GOVERNED AI
Put one valuable Microsoft agent workflow into production with evidence Risk can review.
Microsoft provides the agent engine. Onyx implements the workflow, grounding, tools, identity, evaluation and operating evidence required to decide whether it is safe and valuable enough to use.
- Scope
- One approved agent workflow.
- Engine
- Appropriate first-party Microsoft agent capability.
- Required participants
- Business owner, AI/Data and Risk/Security.
- Duration
- Agreed after workflow and risk qualification.
- Commercial status
- Quote after workflow and risk qualification.
- Transaction route
- Professional services; route confirmed for the specific engagement.
The problem
An agent workflow has value, but it cannot yet pass risk review. Implement one Microsoft agent workflow with grounding, least-privilege access, evaluation, trace evidence and measurable acceptance.
Who this offer fits
This offer fits when
- one business workflow and owner are defined
- source data and access routes are available
- the customer wants a first-party Microsoft agent path
- grounding, evaluation, traceability or security blocks production
- acceptance can be measured
It does not fit when
- the request is generic “AI strategy” without a workflow
- no ground truth or evaluation route exists
- uncontrolled autonomy is required
- the real need is conventional analytics, document processing or custom application work without an agent job
What you receive
- workflow and decision boundary
- identity and least-privilege access design
- grounding and tool design
- prompt/instruction and policy controls
- evaluation dataset and acceptance criteria
- trace and review evidence
- implementation and production plan
- measured acceptance report
- handover and operating recommendations
The evidence ladder
-
01
Illustrative demonstration
Synthetic or demonstration values are visibly labelled.
-
02
Technical evaluation
The agent is measured against a defined dataset and criteria.
-
03
Business acceptance
The process owner confirms whether the workflow result is usable.
-
04
Risk acceptance
Security/Risk reviews the implemented controls and residual risk.
-
05
Production operation
The customer accepts the operating model, monitoring and ownership.
No layer is silently upgraded into “production ready” or “approved.”
Relationship to FabOps
FabOps may supply relevant platform-accountability or evidence controls where the engagement requires them. It is not a mandatory dependency and is not included unless explicitly scoped.
Frequently asked questions
Does Onyx provide its own proprietary agent platform?
The current offer is built around appropriate first-party Microsoft agent capabilities. Onyx supplies the governed implementation and evidence layer.
Do you guarantee accuracy or regulatory approval?
No. The engagement defines and tests acceptance criteria, exposes residual risk and produces reviewable evidence. The customer retains decision authority.
Can the agent take actions?
Actions are scoped to the approved workflow, permissions and guardrails. Unbounded autonomy is outside the standard offer.