AI Playbook for Accounting Teams

• Start with: invoice exception review or invoice coding recommendation
• Save as a skill: matching rules, coding conventions, approval matrix, and stop rules
• Connect: AP inbox, ERP/AP tool, PO/receipt data, vendor master, policy
• First agent: AP invoice exception coordinator
• Measure: exception cycle time, match rate, and duplicate-payment prevention

# Reconciliation Exception Triage

## Job to be done
Turn a raw list of reconciliation breaks into a focused investigation queue.

## When it runs
When a new unmatched item appears, an item ages beyond threshold, or the reconciler requests triage.

## Approved context
Reconciliation tool, GL detail, account support, prior-period notes, account-owner roster, materiality rules.

## Operating loop
Classify each break by likely cause; calculate age and amount; identify evidence missing; rank by materiality and risk; assign an accountable owner and proposed next step.

## What it delivers
An exception worklist with issue type, amount, age, evidence needed, owner, priority, and review question.

## Human owner
Account owner investigates and approves the resolution.

## What it must never do
Never clear, write off, book, or reclassify a reconciliation item. Escalate potential fraud or material breaks immediately.

## How to measure it
Aged breaks, time-to-resolution, repeat exception rate, and percentage resolved with complete evidence.

## Operating rules
1. Use only approved sources and follow the linked accounting skill.
2. Cite source records and separate facts from hypotheses.
3. Route missing evidence, threshold breaches, and material exceptions to the human owner.
4. Keep an auditable record of trigger, inputs, output, reviewer, decision, and override.
5. Stop and ask for human review when the policy, evidence, or approval authority is unclear.

# Cash Forecast Watcher

## Job to be done
Surface material cash and liquidity changes early enough for treasury to act.

## When it runs
New bank feed, balance threshold breach, forecast refresh, or material AR/AP timing change.

## Approved context
Bank balances, cash forecast, AR and AP timing, payment calendar, approved forecast assumptions, covenant thresholds.

## Operating loop
Compare current cash and expected flows to the approved forecast; identify timing shifts and threshold breaches; list the assumptions driving the movement; prepare questions for owners.

## What it delivers
A cash alert with variance amount, timing, likely drivers, liquidity impact, assumptions, and questions requiring validation.

## Human owner
Treasury validates timing, funding decisions, and any proposed action.

## What it must never do
Never move money, schedule payment, alter a forecast assumption, or communicate liquidity guidance externally.

## How to measure it
Forecast accuracy, alert lead time, number of material surprises, and time to resolve timing uncertainty.

## Operating rules
1. Use only approved sources and follow the linked accounting skill.
2. Cite source records and separate facts from hypotheses.
3. Route missing evidence, threshold breaches, and material exceptions to the human owner.
4. Keep an auditable record of trigger, inputs, output, reviewer, decision, and override.
5. Stop and ask for human review when the policy, evidence, or approval authority is unclear.

# Month-End Narrative Drafting

## Job to be done
Reduce blank-page time for management reporting while keeping finance accountable for the final message.

## When it runs
Final review package is available or the reporting calendar reaches the drafting milestone.

## Approved context
Approved financials, variance drivers, KPI dashboard, prior narrative, operating-owner notes, reporting style guide.

## Operating loop
Extract the relevant approved facts; compare with budget and prior period; organize headline, drivers, risks, and decisions; draft cautious language; list unsupported claims for review.

## What it delivers
A first draft of the management report plus a fact-check list and open questions.

## Human owner
Finance leader edits, fact-checks, and owns the final message.

## What it must never do
Never distribute the narrative, add unapproved forward-looking statements, or infer a business cause without support.

## How to measure it
Draft preparation time, edit rate, factual-correction rate, and on-time reporting.

## Operating rules
1. Use only approved sources and follow the linked accounting skill.
2. Cite source records and separate facts from hypotheses.
3. Route missing evidence, threshold breaches, and material exceptions to the human owner.
4. Keep an auditable record of trigger, inputs, output, reviewer, decision, and override.
5. Stop and ask for human review when the policy, evidence, or approval authority is unclear.

