The Agentic CLV, Part 2: How to calculate customer lifetime value

Part 1 argued that customer lifetime value is moving from a number on a slide to a signal that systems act on every day. This part is about the number itself: how to calculate customer lifetime value, which version of it you need for which decision, where the calculation usually goes wrong, and what has to change once an agent, not an analyst, is the one using it.

Lifetime is the wrong unit

Part 1 used the textbook definition: the profit a customer brings over the whole relationship. It is a good intuition and a poor thing to calculate. Nobody can tell you how many years a customer will stay. Customers rarely announce that they have left; in most businesses they simply stop buying, and you notice months later.

What you can estimate is narrower and far more useful: how much margin a customer is likely to bring in a fixed window ahead, for example the next 12 months, and when they are likely to buy next. That is the definition we use in production, and the rest of this article builds on it.

The customer lifetime value formula, with an example

In its simplest form, the customer lifetime value of one customer is:

• Margin per order, not revenue: what is left after the cost of goods, shipping, payment fees and returns.
• Expected orders: how often the customer is likely to buy within the window.
• Cost to win and serve: the share of acquisition, service and retention spend that belongs to this customer.

Worked example: a customer whose orders leave €155 margin on average, who is expected to order 9 times in the next 12 months, and on whom you spend about €155 to win and serve. 155 × 9 − 155 = €1,240.

The predictive form: two questions per customer

"Expected orders" hides two different questions: will this customer buy at all in the window, and if so, how much? Predictive CLV answers them separately:

CLV = P(active | customer data) × E(margin | customer data, active) − cost

Same customer: there is a 90% chance they buy again in the next 12 months, and if they do, they are expected to place about 10 orders at €155 margin, so €1,550. 0.9 × 1,550 − 155 = €1,240. Section 5 shows how both terms are predicted for every customer.

Two simpler variants you will meet:

• Average order value × purchase frequency × customer lifespan. Easy to calculate, but it uses revenue instead of margin and a lifespan nobody knows.
• Margin × retention rate ÷ (1 + discount rate − retention rate). Works for subscriptions with a clean, stable retention rate; much less well where customers buy irregularly, as they do in retail and most B2B.

Should future margin be discounted? Over a 12-month window the effect is small. We do not predict beyond 365 days, so discounting changes little; for longer planning horizons, discount with your cost of capital.

Why one average CLV misleads

The formula is usually filled in with averages: average order value, average frequency, average cost. The result is one number for the whole customer base, and that number describes almost nobody.

Customer value is highly skewed. Typically the top 20% of customers bring 75 to 85% of revenue, while a long tail buys once and never returns. An average sits between the two groups and fits neither. Decisions taken on it overpay for the tail and underinvest in the customers who matter.

Three customers, same formula, their own figures (illustration):

The average of €1,338 is close to customer B by coincidence and far from A and C. Bidding €400 to win "an average customer" would be cheap for an A and a heavy loss on a C. The useful version runs the formula once per customer, with that customer's own margin, rhythm and cost.

Three levels of CLV, and where each still has a role

In practice, companies work with three levels of sophistication. None of them is wrong; they answer different questions.

Most companies need all three. Historical value keeps the prediction honest, segment formulas support planning, and the per-customer prediction drives daily decisions. Many start at level 1 or 2; a good segmentation is often the right first step before a full prediction.

How a per-customer prediction works: two models

Predictive CLV is the product of two answers per customer, each from its own machine learning model, minus the cost:

• Model 1, activity: will this customer buy at all in the next 12 months? A classification model returns a probability between 0 and 1.
• Model 2, value: if the customer does buy, how much margin will they bring in the window? A regression model, trained only on customers who were active.
• Cost: not predicted but allocated per customer, from acquisition channel, service contacts and the discounts they actually used.

How the models learn: a split in time

Pick a cut-off date in the past. Everything a customer did before it, typically the 24 months before, becomes their features; what they actually did in the 12 months after it becomes the target: active or not, and the margin they brought. Repeat this for several cut-off dates, and the models learn from many customers at many stages of their relationship.

The features start with the classic RFM trio: recency (days since the last order), frequency (number of orders and the intervals between them) and monetary value (margin per order, after returns). Then come product categories, discounts used, returns, service contacts, acquisition channel, customer master data and, where available, site and app behaviour.

The algorithm: gradient-boosted trees

Both models are gradient-boosted decision trees such as XGBoost. They cope with many mixed features, pick up non-linear patterns (three silent months mean a lot for a monthly buyer and little for a yearly one) and show which features drive each prediction, which helps marketing teams trust the values. Classical probabilistic purchase models (Pareto/NBD, BG/NBD with Gamma-Gamma) remain a useful benchmark to beat.

Checked before anyone acts on it

Before the values are used, the models are tested on a cut-off they have not seen: predict the following 12 months, then compare with what actually happened. After go-live, every customer is scored each night on today's features, and the models are retrained on a schedule, for example monthly.

The timing of the next purchase

The same setup with a shorter target window, such as the next 30 days, gives the timing: how likely a customer is to buy soon. Marketing teams value this most. It says not only who is worth an offer, but when. Contacting a customer just before they would have bought anyway gives away a discount for nothing; contacting them when their usual rhythm has broken can still win them back.

