Skip to main content

Finance Subagents: Specialized AI for Financial Workflows

Your finance team has specialists. The FP&A analyst builds forecasts. The accountant handles reconciliations. The controller reviews journal entries. The CFO synthesizes everything into board narratives. Each person has deep expertise in their domain, and complex work flows between them.

What if your AI assistant worked the same way?

In Chapter 5, you learned about subagents: specialized AI assistants with focused expertise and isolated context. You saw how the Plan subagent researches your codebase before creating multi-step strategies. You created custom subagents for specific workflows.

Now we apply that same pattern to finance. Instead of one general AI trying to do everything, you orchestrate specialists: a Financial Modeler for forecasts, a Scenario Analyst for what-if analysis, a Risk Assessor for stress testing, a Validator for integrity checks, and a Narrative Generator for stakeholder communications.

This mirrors how finance teams actually work, but with automation and consistency.


Building on Chapter 5: The Subagent Foundation

Before diving into finance-specific applications, let's recall the core concepts from Chapter 5, Lesson 11:

What is a subagent? A specialized AI assistant with its own instructions and isolated context window. Each subagent is an expert at one type of task.

The execution model: A subagent is invoked once for a specific goal, completes its work, and returns results to the main Claude Code session. Then control returns to you.

Why isolation matters: Each subagent starts with clean context, uncluttered by previous conversations. This produces more focused, higher-quality results.

Orchestration: The main Claude Code session coordinates multiple specialists toward a complex goal, like a project manager delegating to team members.

Now we apply these patterns to finance work.


The Finance Team Analogy

Every finance department has specialists. Compare a typical finance team to its AI equivalent:

Human RoleAI SubagentPrimary Focus
FP&A AnalystFinancial ModelerBuilds forecasts, refactors models, normalizes assumptions
Strategy AnalystScenario AnalystCreates what-if scenarios, compares alternatives
Risk ManagerRisk AssessorIdentifies risks, runs stress tests, flags exceptions
ControllerValidatorChecks integrity, validates assumptions, ensures consistency
CFO / IR LeadNarrative GeneratorCreates stakeholder explanations, board summaries

When your CFO asks for a quarterly analysis, the work flows: FP&A updates the model, Strategy runs scenarios, Risk flags concerns, Controller validates the numbers, and someone synthesizes it into a board presentation.

AI subagents can mirror this workflow, each specialist contributing their expertise while the orchestrator (you, with Claude Code) coordinates the overall process.


The Five Finance Subagent Roles

Based on research into agentic financial planning, five specialized roles emerge as particularly valuable:

1. Financial Modeler

Specialty: Builds and refactors financial models, translates business intent into structured financial logic.

When to use:

  • Creating new revenue forecasts or expense projections
  • Refactoring existing spreadsheets for clarity
  • Normalizing assumptions across multiple models
  • Detecting logical inconsistencies in formulas

Example prompt to invoke:

"Use the financial-modeler to create a 12-month revenue projection based on these assumptions: 15% monthly growth rate, $50K starting MRR, 3% monthly churn."

What it returns: Structured model with formulas, assumption documentation, and dependency explanations.


2. Scenario Analyst

Specialty: Creates and compares alternative scenarios, explores what-if questions.

When to use:

  • Exploring different growth trajectories (conservative, base, aggressive)
  • Analyzing impact of strategic decisions (pricing changes, new market entry)
  • Comparing acquisition targets or investment options
  • Answering "what happens if..." questions from leadership

Example prompt to invoke:

"Use the scenario-analyst to compare three pricing strategies: current pricing, 20% increase, and freemium model. Show impact on revenue, customer count, and LTV over 24 months."

What it returns: Comparative analysis with scenarios side-by-side, key metrics highlighted, and sensitivity to assumptions.


3. Risk Assessor

Specialty: Identifies risks, runs stress tests, flags unusual items for review.

When to use:

  • Risk review before board meetings
  • Audit preparation and control testing
  • Identifying exposure to economic scenarios
  • Flagging transactions or trends that need investigation

Example prompt to invoke:

"Use the risk-assessor to stress test our cash flow under these scenarios: 30% revenue decline, 60-day payment delays from top 3 customers, loss of our largest customer."

What it returns: Risk analysis with quantified impacts, early warning indicators, and recommended mitigations.


4. Validator

Specialty: Checks integrity of data, validates assumptions, ensures consistency before finalizing outputs.

When to use:

  • Before publishing any financial report
  • After merging data from multiple sources
  • When assumptions have changed and models need verification
  • Quality assurance on AI-generated outputs

Example prompt to invoke:

"Use the validator to check this month-end package: verify all transactions posted, confirm no duplicate entries, validate that actuals match the bank reconciliation."

What it returns: Validation report with pass/fail status, exceptions identified, and specific items requiring human review.


5. Narrative Generator

Specialty: Creates stakeholder-appropriate explanations, translates numbers into stories.

