Segmentation is grouping customers or activity by shared risk traits, such as product, channel, geography, or tenure, so controls and thresholds can fit each group instead of everyone at once. Done well, it lets a team catch more fraud and block fewer good customers at the same time.
What is segmentation, in plain English?
Segmentation is the practice of cutting your population into groups that behave differently, then treating each group on its own terms. A first-time user paying with a new card in a high-risk corridor is not the same risk as a five-year customer buying their usual order, so applying one rule to both is a mistake that hurts you twice.
The traits you segment on are the ones that move risk: product type, channel such as web versus in-app, geography, customer tenure, payment method, and value band. Once the groups are defined, you can set thresholds, rules, and even models per segment, so a control that is right for risky new signups does not punish safe long-standing customers.
The reason this matters is simple. A single global threshold is a compromise that over-blocks the safe segments while under-protecting the risky ones. Segmenting removes that compromise: you can lift fraud capture where it is needed and loosen friction where it is not, at the same time.
One threshold versus segmented thresholds
What changes | One global threshold | Segmented thresholds |
Safe customers | Over-blocked by a rule tuned for the worst case | Lighter friction that fits their low risk |
Risky segments | Under-protected by a rule softened for everyone | Tighter controls where fraud concentrates |
False positives | High, because one size fits no one | Lower, thresholds match real behavior |
Tuning | Simple, one number to move | More work, each segment needs monitoring |
Data needs | Plenty of volume in one bucket | Each segment must stay big enough to measure |
Who uses segments?
Who | Their role |
Risk strategy team | Defines the segments and sets thresholds and rules for each one. |
Data science | Tests which traits actually separate risk and checks each segment has enough volume. |
Analysts and investigators | Work queues and alerts that are already sorted by segment risk. |
Product and growth | Care that low-risk segments get a smooth experience, not blanket friction. |
What it looks like in practice
In practice
A payments team runs one review threshold across all checkouts. Complaints pile up: loyal customers on repeat orders keep getting held, while a wave of fraud slips through on brand-new accounts. The single number is wrong for both groups.
They split traffic into segments by tenure and channel. Established customers on the app get a high threshold and almost never see friction. New accounts on the web, where the fraud actually sits, get a much tighter threshold plus a step-up check. Fraud capture rises on the risky segment and false positives drop on the safe one, from the same total review budget.
Why it matters to operators
Segmentation is one of the highest-leverage moves a risk team has, because it breaks the false trade-off between catching fraud and keeping good customers. Instead of dialing one threshold up or down and losing on one side no matter what, you tune each group toward its own best point.
The cost is complexity and thinner data. Every segment you create is another thing to monitor, and if you slice too finely each group lacks the volume to tune or watch reliably, so thresholds get noisy and unstable. The skill is finding groups that are meaningfully different yet still large enough to measure, rather than chasing ever-smaller buckets.
What to watch when you segment
- Over-slicing. Segments too small to produce stable numbers give you noise dressed up as precision. Merge them.
- Stale definitions. Traits that separated risk last year may not now; re-check that each segment still behaves distinctly.
- Leakage between groups. Fraudsters migrate to whichever segment has the softest controls, so watch for risk shifting after you tighten one group.
- Proxy traits. Segmenting on geography or similar attributes can drift into unfair treatment; keep the basis tied to genuine risk and reviewable.
- Monitoring gaps. More segments means more dashboards; an unmonitored segment is where surprises grow.
Quick questions
How is segmentation different from a model?
Segmentation groups the population so you can apply the right controls to each group; a model scores individual risk. They work together: you often build or tune separate models per segment so each learns from behavior that is actually similar.
What traits make good segments?
Traits that genuinely separate risk and that you can observe reliably, such as product, channel, tenure, payment method, value band, and geography. A good segment is different enough to justify its own thresholds and big enough to measure.
How many segments is too many?
When a segment no longer has enough volume to tune or monitor with confidence, you have gone too far. There is no fixed number; it depends on your traffic. Merge segments that are small or that behave the same.
Can segmentation be unfair?
It can if segments rely on protected or proxy attributes rather than real risk. Keep the basis defensible, documented, and tied to behavior, and be ready to explain why a group is treated the way it is.
Does segmentation replace one global threshold entirely?
Usually you keep a sensible default for anything that does not fit a defined segment, then layer segment-specific thresholds on top. The default is a safety net, not the main tool.
How do I know segmentation is working?
Compare fraud capture and false positive rates before and after, per segment and overall. Success looks like higher capture on risky segments and lower false positives on safe ones without raising your total review load.
Go deeper
- FFIEC BSA/AML Examination Manual ↗ — The manual US examiners use to assess BSA and AML programs.
- FTC Consumer Advice: Scams ↗ — US consumer guidance on current scams and fraud, and how to report them.