Customer data glossary

Real-time decisioning

By Art Kukharenko · Published · Updated

Browse learn glossary resources

Real-time decisioning selects the next action for a customer, such as an offer or message, within milliseconds to seconds of an event, using current data and rules or models.

Part of the Learn hub and the customer data glossary.

What it means in practice

Real-time decisioning chooses an action while the customer is still in the interaction: which offer to show in the app, which message to send after a missed login. A complete loop has five parts. An event arrives from an event stream. The platform looks up the current profile and features. Eligibility, consent and suppression checks run. Rules or models pick the action. The decision is delivered and logged with its reason.

"Real time" is not a standard. Set a latency budget for each use case, measured end to end, and decide what happens if it is missed: a default action, no action, or a fall-back to batch. In-session personalisation may need milliseconds, triggered emails can take seconds or minutes, and most audience building can be batch. Cost and complexity rise steeply as the budget shrinks.

Rules beat models when a decision needs a clear reason, data is sparse, or a regulator may ask why. Models earn their place where many weak signals matter. Many deployments combine both, with rules as guardrails around the model's choice.

Everything evaluated must be current at decision time. A cached consent flag from last night is a risk.

Example (retail banking): when a customer opens the mobile app, the platform checks product eligibility, open complaints, vulnerability flags and consent, then picks one banner or none. The inputs and rule versions are logged so the bank can explain the choice later. A credit decision is regulated differently from a banner and needs its own governance.

Where it matters

Frequently asked questions

How fast is real time?

There is no standard. Define the latency budget per use case, end to end, and test under load. Milliseconds, seconds and minutes are different architectures.

Rules or machine learning?

Start with rules where explanation, auditability or sparse data matter, and add models where many signals improve the result. Keep rules as guardrails.

Does it need a streaming platform?

Often, but not always. Some use cases are served by an API call at the moment of interaction; streams help when decisions react to events as they happen.

Back to the customer data glossary or the Learn hub.

Further reading

Published: · Updated:

Publication date recorded for the approved release. Editorial updates do not establish legal approval or a customer result.

Need help?

Ask the PrivyCDP Assistant