Finance Data Agent
A pre-configured Fabric Data Agent sitting on a Finance semantic model — with the curated instructions, example queries, and evaluation harness that make it answer questions the CFO would actually accept.
The hard part of a Data Agent isn't standing one up — it's getting it to answer correctly. The difference between a demo-ready agent and a production-trusted one is 50 hours of curated instructions, example pairs, glossary entries, and an evaluation harness.
This accelerator gives you all of that, scoped to a Finance domain. Drop in your own semantic model, swap the glossary entries to your terminology, and you have a CFO-grade agent in a week instead of a quarter.
Who this is for
- Finance + data teams who want to give controllers and FP&A analysts conversational access to actuals, forecasts, and variances — without writing reports for every question.
- Consultancies who want to drop a working Data Agent into client engagements and customize from there.
- Teams already on Fabric who experimented with a Data Agent, got mediocre answers, and want to see what a tuned one looks like.
What's in the box
| File | Type | What it does |
|---|---|---|
semantic-model/finance.bim | TMDL | Finance star schema: GL accounts, cost centers, entities, scenarios, time intelligence. ~80 documented measures. |
semantic-model/glossary.md | MD | Business glossary: revenue, OpEx, EBITDA, accruals, FTE — defined in your auditor's language. |
agent/instructions.md | MD | 15,000-character system prompt: routing rules, terminology, refusal guidance, formatting standards. |
agent/examples.json | JSON | 25 curated question→DAX pairs covering YoY, variance, drill-down, period close, ratios. |
eval/eval_set.json | JSON | 50 questions with expected answers + acceptance criteria. Re-run any time you change instructions. |
eval/run_eval.ipynb | Python | Notebook that runs the eval set against your agent and produces a pass-rate scorecard. |
integration/teams_app.json | JSON | Teams app manifest: pin the agent in any channel. |
integration/copilot_studio.json | JSON | Copilot Studio custom skill definition. |
rls/dynamic_rls.dax | DAX | Entity + cost-center RLS using USERPRINCIPALNAME() + bridge table — tested across 5 user personas. |
data/seed_data.csv | CSV | Synthetic finance data so you can run the agent end-to-end before plugging in your real data. |
docs/CFO_GRADE_CHECKLIST.md | MD | The 14-point checklist we use to certify an agent for executive consumption. |
The CFO-grade checklist
An agent is "demo-grade" when it answers most questions. It's "CFO-grade" when:
- It refuses confidently when the data can't support a precise answer.
- It always cites the underlying measure and time period.
- It applies the right time-intelligence pattern (calendar vs fiscal, period-end vs running).
- It honors RLS — a regional VP doesn't see other regions.
- It uses your terminology, not Microsoft's defaults ("OpEx" not "operating expenses").
- It handles "show me" in Teams the same as "what's" — both work.
- It produces numbers that tie to the close, not approximations.
Every one of these is hard-won, embedded in the included instructions and example set.
Pricing
Frequently asked questions
What about Azure OpenAI token costs?
My finance data doesn't look like a star schema — what then?
How accurate is the agent really?
Does it work with regional finance terminology?
Can I deploy multiple agents (one per region or function)?
Refund policy?
The CFO asks "what was OpEx variance last quarter?" — and gets the answer in Teams
That's the promise. The work between zero and that moment is what this accelerator covers.