Governed answers, not an ungoverned chat interface
Brillio undertook the transformation internally, using its own FP&A function as the reference implementation for ADAM Agentic AI in decision-support Finance. ADAM offered reusable, purpose-built agents, composable orchestration, contextual memory, explainability and human-in-the-loop policy governance, all mandatory when the payload is the company P&L, resource-level margin and board packs.
Ungoverned large-language-model interfaces were insufficient here. Finn binds to BDP Synapse presentation tables, the same governed data as the official packs, enforces Azure AD and policy-based roles with fail-closed row and column access control, and uses a metric registry and parameterized queries for signed report IDs rather than unconstrained SQL. Report owners sign off wave by wave, so the assistant cannot silently redefine a pack. Embedding Finn in BrilliAI places it inside the workflow employees already use. BDP on Azure Synapse remains the analytical system of record and NetSuite the ledger; formal reporting continues to live in Excel packs and Power BI.
One assistant across 132 signed reports
Authorized employees ask natural-language questions and receive governed answers with metric, period and scope cited, from the same platform used for formal reporting. The signed estate sits in one stakeholder catalogue: Revenue Flash (10 reports), P&L (36), Extended PP (26), Phase 2 (21) and Phase 3 (39).
Coverage runs from the Monday flash call to the board deck. The CFO office and vertical leads spot forecast misses by account; controllers and geo finance explain margin movement and watch SG&A; practice heads, delivery managers, talent acquisition and country finance recover below-threshold people margin and cut bill-days leakage; practice finance and the deal desk close CRM-to-flash-to-PP gaps and stop sold-versus-delivered leakage; board, CEO, CSO and MIP owners run governance, control travel and re-forecast the corporate P&L.
Users authenticate with Azure AD and ask in the language of their own pack. Answers export to Excel or PDF, voice input supports meeting use, and role-based access ensures two users asking the same question receive appropriately scoped results.
Fail-closed access on every resource-level answer
A query supervisor routes intent across the report families, and specialist tools serve each family through template-driven queries behind metric contracts. Access control enforcement is fail-closed, including on resource-level and compensation-sensitive reports. Query guardrails, answer audit and insight synthesis complete the pattern, with atomic agents under a single orchestrator.
The supporting estate: an enterprise web interface in BrilliAI; API services on Azure App Service; Azure OpenAI; LangGraph orchestration; session persistence; Synapse BDP for flash, allocations, people-productivity and performance walks; governed prompt and glossary storage; and policy-based access control. The intelligence layer covers pattern routing, the metric registry, parameterized queries keyed to signed report IDs, and answer validation.
Owner sign-off as the acceptance baseline
Deployment was a finance-led, phased program with report owners as stakeholders, not a data-platform replacement and not a flash-only release. Discovery inventoried the Excel and email packs in the monthly cycle, mapped personas to packs, reviewed the Synapse data foundation, and ran access control and personally identifiable information review with security, particularly on resource-level people-productivity and allocations.
FP&A owns the metric contracts, and waves were locked through formal owner sign-off. Delivery sequenced core Flash, P&L and walks first, then Extended PP, Phase 2 reconciliation and forecast, then Phase 3 governance packs, with the full estate in scope from program start and engineering waves released incrementally. Build covered domain pipelines per report family, BrilliAI integration, access administration, validated question sets per wave and an operational runbook. Validation ran by persona, with owner sign-off as the acceptance baseline.
The operating model changed in parallel: analysts moved from pack assembly to metric ownership, leaders consume self-serve reporting within access policy, and engineering cannot redefine core finance metrics without owner approval. Month-end and flash close windows were respected. Salesforce remains the system of record for commit and reconciliation, which Finn reconciles against rather than replaces.
