RFM is not a complicated model. Recency, frequency, and monetary value describe how recently someone bought, how often, and how much. Teams adopt it because it is legible. Then the segments come out wrong, and the model gets blamed.
The usual failure is upstream of the score.
One person, counted once
If the same customer is an ecommerce account, a store receipt, and a CRM contact, a naive RFM pass scores three people. The store visits look infrequent. The ecommerce account looks low value. The CRM contact looks dormant. None of those descriptions is the customer. Identity resolution is the step that makes the three histories one history, and only then is the score about a person.
Duplicates the other way are just as damaging. Two people sharing an email, or a household collapsed into one id, produce a "best customer" who is really two ordinary ones. I would rather leave a match unmade than invent a person the business will market to as if they were real.
Transactions that were never identified
A large share of retail history is anonymous at the moment of sale. RFM cannot see it, because there is no person to attach it to. That is a fact about the data, not a defect in the formula. When that missing identification is the reason the segments are thin, a loyalty or other first-party identity capture can create the signal. The score still runs on the unified history afterward. The program is not a substitute for the model.
What a usable segment needs
I treat a segment as usable when someone outside the data work can act on it. A high-value group, a group that is starting to drift, and a win-back group are enough to start, provided each one can be handed to the tool the team already uses. A score with no path into that tool is a column in a table.
Behavioral cuts can sit on top of RFM once the identity is sound. They should not be asked to compensate for a broken key.
The build this depends on is the customer intelligence work: resolve the person, keep the history current, and produce audiences the team can send. If you already know the segments are off, say which systems hold the purchases and what you were hoping the score would decide. The intake or the form is enough.