Ask Your Data
Workflow
A grounded AI workflow built on top of your semantic model so teams can ask plain-English questions, explain results faster, and trace every answer back to the source data.
If your team is still spending too much time
explaining the numbers.
Finance teams writing variance commentary from scratch every month. Operations managers pulling together cross-region summaries by hand. Commercial leaders asking an analyst the same five questions every Monday.
This is the offer where AI starts earning its keep on your existing Microsoft stack — not as a chatbot bolted on top, but as a governed workflow that your team trusts.
A production-grade workflow.
Not a proof-of-concept.
- A grounded AI workflow — Copilot Studio, Claude, or Azure OpenAI depending on your governance posture — over your existing semantic model.
- A business glossary and definition layer that makes the agent answer consistently and in your team's own language.
- A 200-question evaluation set, co-authored with your team, that we run as a regression test before every release and hand to you so you can run it forever.
- Source-backed answers. Every response from the workflow cites the specific measures and rows it used, so reviewers can trace back to proof.
- A measurable accuracy threshold. We do not ship the workflow until the eval set passes at the level we agreed up front in writing.
- Training for your data team and 2–3 power users. Documentation that actually matches what was built.
- Full handover. You own the prompts, the glossary, the eval set, the infrastructure, and the right to modify any of it without re-engaging us.
Four to six weeks.
No mystery.
Workshop to name the specific decisions the workflow is meant to improve, the people who make them, and the failure modes we need to design around. Nothing is built until this is signed off.
Semantic model review and strengthening where needed. Glossary and definition layer. First draft of the workflow with Claude Code + MCP. Initial eval set co-authored with your team.
Extended evaluation against the 200-question set. A second Decision Engineer red-teams the workflow — tries to break it, find the blind spots, surface edge cases. Every failure becomes a regression test.
Rollout to the initial user group. Training sessions. Written handover. Option to continue under a Decisions Desk retainer or leave us behind entirely.
The questions buyers actually ask.
How is this different from just turning on Copilot?
Copilot out of the box is generic. It reads whatever model it is pointed at, including every inconsistency and every ambiguous measure. An Ask Your Data Workflow includes a business glossary, definition layer, evaluation harness, and governance review — none of which come in the box. It is the difference between a demo and a production system.
What happens if the AI hallucinates?
We ship with source-row traceability on every answer, an agreed accuracy threshold, and a 200-question eval set you keep and can rerun whenever you update the model. Hallucinations are not eliminated — but they become measurable, testable, and catchable before they reach a board pack.
Which LLM do you use?
Depends on your governance posture. Microsoft-aligned clients usually land on Copilot Studio + Azure OpenAI inside their own tenant. Clients with tighter sovereignty needs sometimes prefer Claude via a private deployment. We make the choice a design decision, not a feature list.
What happens to costs once we go live?
Azure OpenAI and Copilot run on your own Azure billing. We size and document the expected monthly cost during the build, with budget caps and alerts set up before launch. Typical mid-market workflows land somewhere between $500 and $4,000/month in model spend, depending on usage.
Can we start without doing a Fabric Readiness Sprint first?
Yes if your semantic model is already in good shape. No if it isn't — building an Ask Your Data Workflow on a fragile model produces a confident-sounding system that gets important things wrong. We'll tell you honestly on the first call which path makes sense.
Show us the decision
that still takes too long.
A free 45-minute call. Bring a workflow, a reporting pain, or a trust issue — we'll tell you quickly whether this is a real fit.
Request a scoped proposal