: aggressive automation blurs the brand, trust drops, and LTV with it. The optimum is human-in-the-loop: the agent writes, a person checks the tone.
What to do this week
- Export 12 months of orders from your CRM or 1C to CSV.
- Move it into a spreadsheet. Five columns are enough: customer, date of last order, order total for 12 months, number of orders, segment (high/medium/low).
- Calculate a simple LTV: what the customer paid over 12 months minus the direct costs of those orders. Sort the base by the result.
- Find the 50 most valuable customers with no purchase in the last 90 days — that is the manual list for your first retention campaign, and it works without any model.
- Connect a no-code churn tool or run RFM in the spreadsheet.
- Set aside a control group of 10–20% with no offer.
The real goal
The goal is not to “cut churn by a percent” but to raise the share of revenue from your existing base. When 21% of customers bring 44% of revenue, marketing has one question: who is your top 1%, and why haven’t they left yet?
A data team is useful, but it is not a condition for starting. The threshold has dropped: the minimum is a CRM export and a spreadsheet of a few hundred rows. From there the system runs on a monthly cycle: export, update the segments, an offer for the list, a control group, a report on recovered revenue. That is one week a month, not a six-month project — and that is what to plan for.
If you delegate: what to ask an agency
Some teams decide not to build the system themselves but to hire a contractor. That is a perfectly good route — but then the first conversation should follow a checklist, not slides. An honest contractor will answer all of it:
- What data you need and how you handle personal data — under Russia’s personal data law (152-FZ): who stores it, who trains the model.
- What happens in the first month — week by week: which exports, which list, which decision you will make based on them.
- How you measure — where the control group is, what “recovered revenue” means, what report arrives at the end of the month.
- The model or the process — what leaves with the contractor and what stays with you (segments, scripts, data).
- What you can’t promise — a contractor who promises a guaranteed LTV result promises the impossible: nobody gives such guarantees, not McKinsey and not your CRM.
After questions like these, most agencies drop out on their own — and that still doesn’t mean you’ve hired anyone yet.
Boson Group is a digital agency: e-commerce, performance marketing, AI agents. We can start with exactly what is described above — a CRM export, a spreadsheet of a few hundred rows and a list of fifty names — or run the whole system for you. Write to team@boson.group.




