Customer segmentation with your own data: from theory to the CRM

Customer segmentation means splitting your client base into groups that behave differently and therefore deserve different treatment. That sounds obvious. What is less obvious is that most segmentations we find in mid-sized companies are never used: they live in a slide deck from eighteen months ago, with pretty names ("the champions", "the dormant ones") and no field in the CRM telling you which group an account belongs to. A segment you cannot filter a list by is not a segment. It is a metaphor.

This guide is for marketing, sales and management teams at companies that already have enough data — a CRM with history, an ERP with invoicing, a shop with orders — and want to segment in order to decide, not to illustrate. You will find which methods fit which level of data maturity, how many segments an organisation can really carry, how to activate them day to day, and the mistakes that send the whole exercise back into a forgotten PDF.

What is customer segmentation actually for?

Segmenting means grouping customers by something that predicts future behaviour, not by something that merely describes them. That line separates a useful segmentation from an inventory.

An example. "Industrial-sector clients with 20 to 50 employees" is a description. It may fill a report, but it does not tell you what to do tomorrow. "Clients who bought three times in the last six months and have not ordered in ninety days" is a segment: it has an attached action (call before they go cold) and a measurable outcome (how many order again).

Customer segmentation earns its keep when it lets you make at least one of these four decisions differently per group:

• Where sales time goes. A book of 800 accounts and a team of three forces a choice. "Whoever calls most" is not a criterion.
• What message you send. A brand-new client and an eight-year client do not need the same email.
• What price or terms you offer. Discounts, shipping, payment windows.
• How hard you fight to keep them. Not every client is worth a visit.

If a segmentation changes none of those four, it is decoration. And if it changes one, it has to reach the place where the decision happens: the CRM, the email tool, the list the rep looks at on Monday morning.

The question to ask before segmenting is not "how do I group my customers" but "what would I do differently if I knew which group each one is in".

Which segmentation method fits your data?

There is no single method. There is a ladder, and the right rung depends on the data you have clean today, not the data you wish you had.

| Method | What it uses | What you need | When it fits | |---|---|---|---| | Firmographic / demographic | Sector, size, region, tenure | A reasonably tidy customer record | A starting point; for prioritising territories or tailoring the pitch | | RFM | Recency, frequency, monetary value | 12-24 months of invoicing or order history | The best opener for B2B and e-commerce with repeat purchase | | Behavioural | What they buy, which channel, which services they use | Order lines or usage data, not just totals | When the catalogue is wide and consumption patterns differ | | By value (CLV) | Accumulated and projected margin per client | Cost to serve, not only revenue | To decide where senior people go and which accounts to defend | | By need or moment | What problem the client is solving now | Interviews, surveys, structured sales notes | To design the value proposition, not to run campaigns | | Statistical clustering | Automatic grouping across many variables | Clean data and someone who can interpret it | Once the methods above are in use and fall short |

The classic mistake is starting at the bottom of the table. Clustering is the method that shines most in a deck and ends up unused most often, because it produces groups nobody can name or explain to a salesperson. If your company has no operational segments today, start with RFM: it is arithmetic, it fits in a spreadsheet or a SQL query, and it produces groups anyone understands.

How do you build an RFM segmentation step by step?

RFM ranks customers by three numbers already sitting in your ERP:

• Recency (R): days since the last purchase. Lower is better.
• Frequency (F): number of orders or invoices in the period.
• Monetary (M): accumulated spend in the period — margin rather than revenue, if you can compute it.

The concrete procedure:

1. Set the period. Twelve months for monthly-purchase businesses; twenty-four for annual or seasonal ones. If your cycle is three years long (machinery, large projects), RFM is not your method: jump to value and behaviour. 2. Compute the three values per customer. One query over the invoice table grouped by client. If the same client appears three times under three different tax IDs, stop and fix that first — we cover it in data quality. 3. Split each variable into bands. Quintiles (1 to 5) are standard. Each customer ends up with a code such as 5-4-5. 4. Group the codes into named segments. Five or six at most. For instance: champions (high R and F), loyal with upside (high F, mid M), at risk (low R, with F and M that used to be high), new (high R, low F), low-value occasional. 5. Write the segment into the CRM as a field. This is the step everyone skips. One field, refreshed monthly by an automated process, visible on the customer record.

One detail that saves arguments: the bands should be recalculated against the full history, but the segment is refreshed periodically, not in real time. If a client jumps between groups three times a month, reps stop paying attention.

How many segments should a mid-sized company have?

Between four and seven. Below four you distinguish nothing; above seven nobody learns them.

The test is simple: ask a rep with two years in the company how many segments there are and what they do with each. If they cannot recite it, you have too many. We have seen twenty-two-group segmentations built with impeccable statistical rigour that did not change a single phone call.

