architecture

Finance AI Agents

AI agents for finance: report narration, expense categorization, and forecasting assistance. Includes off-the-shelf options, custom build approaches, and training examples.

Overview

Finance roles deal with data synthesis, reporting, reconciliation, and analysis. Many tasks follow patterns that AI can accelerate while maintaining accuracy requirements.

Efficiency Potential: 4-8 hrs/week at Adoptive tier

Agent Summary

Agent Efficiency Target Complexity Priority
1-2 hrs → 20 min Low-Medium High
90% auto-categorization Medium Medium
4 hrs → 1 hr Medium Medium

Financial Report Narrator

Problem Statement

Finance produces reports full of numbers. Executives want narratives—what do the numbers mean? Translating data to insight is time-consuming and often rushed.

Efficiency Target: Report narrative from 1-2 hrs → 20 min

Training Example: Monthly Report Narrative

Generate an executive narrative for this financial report.

REPORT TYPE: {{report_type}} (Monthly P&L, Cash Flow, Budget vs. Actual)
PERIOD: {{period}}
AUDIENCE: {{audience}} (Board, Exec Team, Department Heads)

DATA:
{{financial_data_json}}

PRIOR PERIOD:
{{prior_period_data}}

BUDGET:
{{budget_data}}

KNOWN CONTEXT:
{{business_context}} (major deals, investments, seasonality)

GENERATE:

## {{report_type}} - {{period}}

### Executive Summary
[3 bullets: headline number, biggest driver, key action/attention area]

### Performance Highlights

**Revenue: ${{revenue}} ({{vs_budget}}% vs. budget, {{vs_prior}}% vs. prior)**
[What drove performance—be specific about deals, segments, trends]

**Expenses: ${{expenses}} ({{vs_budget}}% vs. budget)**
[Major variances and why; one-time vs. ongoing]

**Profitability: {{margin}}% margin ({{vs_prior}} vs. prior)**
[What's driving margin change]

### Key Variances Explained

| Line Item | Actual | Budget | Variance | Explanation |
|-----------|--------|--------|----------|-------------|
[Top 5 variances with business explanation]

### Forward Indicators
- [What this month suggests about next month/quarter]
- [Risks or opportunities emerging]

### Recommended Actions
- [If any financial actions needed]

### Questions This Raises
- [What leadership might ask; prep talking points]

Expense Categorizer

Problem Statement

Expense reports, invoices, and transactions need categorization. Anomalies (fraud, errors, policy violations) hide in volume. Manual review is tedious and inconsistent.

Efficiency Target: 90% auto-categorization; flag anomalies for human review only

Architecture Pattern

Consider ENT-FINE-TUNING for consistent categorization:

# Fine-tuned model for expense classification
# Training data format:
{
    "messages": [
        {"role": "system", "content": "Categorize this expense."},
        {"role": "user", "content": "Vendor: Delta Airlines, Amount: $847.23, Description: Flight to NYC"},
        {"role": "assistant", "content": "{\"category\": \"Travel - Air\", \"confidence\": 0.98}"}
    ]
}

Training Example: Expense Analysis

Analyze these expense transactions.

COMPANY EXPENSE POLICY:
{{policy_summary}}

EXPENSE CATEGORIES:
{{category_definitions}}

TRANSACTIONS:
{{transaction_data}}

HISTORICAL PATTERNS:
{{historical_baseline}}

FOR EACH TRANSACTION, DETERMINE:

1. CATEGORY: [best fit from defined categories]
   Confidence: [high/medium/low]

2. POLICY COMPLIANCE: [compliant / needs review / violation]
   If not compliant: [specific policy concern]

3. ANOMALY FLAG: [normal / unusual / flag for review]
   If flagged: [what's unusual]
   - Amount vs. typical for category
   - Merchant vs. typical
   - Timing/pattern concerns
   - Duplicate potential

OUTPUT:

## Expense Analysis Summary

### Auto-Categorized (High Confidence)
| Transaction | Amount | Category | Notes |
[Can be processed automatically]

### Needs Review
| Transaction | Amount | Issue | Recommended Action |
[Requires human decision]

### Anomalies Detected
| Transaction | Concern | Severity |
[Potential issues to investigate]

### Category Summary
| Category | Total | vs. Budget | vs. Prior |
[Aggregate view]

Forecasting Assistant

Problem Statement

Financial forecasting requires synthesizing historical data, market conditions, pipeline data, and business context. Models need constant updating and narrative explanation.

Efficiency Target: Forecast model updates from 4 hrs → 1 hr; better narrative

Training Example: Revenue Forecast

Generate a revenue forecast with supporting narrative.

HISTORICAL DATA:
{{historical_revenue}}

CURRENT PIPELINE:
{{pipeline_data}}

SEASONALITY PATTERNS:
{{seasonality}}

MARKET/BUSINESS CONTEXT:
{{context}}

ASSUMPTIONS TO APPLY:
{{assumptions}}

GENERATE:

## Revenue Forecast: {{period}}

### Summary
| Scenario | Forecast | vs. Plan | Confidence |
|----------|----------|----------|------------|
| Conservative | $X | -Y% | [basis] |
| Base | $X | Z% | [basis] |
| Optimistic | $X | +W% | [basis] |

### Base Case Build-Up

**Recurring Revenue: ${{recurring}}**
- Starting ARR: $X
- Expected churn: Y% (reasoning)
- Net: $Z

**New Business: ${{new_business}}**
- Pipeline: $X at Y stages
- Historical conversion: Z%
- Adjusted for: [current factors]
- Expected close: $W

**Expansion: ${{expansion}}**
- Expansion pipeline: $X
- Historical expansion rate: Y%
- Expected: $Z

### Key Assumptions
1. [Assumption] - Sensitivity: [if wrong by X, impact is Y]
2. [Assumption] - Sensitivity: [if wrong by X, impact is Y]

### Risks to Forecast
- [Downside risk] - Probability: [X] - Impact: [$Y]

### Upside Potential
- [Opportunity] - Probability: [X] - Impact: [$Y]

### Recommended Actions
- [To de-risk forecast]
- [To capture upside]