Monthly close, auto-run
Monthly P&L, balance sheet, and anomaly report delivered by the 5th of the month.
The Bookkeeping Agent ingests transactions from Plaid / Mercury / Wise feeds, auto-categorizes every transaction with a confidence score, routes anything below 95% confidence to human review, reconciles monthly, and produces P&L, balance sheet, and a narrative report. Every categorization decision is logged to an immutable event ledger.
Every external system the agent touches is a typed, observable MCP tool server. Rate-limited, retryable, and rate-cost attributed back to the originating AgentRun.
Asserted in CI. Deployments that exceed the p95 target fail to ship — this is how we keep unit economics honest as prompts and models evolve.
Every call routed through the platform wrapper with prompt caching on stable context blocks. Each decision recorded as an AgentRun with tools, output, cost, and latency.
<95% confidence categorizations, reconciliation breaks, and anomalies above a configurable threshold route to human review.
Multi-currency consolidation (roadmap). Transfer pricing. Audit defense. Inventory accounting for e-commerce (roadmap).
The Bookkeeping Agent meets the FormBridge autopilot contract — a defined outcome, its own tools, reproducible decisions, a published cost budget, a human escalation path, and cassette evals replayed in CI.