Run a standing finance function with agents that reconcile transactions, track runway, flag anomalies, and prepare reporting for a human to approve. Use when the money side of a business needs steady attention rather than a quarterly scramble.
Scanned 9/5/2026
Install to Claude Code
npx -y skills add Amey-Thakur/AI-SKILLS --skill agent-finance-desk --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Agent Finance Desk?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/amey-thakur-agent-finance-desk)More formats (shields.io, HTML) on the badges page.
---
name: agent-finance-desk
description: Run a standing finance function with agents that reconcile transactions, track runway, flag anomalies, and prepare reporting for a human to approve. Use when the money side of a business needs steady attention rather than a quarterly scramble.
---
# Agent finance desk
Finance fails quietly. Categorisation drifts, a subscription renews
unnoticed, invoices age, and nobody sees it until the quarter closes.
A finance desk of agents is worth having because the work is
repetitive, rule-bound, and unforgiving of gaps, which is exactly what
a scheduled agent loop does well and a busy founder does badly.
## Team
- **Bookkeeper**: ingests transactions, categorises, flags anything it
cannot classify confidently.
- **Analyst** (`data-scientist-role`, `saas-metrics`): computes
runway, burn, margin, and the movement since last period.
- **Controller**: checks the analyst's numbers against source records
and challenges anything that moved without explanation.
Shape: a pipeline on a weekly close, bookkeeper to analyst to
controller, ending in a human sign-off.
## Method
1. **Fix the chart of accounts before automating.** Agents categorise
consistently against a defined scheme and badly against a vague one,
so the scheme is the prerequisite, not the output.
2. **Classify with an explicit unknown bucket.** Anything the
bookkeeper cannot place with confidence goes to review rather than a
plausible guess. A wrong category silently distorts every number
downstream.
3. **Reconcile against the source, never the summary.** The controller
compares computed figures to bank and processor records, because a
spreadsheet that agrees with itself proves nothing.
4. **Report movement with a reason.** Every metric that changed
materially needs a named cause, and unexplained movement is the
finding, not a rounding note.
5. **Track commitments, not just spend.** Contracted renewals, annual
plans, and unbilled work determine runway more than last month's
card charges (see agent-procurement-desk).
6. **Escalate on threshold, not on schedule.** Runway below a set
number of months, an anomalous charge, or an aging receivable
interrupts immediately rather than waiting for the weekly file.
7. **Keep every figure traceable to a record.** Each number in the
report links to the transactions behind it so a human can audit any
line in seconds.
## Run it
In Claude Code, run the three roles as sequential subagents over
exported statements in a dated folder, with the controller's checks as
an explicit pass and the final report requiring human approval before
anything is filed or paid. Never grant payment credentials to an agent.
Port to CrewAI as a sequential crew, or to LangGraph with a
verification node that can send the analysis back for rework.
## Signals it works
- Unclassified transactions surface as questions rather than guesses.
- Every reported number can be traced to source records.
- Runway and anomaly alerts arrive when they happen, not at close.
## Boundaries
This desk prepares and checks; it does not file, pay, or advise. Tax
treatment, statutory filings, and audit opinions are the work of a
qualified accountant, and nothing produced here is financial or tax
advice. Agents must never hold payment credentials or move money.
Regulatory reporting has legal consequence for the human who signs it.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!