Governed Enterprise AI Assistant Replacing BI Ticket Queues

Barton Malow's operational truth lived across a data warehouse, BI dashboards, and a data catalog - CRM opportunity pipelines, financial records, project status reporting, and market intelligence. Answering a routine question meant filing a request with a BI analyst and waiting. Off-the-shelf AI assistants could not close the gap: they had no path to the warehouse, no concept of who was permitted to see which data, and no knowledge of the domain rules that make construction numbers correct - a fiscal year that does not follow the calendar, entity-level filters, and mandatory joins between certified tables. Worse, they answered anyway.

Working inside Barton Malow's Data & Automation Engineering team, we helped build Moneypenny - a company-owned, multi-model AI platform where every answer is produced by executing a tool, never by recalling one. The architecture separates the conversation layer from the capability layer using the Model Context Protocol, so warehouse search, SQL execution, BI reporting, catalog lookups, web automation, charting, and deterministic math all run as governed tools. Models from four providers are selectable per conversation, and data access is resolved per person rather than through a shared account.

Rather than asking a model to write SQL directly, we built a two-stage pipeline. A metadata search tool first resolves the question to the correct certified tables, returning their join specifications, entity filters, and fiscal-year semantics. Only then does a second tool compose and execute the query. That ordering is what separates a plausible query from a correct one: in benchmark testing, a question about open projects for a specific company returned 21,314 records through this pipeline versus 370 from direct natural-language generation, which had selected the wrong table and dropped the company filter entirely. Arithmetic is handed to a deterministic expression parser rather than the model, removing an entire class of silent error.

Most AI-on-data projects connect through a single privileged service account and attempt to filter results afterwards. We inverted that. Every user is provisioned their own warehouse service principal with OAuth token refresh, and catalog grants are derived from their assigned data domain, with custom agents constrained to a single domain by design. Governance is therefore enforced by the warehouse itself rather than by prompt instructions - an employee without finance access does not receive a filtered answer, the query simply never runs. That made it possible to open the assistant to the whole organization instead of a vetted pilot group.

Moneypenny was built so the tool layer could outlive any single interface. Users compose custom agents with their own instructions, tool sets, and embedded reports. Scheduled prompt automations run saved questions on a timezone-aware schedule and email the results, turning ad-hoc queries into standing reports. Browser automation handles sources with no API at all. The same governed capabilities are exposed four ways - web application, Microsoft Teams, a versioned REST API, and an OAuth 2.1-secured MCP endpoint for external agents - with streaming state shared across Kubernetes pods so a response survives a restart mid-answer.