AI Finance Ops Command Platform
A working build of how a finance team runs AI in live operations: agents do the heavy lifting on reconciliation and forecasting, humans hold the approval gates, and the evidence trail writes itself.
- Role
- Builder / operator
- Scope
- AI P&L, reconciliation, FP&A cadence, controlled recommendations, approval gates, evidence, observability, dbt lineage
- Data posture
- Synthetic data and generalized patterns
- Code
- Repository in hardening — available on request via contact
The console
A live cut of the control plane, on synthetic data. Pick a task from the queue, read the evidence behind the agent's work, then approve or escalate it — and watch the action land in the audit trail.
Finance ops consolePeriod 2026-06 · Close day 3
Synthetic data · interactive demo
Cash runway
26.4 mo
vs 24.1 plan
Forecast variance
-1.8%
Q3 opex, favorable
Close progress
day 3 of 5
8 of 11 recs done
Awaiting review
3 items
2 recs · 1 forecast
REC-0412 · recon-agent · confidence High
Reconcile payroll clearing account
Matched 214 of 216 transactions. Two exceptions: duplicate reversal ($4,210) and a timing break clearing on the 3rd. Proposed adjusting entry drafted.
Trigger
- Nightly close run · payroll clearing balance outside $1k tolerance
Sources
- GL extract 06-30
- Payroll register 06-30
- Bank feed (synthetic)
Assumptions
- Reversal pairs match on amount + memo
- 3-day settlement window
Checks passed
- Sum of proposed entries nets to zero
- No entry exceeds $25k auto-limit
Nothing executes without this step.
Activity — every action lands in the audit trail
- recon-agent posted REC-0412 for review
- controller approved REC-0409
- scenario-agent auto-escalated FCT-0210 (> $1M threshold)
What's inside
AI-assisted P&L and reconciliation
Agents draft matches, exceptions, and adjusting entries against the ledger. A human posts them.
FP&A cadence
Forecast refreshes and variance narratives on a weekly loop, tied to the drivers that moved.
Controlled recommendations
Every AI recommendation carries its confidence, its dollar impact, and the gate it must clear.
Approval gates
Thresholds decide what a reviewer can clear and what auto-escalates. Nothing executes below a signature.
Evidence and audit trail
Trigger, sources, assumptions, and checks captured per task — reconstructable after the fact.
Observability and lineage
Agent behavior is watched like production software, and every number traces to its source via dbt lineage.