A per-customer model is a detection model that learns the normal behavior of one organisation rather than applying a generic pattern across all tenants. This approach improves precision because risk is judged against that customer’s own communication norms, business relationships, and historical email activity.
How a Per-Customer Model Works
A per-customer model is trained to understand one tenant’s own baseline, so the same event can look normal for one organisation and suspicious for another. That makes it especially useful when communication patterns, partner networks, or volume profiles vary sharply across customers.
Its value comes from reducing false positives that often appear when a single generic model is forced across very different environments. In practice, the model is comparing activity against the customer’s own history, not against an abstract population average.
Why It Improves Detection Precision
Precision improves because the model can learn local context that shared models often flatten away, such as recurring vendors, internal naming conventions, seasonal bursts, or normal outbound destinations. A per-customer baseline can therefore distinguish legitimate business variation from unusual behaviour more reliably.
This approach is strongest when the organisation has enough stable historical data to describe normal behaviour with confidence. It is less helpful when the tenant is new, highly volatile, or has too little data for the model to form a dependable baseline.
Common Trade-Offs and Limits
Per-customer models usually trade scale and simplicity for accuracy. They require tenant-specific training, ongoing maintenance, and careful threshold tuning, which can increase operational overhead compared with one-size-fits-all detection.
They can also be sensitive to concept drift. If the customer’s business changes quickly, the model may either miss emerging anomalies or overreact to legitimate change until the baseline is refreshed.
Where It Fits in Security Operations
Per-customer modelling is most useful in detection workflows where context matters more than uniform pattern matching, such as email abuse, unusual sender relationships, account behaviour, or tenant-specific fraud signals. It is a contextual analytics approach, not a replacement for broader detection engineering, investigation workflow, or policy controls.
Because the model is tuned to a single customer’s environment, it is often best used as one signal in a larger triage process. Analysts still need to validate whether an alert reflects abnormal behaviour, an expected business change, or a model blind spot.
Risk and Threat Considerations
Per-customer models reduce noisy alerts, but they also concentrate trust in the quality of one tenant’s historical data. If the baseline is incomplete, poisoned, or built during an abnormal period, the model can normalise risky behaviour and miss the very activity it is meant to surface.
Failure mechanism: An attacker or internal misuse pattern can hide inside the customer’s own “normal” if the model has already learned compromised relationships, atypical sending patterns, or an inflated tolerance for rare events.
Impact: Detection quality degrades in a customer-specific way, which can delay investigation, weaken fraud or abuse visibility, and allow tenant-local anomalies to persist longer than they would under a more conservative control.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST CSF 2.0, NIST SP 800-53 Rev 5 and CIS Controls v8 set the technical controls, while ISO/IEC 27001:2022 defines the regulatory obligations.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST CSF 2.0 | ID.RA-03 — Cybersecurity Risk Assessment | Per-customer baselines must be evaluated for tenant-specific detection risk and model drift. |
| Recommendation — Reassess tenant-specific detection risk whenever the customer baseline changes. | ||
| NIST SP 800-53 Rev 5 | AU-6 — Audit Record Review, Analysis, and Reporting | The model exists to improve review and analysis of tenant-specific activity signals. |
| SI-4 — System Monitoring | Per-customer modelling is a monitoring approach that detects abnormal behaviour against a local baseline. | |
| AC-2 — Account Management | Customer-specific behaviour often depends on account and sender history that must stay current. | |
| Recommendation — Use AU-6 to tune review logic around customer-specific alert patterns. Apply SI-4 to monitor tenant activity against the customer’s learned baseline. Keep account and relationship data current so the model reflects legitimate usage. | ||
| ISO/IEC 27001:2022 | A.8.16 — Monitoring activities | Per-customer models are monitoring controls that depend on ongoing observation and tuning. |
| Recommendation — Review monitoring outputs regularly so tenant-specific anomalies stay visible. | ||
| CIS Controls v8 | CIS-8 — Audit Log Management | Detection models rely on tenant activity records as the evidence base for learned normal behaviour. |
| Recommendation — Collect and retain the customer’s relevant activity logs to support baseline-driven detection. | ||
Practitioner Guidance
What to watch for: Treat baseline quality as part of model governance, not just model performance. The most useful question is whether the learned normal still reflects the tenant’s current operating reality, especially after mergers, seasonal shifts, migrations, or major process changes.
Governance implication: Per-customer models work best when there is clear ownership for retraining, threshold review, and exception handling, so that “normal” does not silently drift away from what security teams consider acceptable.
Related resources from NHI Mgmt Group
- What breaks when customer identity is forced into a shared platform model?
- What breaks when AI access to customer records is not logged per record?
- What breaks when AI model access is managed without logging, budgets, and per-team controls?
- When does a self-service customer success model create more friction than it removes?
