A certified financial data foundation, built in three weeks.
A venture-backed healthcare company landed its GL in Databricks and gave Claude a semantic layer: one agreed definition for each of more than 100 financial metrics. It proved those definitions with an evaluation suite, a bank of test questions with known answers, and opened a Claude environment where any team can build on the same numbers.
At a glance
| Industry | ERP | Data warehouse | AI layer | Foundation build | Delivered |
|---|---|---|---|---|---|
| Healthcare, venture-backed, $100M+ annual revenue | Workday | Databricks | Claude Enterprise | 3 weeks | Semantic layer, evaluation suite, company-wide Claude environment |
The challenge: Teams building fast. No shared definitions.
The finance team spent 8 to 12 hours each month assembling the management P&L. GL data came out of Workday, ran through a third-party mapping tool, and was finished by hand in Excel. The rules for which accounts roll into which management P&L line lived in that tool and a planning spreadsheet, where no other system could read them.
The company had rolled out Claude, and the finance team had started using it to build Skills and workflows. Each one needed the numbers behind it, and with no shared source, each builder defined them independently. Two Skills could disagree about contribution margin and nobody would know, and the risk grew with each new person building on Claude.
AMW was brought in to build the foundation underneath: one place the numbers come from, and one definition per metric for every Skill, dashboard and agent to use.
What we built: Connect. Define. Certify. Build on it.
01Connect: Workday data in Databricks, refreshed monthly.
The team chose a data warehouse over a direct ERP connection. Databricks gave them more control over the data model, and room to bring in data beyond GAAP financials later without going back to the ERP. Each month, three Workday exports land in Databricks through a documented load: GL transaction detail, the trial balance and the income statement. Employee names and identifiers attached to journal lines are stripped during the load, so they never reach Databricks or Claude.
02Define: 100+ governed metrics, owned by accounting and FP&A.
The semantic layer is the single, agreed set of metric definitions that sits between the raw data and everyone who uses it. Here it is a set of governed metric views in Databricks, with more than 100 definitions in two layers. The GAAP layer lets anyone trace a GAAP P&L number end to end, from the reported line down to the journal entries behind it. The management P&L layer routes GL transactions to management P&L lines using the mapping rules that used to live in the third-party tool. Accounting owns the GAAP definitions and FP&A owns the management definitions. Both teams reviewed the mapping logic and signed off on the outputs before certification. Claude's instructions say when to pull a metric and when to ask which one the user means. Asked about revenue, for example, it checks whether the user means GAAP or management revenue before it queries.
03Certify: An evaluation suite that proves the definitions.
An evaluation suite works the way an auditor tests a control: a fixed set of questions, an expected answer for each, and every exception traced and fixed. AMW built one with more than 100 questions. It covers GAAP metrics, management P&L metrics, and guardrails: questions where the right response is to ask for clarification rather than answer. The first full run scored 94%. Each miss was traced to a gap in the semantic layer or in the SQL Claude generated, fixed at the source, and rerun until the suite passed at 100%. For comparison, Claude querying the raw GL, trial balance and P&L tables with no semantic layer scored 32%. It handled most GAAP questions and missed most management P&L questions, because those definitions exist nowhere outside the governed layer. The gap is about company-specific definitions, not Claude's ability. Without a place to look them up, any model would miss them. The suite is built to be rerun after each month-end close and each new Claude model, so the definitions stay certified as the business changes.
04Build on it: A Claude environment any team can build on.
On top of the foundation, AMW configured a Claude environment that anyone at the company can use for financial data. It reads directly from the semantic layer through the Databricks SQL MCP connector, the link that lets Claude query Databricks, so every number comes from the governed definitions rather than from a prompt or a spreadsheet. Each person signs in to that connector as themselves, so Claude can only pull the data they are already allowed to see in Databricks. Access is managed in one place, and it carries through to whatever they build. Two Skills sit on top. A Skill builder guides people building new Skills, workflows and agents, and makes sure whatever they build pulls its numbers from the semantic layer. An executive briefing Skill helps people run analysis and turn it into interactive dashboards for executive decisions. This is the part that scales. A salesperson building a territory dashboard and an operations lead running a cost analysis work from the same certified numbers finance uses. What each team builds adds to the rest, and the work compounds instead of drifting apart.
Results: What changed
| Result | What it means |
|---|---|
| 8 to 12 hrs / month | Manual assembly of the management P&L removed. It now comes straight from the semantic layer, using the same rules the mapping tool applied. |
| $30K / yr | Third-party mapping tool no longer needed. The semantic layer now performs the mapping it did. |
| 32% → 100% | Evaluation suite pass rate across 100+ questions: raw Claude on the same tables versus Claude on the governed semantic layer. |
Definitions are changed in plain language through Claude Code and the Databricks SQL connector, following a documented runbook, and the evaluation suite is rerun to recertify them.
Build the foundation your teams build on.
Start with a 30-minute intro call. We'll look at where your key numbers live today and what it would take to certify them.