Architectural deep dive

Inline AI

Invoke focused cloud functions or reasoning agents inside the pipeline, then use their responses immediately to transform, validate, route, or act on data in motion.

Cloud functions AI agents Contextual decisions Private endpoints

The difference

AI participates in the data flow

AI is frequently implemented as a separate workflow after data has already landed. DataZen can call intelligence from within the pipeline itself. The response becomes pipeline data, so the next operation can evaluate it just like any other field.

Deterministic cloud functions

Call focused Azure Functions, AWS Lambda functions, or other HTTP endpoints for repeatable extraction, classification, enrichment, validation, and custom logic.

Agentic reasoning

Invoke an AI agent when the result depends on context, interpretation, tool use, or a decision that cannot be fully expressed as fixed rules.

Row-aware prompts and requests

Build requests from pipeline values, send the context needed for the task, and merge structured responses back into the current data set.

Architecture-controlled exposure

DataZen connects to the AI endpoint you configure. That endpoint can be public, private, or hosted within your own environment according to your security model.

Execution model

Reason, evaluate, and continue

Prepare context. Select only the fields and records the function or agent needs.
Invoke intelligence. Call a configured cloud function or use the pipeline's agent operation with explicit instructions and limits.
Extract the response. Parse text or structured output into fields that downstream operations can consume.
Act inline. Branch, validate, enrich, route, notify, or persist based on the returned result.

Choose the smallest intelligent component that fits

Use a cloud function when the behavior should be deterministic and testable. Use an agent when the task needs interpretation or reasoning. Both can occupy the same pipeline without moving orchestration into a separate product.

Design considerations

Control context, latency, and data exposure

Minimize payloads: filter columns and rows before the call to reduce latency, cost, and unnecessary disclosure.
Prefer structured output: define predictable response shapes when downstream pipeline logic depends on the result.
Set operating limits: apply timeouts and clear failure behavior so an external model cannot stall the entire flow.
Match deployment to policy: select an endpoint whose networking, retention, residency, and model controls meet your requirements.

Inline AI keeps data preparation, intelligence, and the resulting action in one observable pipeline.

Back to DataZen