A five-gate hallucination gateway
Answers are evaluated through an ordered guard chain covering freshness, ontology, validation, grounding and feedback. The control plane is the minimum viable deliverable, not a retrofit.
Microsoft AI agents in production
AI agents designed to withstand risk review.
We define the business outcome first, then add grounding, access controls, evaluation and trust evidence for governed production.
Grounded measured reviewable
Ask Lisa governed agent
“What was net revenue in Q3 for the maritime division?”
£12.4M
verified against source
Product demo
Watch the governed agent answer, live.
A real question, a governed query tool, and an answer grounded in your Microsoft Fabric data, with a citation to the exact table behind every figure.
- Real question → real, grounded answer
- Cited to the source warehouse table
- A separate test set and measured result for each implementation
The Microsoft AI Agent Accelerator puts an agent into one valuable workflow on Microsoft Fabric, with measured performance and evidence the risk team can review. We define the business outcome first, then add grounding, access controls, evaluation and trust evidence for governed production.
The agent gap
The engine is Microsoft's. The trust is the hard part.
Microsoft Fabric now ships first-party conversational agents over your data, the Fabric Data Agent and the Real-Time Intelligence Agent. The model is the easy part. An agent that invents a number, reaches data it should never see, or can't show how it got an answer will not survive a risk review, and never reaches production.
The governed layer
What we deliver around the agent
Answers that show their working
Every figure carries a verification flag, the literal SQL or KQL it ran, and the row count. Ambiguous questions get a clarify-first response; out-of-scope ones get a clean refusal.
Accuracy measured, not asserted
A scored grounding test set built for your data, run through a three-instrument harness, deterministic checks, an independent semantic judge, and numeric ground-truth verification, so the agent's accuracy on your estate is a number, not a claim.
Scoped by construction
Schema scoping at the serving layer plus row-level-security automation, the agent cannot ask for what it cannot see. PII and commercial columns are excluded before the agent is ever pointed at the data.
A trust report your risk function can review and use in its approval process
Accuracy results, grounding coverage, refusal behaviour, residency considerations, and the honest bounds of a probabilistic system, the document that turns a demo into a deployment.
Your tenant, your capacity
Delivered governance-first by Onyx Data. The deployment architecture, residency and data boundary are documented for risk review against the applicable customer configuration.
The reference implementation passed all 50 questions in its defined grounding benchmark. The benchmark applies to that reference implementation. Each customer receives its own test set; results depend on the customer’s data and workflow, and approval remains the customer risk function’s decision.
Where it lives
Bound to where your people already work
- Microsoft Copilot StudioValidated for Microsoft Teams, with user identity carried through.
- Microsoft Foundry agentsThe path your business users reach, with least-privilege and trace evidence.
- Model Context Protocol endpointThe autonomous surface for your own AI systems, identity-aware by design.
FAQ
Questions buyers ask
What is the Microsoft AI Agent Accelerator?
It is Onyx Data's governed deployment of AI agents on Microsoft Fabric. Microsoft's first-party agents provide the engine; we deliver the governed layer that gets them into production, with grounded answers traceable to source, measured accuracy, access controls and evidence your risk team can review.
Does it use Microsoft's agents, or build a new one?
It uses Microsoft's first-party agents, the Fabric Data Agent and the Real-Time Intelligence Agent. We do not replace the engine; we add the governance, grounding and evidence around it that a risk function requires before go-live.
How do you stop the agent from hallucinating?
Answers are evaluated through a five-gate gateway (freshness, ontology, validation, grounding and feedback) and can carry their working, the query and row count, with a verification flag. Ambiguous questions get a clarify-first response and out-of-scope ones a clean refusal. Accuracy on your data is measured through a customer-specific verification harness, not assumed.
Where can people use the agent?
Bound to where your people already work: Microsoft Copilot Studio (validated for Microsoft Teams), Microsoft Foundry agents, and a Model Context Protocol endpoint for your own AI systems, each with identity passthrough and least-privilege guidance.
Get a governed agent into production.
Book a 30-minute strategy call. We'll map the route from a Fabric agent demo to a governed deployment designed to withstand risk review.
Book a Strategy Call30 mins no pitch leave with a plan