Sources with an owner
Identify authoritative records, source freshness and who can approve their use. Assess data quality before indexing.
Azure AI apps and agents
Build the agent.
Engineer the system around it.
We build Azure AI applications and agents with Microsoft Foundry, grounded in the data they may use and evaluated against the work they must do.
Solution area Cloud & AI PlatformsMicrosoft FY27 conversation Ubiquitous innovation
Build a unified, governed data and AI estate
Build a unified, governed data and AI estate
Data and context
Useful answers start
with governed context.
Design the data access path before the answer. Use Foundry IQ and Fabric where they fit the estate, with identity, retrieval boundaries and source ownership made explicit.
Access in retrieval
Enforce source access when retrieving context, alongside the application’s identity checks.
Answers with provenance
Connect the response to the sources it used and define how the system handles absent or conflicting evidence.
Reference architecture
Evaluation belongs
in the architecture.
Connect the application to its agent, approved tools and governed data. Keep the runtime and controls within the agreed customer boundary.
Application and identityUI or API · authorized user · scoped request
Agent, model and retrievalMicrosoft Foundry · Foundry IQ · approved tool boundary
Governed data and operating evidenceFabric or approved data sources · evaluation · telemetry
Ingestion, indexing and retrieval, model calls and tool execution are separate boundaries. Components and data paths are selected for the workload, availability and existing estate.
| Artifact | Purpose |
|---|---|
| Decision memo | The task, permitted data, approval boundary and acceptance conditions. |
| Evaluation record | Test-set provenance; groundedness, permission leakage, task completion, harmful output and failure handling. |
| Operational trace | Model, prompt and configuration versions; observable outcomes with sensitive content redacted before diagnostics. |
| Runbook | Ownership, escalation, recovery, release controls and the evidence needed to operate the system. |
The engagement
Frame. Baseline.
Build. Hand over.
Agree the outcome and operating boundary, then leave your team with the capability and the evidence.
Frame
Agree the workload, the decision it supports and the identity and data boundaries.
Baseline
Assess the current task, source quality and representative evaluation cases.
Build
Build the application, retrieval, approved tools, evaluation and telemetry.
Hand over
Transfer the running system, reproducible evaluation and operating runbook.
How we commit
Every promise
has a measure.
- Scope before access. Sources, identities and permitted actions are agreed before the system uses them.
- Quality with context. Evaluation records its test conditions and limitations alongside the result.
- Availability verified. Selected Microsoft components are checked in the customer’s environment.
- Evidence you can operate. Failure handling, diagnostics and a named owner accompany the handover.
Start with the work
Bring the application
that needs to act.
Tell us the task, its data boundary and what a useful result looks like.