There is one legitimate exception: automated operational segmentation. A recommendation engine or an email platform can handle hundreds of microsegments without trouble, because no human is memorising them. What you cannot have is hundreds of segments and a human team expected to use them. Separate the two layers explicitly:

• Management segments: few, named, memorable, for people.
• Automated rules: as many as you like, invisible, for systems.

How do you activate a segment day to day?

This is where the work either pays off or dies. A segment is activated when each group has a concrete action with an owner and a frequency. Without that, you are back to the PDF.

| Segment | Action | Owner | Frequency | What you measure | |---|---|---|---|---| | Champions | On-site visit and proposal to widen the catalogue | Assigned rep | Quarterly | % that adds product lines | | Loyal with upside | Email sequence with sector cases + call | Marketing and sales | Monthly | New orders per client contacted | | At risk | Call from a manager, not the usual rep | Sales lead | Weekly against a fresh list | % reactivated within 60 days | | New | Onboarding: guided second purchase | Customer service | 30 days after first purchase | Second-purchase rate | | Low-value occasional | Digital channel only, no sales cost | Marketing | Automated | Cost per order |

Three conditions make this table happen rather than hang on a wall:

• The segment travels to the CRM as a filterable field. If the rep has to ask someone for the list, they will stop asking.
• The list arrives on its own. A scheduled Monday report with "clients who entered the at-risk group this week" works far better than a dashboard someone has to open. Same logic as an executive dashboard: the data has to go find the person.
• Results are measured per segment, not in aggregate. If you only look at total revenue, you will never know whether the at-risk play works.

When does machine learning make sense for segmentation?

Once the rules are running and demonstrably falling short. Not before.

Clustering (k-means and relatives) finds groups across many variables at once without you naming them. It is useful when you suspect there are consumption patterns you cannot see: product combinations, crossed seasonalities, usage profiles. For it to work you need:

• Clean, sufficient data. As a rule of thumb, several hundred customers with complete history. With eighty clients, clustering invents boundaries that do not exist.
• Variables chosen on business grounds. Throwing in fifty columns because they happen to exist produces incomprehensible groups. Twenty well-chosen ones work better.
• Someone to translate the output. A cluster does not come with a name. You have to look at it, understand it and christen it — and sometimes the honest conclusion is that the group is not actionable.

A sturdier use of modelling is not segmenting but predicting within the segment: which loyal client will slip into at-risk next month, or which new account looks like a future champion. That is closer to lead scoring and usually returns more than redrawing the groups.

One warning that is not minor: if you are going to segment using personal data and use those groups to set prices or terms, check the legal basis and the data-minimisation principle beforehand, not afterwards. We cover the viable minimum in data governance.

What mistakes sink a segmentation?

The ones we see most, in order of frequency:

• Segmenting on dirty data. Duplicate clients, corporate groups split across five records, a year of invoicing that loaded badly. The result looks rigorous and is noise.
• Confusing description with segment. "SMEs on the east coast" is not a segment if nobody acts differently because of it.
• No control group. If you run a campaign on at-risk clients and 20% buy again, you know nothing: maybe that 20% would have come back anyway. Hold out a random sample.
• Freezing the segment. A client base moves. If the CRM field is not recalculated monthly, within a year you are working from a false photograph.
• Measuring only the easy thing. Email opens per segment are convenient and almost never what matters. Orders, margin and retention matter, as we argue in business KPIs.
• Starting by buying a tool. Segmentation is tested in a spreadsheet with an ERP export. If it works there, then you industrialise it.

Frequently asked questions

How much data do you need to segment customers?

For RFM, twelve to twenty-four months of invoicing history and a reliable customer identifier are enough. For behavioural segmentation you need order lines or usage data, not just totals. With fewer than a hundred clients you probably do not need a model — you need a tidy sheet and commercial judgement.

How often should segments be reviewed?

Each customer's segment value should be recalculated monthly and automatically. The definition of the segments — thresholds and names — is reviewed once or twice a year, or whenever the business model changes: a new catalogue or pricing policy can make the bands obsolete overnight.

Does customer segmentation work in B2B with few accounts?

Yes, but with a different method. With fifty or a hundred large accounts, RFM adds little; what works is segmenting by value (margin and cost to serve) and by growth potential, reviewed in a quarterly meeting. Discipline matters more than the algorithm.

Can you segment without a CRM?

You can compute it, but activating it costs far more. Without somewhere for the segment to live as a queryable field, every action requires someone to prepare lists by hand, and that gets abandoned within two months. If there is no CRM, at least keep a shared table that updates automatically.

---

If you have the history and the suspicion that you are treating very different clients the same way, the first step is not to buy anything: it is to check whether your customer data survives a serious query. Our audit looks at exactly that — what exists, how clean it is, and which decision you could already make with it — and if you would rather talk it through, let's talk for thirty minutes and we will tell you whether it is worth it or whether a well-built spreadsheet will see you through this year.