Five inputs teams get wrong

• Revenue instead of margin. A customer who buys high-revenue, low-margin products, or never buys without a discount, can look like your best customer and be one of your least profitable.
• Returns left out. In fashion and electronics, returns can wipe out much of the margin of some customer groups. Net them out per customer, not as one global rate.
• Cost as one average. Acquisition cost differs widely by channel, and service-heavy customers cost more to keep. Allocate what you can per customer and the rest per segment.
• Leaving out the customers who left. Train only on today's active customers and everyone looks loyal. The customers who stopped buying are the most informative ones.
• The wrong unit. In B2B the customer is an account, not a contact; in households one person often orders for several. Decide what a customer is before calculating anything, and keep the IDs stable across systems.

None of these is a modelling problem. They are data preparation, which is why preparing the data usually takes longer than building the model.

What changes when an agent uses CLV

A value that an analyst looks at once a quarter can be rough. A value that a system acts on every day cannot. Five things change:

• Fresh by default. All customers are recalculated every night, and events that change a value at once (a large order, a return, a complaint, a cancellation) update it during the day.
• Fast to look up. A bid or a send decision cannot wait for a model to run. Values are calculated ahead and stored, so a lookup takes milliseconds. For Google and Meta, the value goes out as a conversion value or a value column in a customer list.
• A range, not just a number. An agent should know whether €1,240 means "between €1,100 and €1,400" or "somewhere between €200 and €3,000". New customers come with wide ranges and deserve careful treatment.
• Checked against reality. Every month, predicted margin is compared with actual margin per value group. When the two drift apart, the model is retrained before decisions go wrong. A CLV that nobody checks slowly turns into fiction.
• People in charge. The team decides what the agent may do on its own (adjust bids within a budget, send a reminder) and what needs a person (large discounts, key accounts). Every action is logged.

B2B and B2C: same idea, different details

What you need to start

Less than most teams expect:

• Order history with customer ID, date and value, ideally with margin or product cost. 24 months is a good basis; 12 can work.
• Returns and cancellations.
• Marketing and acquisition cost per channel, at least per month.
• Optional: service contacts, product categories, site or app activity, CRM fields.

A modern data warehouse makes the start faster but is not a precondition.

Frequently asked questions

How do you calculate customer lifetime value?

Simple version: margin per order × expected orders in a fixed window (for example the next 12 months) − cost to win and serve, e.g. €155 × 9 − €155 = €1,240. Predictive version, per customer: probability that the customer is active in the window × expected margin if active − cost, e.g. 0.9 × €1,550 − €155 = €1,240. Both terms come from machine learning models trained on the customer's own behaviour and are refreshed daily.

Which model predicts customer lifetime value?

Two models: a classification model for the probability that a customer is active in the window, and a regression model for the margin if active. Gradient-boosted trees such as XGBoost work well for both. They learn from a split in time: features from the 24 months before a cut-off date, the target from the 12 months after it. Probabilistic models such as BG/NBD with Gamma-Gamma are a useful benchmark.

Can I calculate CLV in Excel?

Yes, for the first two levels: margin per customer from an order export, or the formula with averages per segment. A prediction per customer that is refreshed every day needs a model and a data pipeline; a spreadsheet cannot keep it current.

Which time window should CLV cover?

90, 180 or 365 days, matched to your planning cycle: 12 months for most marketing and budget decisions, 90 or 180 days in fast-moving retail and for campaign planning. We do not predict beyond 365 days: the further ahead, the less a forecast can be checked against actual results. In B2B with long buying cycles, the 365-day value is combined with the account history.

What is a good CLV?

There is no universal benchmark. Compare CLV with what you pay to win a customer: a ratio of about 3 to 1 between CLV and acquisition cost is a common rule of thumb. More useful is how the ratio differs between channels and segments.

How accurate is a CLV prediction?

For a single customer it is a forecast with a range. For groups, such as the top 20%, it is usually accurate enough to steer budgets. Measure it: compare predicted and actual margin every month.

How often should CLV be updated?

Daily if marketing acts on it; monthly is enough if it is only reported.

How does annual contract value relate to CLV in B2B?

Annual contract value (ACV) is the revenue a contract brings per year. CLV builds on it: the margin from that contract, plus whether the customer renews or expands in a fixed window ahead (for example the next 12 or 24 months), minus the cost to win and serve the customer. A higher ACV raises CLV only if the margin holds and the customer stays active in that window.

Key takeaways, and what comes in Part 3

• Calculate CLV for a fixed window ahead, not for an unknown lifetime.
• CLV = margin per order × expected orders − cost to win and serve. Margin, not revenue.
• Predictive CLV = P(active) × expected margin if active − cost, from two models such as XGBoost, trained on a split in time.
• One average describes almost no one; calculate per customer.
• Historical, segment and per-customer values each have a role.
• The timing of the next purchase is the most useful part for marketing.
• An agent needs fresh values, fast lookups, ranges, monthly checks and people in charge.

Part 3 covers the operating model: feeding customer value back into ad platforms, building retention around the timing of the next purchase, and how marketing, sales and service teams organise around a value that changes every day.