# Data-Quality Watchdog

## Job to be done
Catch source-data issues before they become reporting, forecast, or control problems.

## When it runs
Scheduled validation, new file load, data-model refresh, or failed reconciliation check.

## Approved context
Data dictionary, exports, validation rules, control totals, prior-period baseline, source-system owner list.

## Operating loop
Run approved validation checks; identify blanks, duplicates, outliers, invalid formats, and reconciliation failures; quantify impact; assign issues to source owners.

## What it delivers
A data-quality report with issue type, affected records, severity, probable source, and recommended source-level fix.

## Human owner
Data owner corrects the source and confirms remediation.

## What it must never do
Never modify production data, overwrite source records, or silently exclude bad data from a report.

## How to measure it
Issue rate, time-to-remediation, repeat-defect rate, and reporting rework avoided.

## Operating rules
1. Use only approved sources and follow the linked accounting skill.
2. Cite source records and separate facts from hypotheses.
3. Route missing evidence, threshold breaches, and material exceptions to the human owner.
4. Keep an auditable record of trigger, inputs, output, reviewer, decision, and override.
5. Stop and ask for human review when the policy, evidence, or approval authority is unclear.

• What AI does: Matches PO → Receipt → Invoice automatically
• Flags exceptions: Quantity, price, or date mismatches
• Reduces: Manual matching time by 70%+

• What AI does: Detects anomalies (unusual vendor, amount, timing)
• Limitations: Does not prevent fraud; it only alerts
• Must have: Human investigation for flagged items

• What AI does: Drafts collection emails, escalation sequences, payment reminders
• Personalization: Tailors message to payment risk & customer segment
• Control: Human review before send; tone & legal review required

• What AI does: Predicts weekly/monthly cash inflows using payment patterns
• Enables: Better working capital planning & liquidity management
• Accuracy: Improves with more historical data (6+ months)

• Prompts: variance narrative, reconciliation exception triage, close gap analysis
• Skills: balance-sheet reconciliation prep; close command center; GL anomaly review
• Connect: ERP/GL, close system, reconciliation support, task tracker
• Agent: daily close monitor; reconciliation exception triage; GL anomaly review queue

• What AI does: Drafts narrative explanations for budget vs. actual variances
• Inputs: KPI table, drivers, one-time items
• Quality: 70-80% requires human editing; catches 90% of obvious drivers

• What AI does: Auto-matches bank transactions to GL, identifies unreconciled items
• Coverage: 95%+ of standard/recurring reconciliations
• Edge cases: Manual review required for unusual/multi-month exceptions

• What AI does: Projects revenue, expense, headcount based on trends & drivers
• Accuracy: 85-95% for recurring items; worse for new/volatile categories
• Control: Always overlay with business logic & management assumptions

• What AI does: Drafts executive summary of monthly/quarterly results
• Format: Highlights key drivers, misses, opportunities
• Caution: Requires 30-50% human editing; verify data & tone

• Save the step-by-step instructions for a task your team repeats
• Add the information, examples, and review steps that make it useful
• Update it when your process or policy changes

• Set a skill to run at a regular time or when something happens
• Gather the information, prepare the work, and flag issues
• Leave approvals, judgment calls, and system changes with people

• Prepare invoice and vendor exception reviews
• Prioritize collection follow-up
• Draft internal worklists and customer-message starters

• Accounting treatment, payments, entries, and policy decisions
• Final review of anomalies, fraud flags, and supporting evidence
• Any action that changes the books or commits the company

• AI requester ≠ AI approver ≠ payment authorizer
• Close owner ≠ variance explainer reviewer
• Map roles to workflows & validate in system controls
• SAOx / audit testing must include AI-assisted processes

• Document all system prompts used for AI analysis
• Version control prompts; track changes (what changed, when, why)
• Publish approved prompts to team; prevent ad-hoc workarounds
• Archive old prompts; audit trail if disputes arise

• AI results contradict known business facts → investigate immediately
• Consistency drop in prediction accuracy → retrain or pause model
• Unexplained variance explanations → remove from draft until fixed
• Repeated same override on same rule → rules need adjustment