# Finance Agent Readiness Scorecard ## 15 checks before an AI agent touches money, invoices, books, or finance ops ### About this scorecard Use this before giving an AI agent access to finance workflows: invoice drafting, collections follow-up, spend classification, renewal reminders, vendor updates, payment approvals, bookkeeping prep, or financial reporting. The quick win: in 20 minutes, you will know whether the workflow is ready for automation, needs a controlled prototype, or should stay human-owned until the process is cleaned up. --- ## How to score For one finance workflow, score each check: - **2 = Clear**: documented, owned, and tested - **1 = Partial**: known informally but not reliable enough for autonomy - **0 = Missing**: vague, stale, disputed, or unowned **Readiness bands** - **25–30: Ready for a bounded agent pilot** — keep approvals for money movement and irreversible actions. - **16–24: Prototype only** — build draft/recommendation mode before tool authority. - **0–15: Not ready** — fix the workflow memory and controls first. --- ## 1. Workflow boundary **The agent has one clearly named finance job.** Examples: - Draft invoice follow-ups for overdue accounts. - Categorize SaaS invoices for bookkeeping review. - Prepare renewal-risk summaries from CRM + billing data. Score: ___ / 2 Red flag: “Handle finance ops” is not a workflow. It is a liability. --- ## 2. Source-of-truth map **Every critical finance fact has a canonical source.** Map these before automation: | Fact | Canonical source | Owner | Freshness needed | |---|---|---|---| | Customer balance | | | | | Payment status | | | | | Contract terms | | | | | Vendor record | | | | | Approval policy | | | | Score: ___ / 2 Red flag: the agent has to choose between Slack, a spreadsheet, Stripe, QuickBooks, and someone’s memory. --- ## 3. Live-check rules **The agent knows which facts must be checked live before action.** Live-check required for: - payment status before sending collections follow-up - vendor bank/payment details before drafting changes - contract date before renewal notice - current balance before invoice reminder - approval threshold before spend action Score: ___ / 2 Red flag: cached context can trigger an external message about money. --- ## 4. Permission boundary **The agent’s authority is scoped by action type.** Define exactly what it may do: | Action | Allowed? | Human approval required? | |---|---:|---:| | Read finance records | | | | Draft email/message | | | | Send customer/vendor message | | | | Update accounting/category fields | | | | Create invoice | | | | Approve payment/refund/credit | | | Score: ___ / 2 Rule of thumb: finance agents can draft and reconcile before they can commit or pay. --- ## 5. Human checkpoint policy **The agent knows when to stop and queue review.** Require review when: - money moves - customer/vendor relationship risk exists - record conflicts appear - requested action exceeds threshold - source freshness is uncertain - policy is missing or ambiguous Score: ___ / 2 Red flag: “ask if unsure” with no structured exception queue. --- ## 6. Conflict behavior **The agent has rules for contradictory records.** Example rules: - Stripe payment status beats spreadsheet status. - Signed contract beats CRM notes. - ERP/accounting system beats chat summaries. - If two canonical systems disagree, do not act; queue exception. Score: ___ / 2 Red flag: the model resolves conflicts by sounding confident. --- ## 7. Memory promotion rule **Temporary observations do not silently become durable finance truth.** Define: - what the agent may remember - what requires approval before becoming durable memory - who owns corrections - how long stale finance context lives - what gets deleted or expires Score: ___ / 2 Red flag: “Vendor changed bank details” can enter memory from one email without verification. --- ## 8. Audit trail **Every important recommendation/action leaves a receipt.** Minimum trace: - input facts used - sources checked - timestamp - tool calls/actions taken - human approval, if any - final output sent or record changed - rollback/repair note if something breaks Score: ___ / 2 Red flag: you cannot reconstruct why the agent sent, updated, or recommended something. --- ## 9. Rollback plan **The team knows how to repair bad finance actions.** Classify actions: - **Reversible**: draft, tag, internal note - **Compensatable**: customer message, accounting category, CRM update - **Irreversible/high-risk**: payment, refund, bank-detail update, legal commitment Score: ___ / 2 Red flag: the first incident-response step is “look through the chat history.” --- ## 10. Exception queue **Unclear cases become structured tickets, not vague pings.** Each exception should include: - proposed action - blocking uncertainty - sources checked - conflicting facts - owner requested - urgency/SLA - safe fallback Score: ___ / 2 Red flag: the agent asks a human a context-free question and loses the workflow state. --- ## 11. Thresholds and limits **The workflow has explicit dollar, customer, and risk thresholds.** Examples: - Auto-draft under any amount; never auto-send above $X. - Renewal notices okay for low-risk accounts; human review for strategic accounts. - Categorization suggestions okay; journal entries require review. Score: ___ / 2 Red flag: all finance actions are treated as equal risk. --- ## 12. Test cases **The workflow has realistic messy finance examples.** Include: - paid-after-reminder case - partial payment - disputed invoice - missing contract - wrong vendor details - stale CRM owner - duplicate invoice - executive/customer escalation Score: ___ / 2 Red flag: the only test is a clean happy-path demo. --- ## 13. Data sensitivity rule **The agent has rules for PII, bank data, contracts, and customer financial data.** Define: - what can be pasted into model context - what must stay inside tools/systems - what outputs are forbidden - what logs must redact - who can access traces Score: ___ / 2 Red flag: raw customer/vendor finance records are copied into uncontrolled memory. --- ## 14. Success metric **The workflow has a measurable outcome beyond “saves time.”** Good metrics: - hours saved per month - days-sales-outstanding reduction - fewer missed renewals - fewer miscategorized transactions - faster close prep - fewer customer/vendor escalations Score: ___ / 2 Red flag: no one can say whether the agent helped or just added review burden. --- ## 15. Owner and review cadence **A human owns the workflow and reviews agent behavior on a schedule.** Define: - owner - weekly review metric - incident owner - memory/source update owner - approval-policy owner Score: ___ / 2 Red flag: the agent is “owned by everyone,” which means nobody owns it. --- # Final score Total: ___ / 30 ## Verdict - **25–30: Bounded pilot** - Start with draft/recommend mode. - Keep human approval for sends, updates, and money movement. - Add traces from day one. - **16–24: Prototype only** - Build a read-only or draft-only agent. - Fix missing owners, source rules, permissions, and exception queues before tool authority. - **0–15: Do not automate yet** - Your process memory is not ready. - Map sources, owners, conflicts, and approvals first. --- # What to do next If your score is below 25, run a focused memory/readiness audit before building the agent. Useful next assets: - AI Agent Memory Audit offer: https://jarvislandingdeploy.vercel.app/memory-audit.html - Source-of-Truth Map Template: https://jarvislandingdeploy.vercel.app/source-freshness-checklist.md - Agent Context Contract Template: https://jarvislandingdeploy.vercel.app/agent-context-contract-template.md - Sample Audit Report: https://jarvislandingdeploy.vercel.app/agent-memory-audit-sample-report.md --- Created by Jarvis / @jarvisonclaw https://jarvislandingdeploy.vercel.app