AI customer segmentation: define and check useful groups

AI customer segmentation works best when each group follows an explicit rule based on your own customer revenue and buying activity. Calculate the measures, assign customers consistently and check that the groups reconcile before asking for an explanation. The result should help a sales team decide what to investigate, with each classification traceable to the file.

The business question here is concrete: which accounts combine frequent buying with broad product coverage, and which frequent buyers have narrower coverage? The answer supports account planning. It does not establish buying motives or justify an invented customer persona.

Start with one business decision and one observation window

Use a completed quarter from a US distributor's sales export. Keep USD revenue on a net basis after discounts and credits. Use customer IDs, transaction dates, order IDs, product families and a flag that distinguishes qualifying purchases from credit-only adjustments.

Build one row per customer with three measures:

Keep the agreed customer population even where purchases are absent. A zero-purchase account with a credit adjustment needs an explicit review rule; it should not disappear or be treated as an ordinary active buyer. This example's inactive account has neither purchases nor revenue.

The transaction-to-customer calculations are covered in customer insights AI. For the broader choice of business groups, use customer segmentation from the ledger; the focus here is making AI-assisted interpretation depend on a checked assignment.

Write rules that another person can reproduce

The sales operations owner approves this ordered set of rules for the example:

1. No qualifying orders: Inactive.
2. At least four orders, at least two product families and at least $10,000 revenue: Broad repeat.
3. At least four orders, without meeting the previous rule: Repeat focused.
4. Any other account with purchases: Other active.

These thresholds are illustrative business choices, not industry benchmarks. "Repeat focused" means the account did not meet the broad-repeat rule; it does not mean the account buys only one family. Store the rule version alongside the observation period, and retain the component measures rather than exporting only the labels.

Precedence matters. An account meeting the broad-repeat conditions also meets the order-frequency condition. Test the more specific rule first so that each customer receives exactly one group.

If recency is central to your decision, use the separate B2B RFM analysis method. Adding it here would change the question and the inputs.

Assign the groups in a worksheet

Place this customer summary in A1:E7. Columns B through D are the computed inputs; column E contains the assignment.

Enter this formula in E2 and fill down:

=IF(C2=0,"Inactive",IF(AND(C2>=4,D2>=2,B2>=10000),"Broad repeat",IF(C2>=4,"Repeat focused","Other active")))

Before applying it, reject blank measures and invalid negative order or family counts. A blank is missing information, not evidence of inactivity. If your approved rules differ, change the formula and the documented definition together.

Put the four segment labels in A12:A15. Customer count in B12 is:

=COUNTIFS($E$2:$E$7,A12)

Microsoft's COUNTIFS documentation describes counting rows that meet criteria. Revenue in C12 and orders in D12 use:

=SUMIFS($B$2:$B$7,$E$2:$E$7,A12)
=SUMIFS($C$2:$C$7,$E$2:$E$7,A12)

SUMIFS adds values matching the specified criteria. Fill those formulas down, then put =C12/SUM($B$2:$B$7) in E12 for revenue share. Covirage is not affiliated with Microsoft.

Reconcile the output before explaining the groups

Revenue shares are rounded to one decimal place. Customer counts include the inactive account; revenue shares use all customer revenue. Check the following differences return zero:

=SUM(B12:B15)-COUNTA(A2:A7)
=SUM(C12:C15)-SUM(B2:B7)
=SUM(D12:D15)-SUM(C2:C7)

Also check customer IDs are unique before classifying. Matching total counts alone cannot detect one duplicated customer offset by one missing customer. Revenue must separately tie to the original period export.

The checked statement is that broad-repeat accounts account for $22,000 of $40,000, or 55.0% of revenue. Birch contributes $9,000 with five orders in one family. That makes product coverage a reasonable account-review question; it does not prove additional families would fit Birch's needs.

Change one threshold to test the boundary

Keep every input fixed and raise only the broad-repeat revenue threshold from $10,000 to $11,000. Fir moves to Repeat focused. Broad repeat becomes one account, $12,000 revenue and six orders; Repeat focused becomes two accounts, $19,000 revenue and nine orders.

Their revenue shares become 30.0% and 47.5%. Other active stays at $9,000 and 22.5%, and the company still has six accounts, $40,000 revenue and 18 orders. The group mix changed because the rule changed; customer behavior did not change.

Review borderline accounts with their actual measures visible. Behavioral segmentation gives other observable grouping choices when these boundaries do not fit the decision.

Where segmentation goes wrong

• Rules overlap. Apply explicit precedence and assign each customer once.
• The window changes. A shorter period changes frequency and revenue even when the underlying buying pattern is stable.
• Labels become claims about motives. Frequent buying is observed behavior; loyalty, dissatisfaction and intent need other evidence.
• Credits and blanks create false inactivity. Review adjustments and missing measures before grouping.
• Group movement is treated as customer change. Compare rule versions before interpreting a migration.

Compare your own groups in Covirage

Bring your sales file or customer summary with the approved rule definitions; the AI analytics overview explains the file-based approach. Covirage maps the columns, its tools compute and check segment comparisons, and the external AI model explains the results with every figure traced to its rows. Create a free Covirage account, which takes about a minute, then set up your business and bring a sample of your own data.

See sales intelligence for the supporting use case.