When to use:

  • Board meeting preparation
  • Investor updates and earnings narratives
  • Executive summaries for leadership
  • Explaining variances in plain language

Example prompt to invoke:

"Use the narrative-generator to create a CFO summary of October results: $1.2M revenue (8% above plan), COGS at 62% (target was 58%), and we closed the Johnson contract worth $400K."

What it returns: Stakeholder-ready narrative with key messages, variance explanations, and forward-looking commentary.


Orchestration Patterns for Finance

When orchestrating multiple subagents, three patterns emerge as most useful for finance workflows:

Sequential Orchestration

Pattern: One subagent completes before the next begins. Output from one becomes input to the next.

Best for: Linear workflows where quality depends on sequence.

Example: Month-end close

Step 1: Validator → Verify all transactions posted
Step 2: Financial Modeler → Update forecast with actuals
Step 3: Scenario Analyst → Compare to plan, identify variances
Step 4: Risk Assessor → Flag unusual items
Step 5: Narrative Generator → Create CFO summary

Each step depends on the previous one completing successfully.


Parallel Orchestration

Pattern: Multiple subagents work simultaneously on independent tasks.

Best for: Time-sensitive workflows where tasks don't depend on each other.

Example: Board meeting preparation

Parallel execution:
- Financial Modeler → Update Q4 forecast
- Scenario Analyst → Refresh 2026 planning scenarios
- Risk Assessor → Update risk register

Then sequential:
- Validator → Check consistency across all three
- Narrative Generator → Create board deck narrative

The first three tasks can run in parallel because they don't depend on each other. Validation and narrative generation must wait for all three to complete.


Conditional Orchestration

Pattern: Subagent invocation depends on conditions or previous results.

Best for: Exception-based workflows where most cases are routine.

Example: Transaction review

Step 1: Validator → Check for anomalies

If anomalies found:
Step 2a: Risk Assessor → Analyze flagged items
Step 2b: Human review required

If no anomalies:
Step 2: Proceed to posting

The Risk Assessor only runs when the Validator finds something unusual.


Month-End Close: A Multi-Agent Workflow

Let's walk through a complete example: month-end close orchestrated by finance subagents.

The scenario: It's October 31st. You need to close the month, update forecasts, and prepare the CFO's summary by end of day November 1st.

Traditional Approach (Manual)

  1. Accountant verifies all October transactions posted (2-3 hours)
  2. FP&A manually updates forecast spreadsheet (1-2 hours)
  3. Controller compares actuals to plan, calculates variances (1 hour)
  4. Someone reviews for unusual items (30 minutes)
  5. CFO writes summary email (30 minutes)

Total: 5-7 hours of focused work, often with errors requiring rework.

AI-Orchestrated Approach

You initiate: "Run month-end close for October."

Orchestrator delegates:

Step 1 - Validator:

"Verify October is ready to close: all transactions posted, no duplicates, bank reconciliation complete, intercompany eliminations applied."

Returns: "October validation complete. 847 transactions processed. Two items flagged for review: $5,200 charge to Marketing with no PO, $340 reconciliation difference traced to timing (check #4521 outstanding). Recommend: approve the reconciliation timing difference, investigate the Marketing charge."

Step 2 - Financial Modeler:

"Update Q4 forecast with October actuals. October revenue: $1.2M. October COGS: $744K. Apply actuals and reforecast November-December."

Returns: "Q4 forecast updated. October actuals incorporated. November projection adjusted from $1.1M to $1.15M based on October momentum. Key assumption change: customer acquisition cost reduced from $350 to $320 based on October data."

Step 3 - Scenario Analyst:

"Compare October actuals to plan. Identify top 5 variances by dollar impact."

Returns: "October variance analysis complete. Revenue +8% ($88K favorable). COGS +4 points ($48K unfavorable). Key drivers: Johnson contract accelerated $60K revenue, AWS costs spiked due to traffic surge. Net impact: operating income $40K above plan."

Step 4 - Risk Assessor:

"Review October for control exceptions, unusual patterns, or items requiring disclosure."

Returns: "Two items flagged: (1) Marketing expense without PO as noted by Validator - requires investigation before posting, (2) AWS cost spike - legitimate but should update forecast assumptions. No disclosure items identified."

Step 5 - Narrative Generator:

"Create CFO summary for October month-end. Include: actuals vs plan, key drivers, forecast update, items requiring attention."

Returns: A complete CFO summary ready for review and distribution.

You review: All five outputs in your inbox. You approve the reconciliation timing, investigate the Marketing charge (turns out it's a valid conference registration that bypassed normal process), and send the CFO summary after minor edits.

Total time: 30 minutes of review and decision-making instead of 5-7 hours of execution.


When Multi-Agent Complexity Adds Value

Multi-agent orchestration isn't always the right answer. Use this framework to decide:

Multi-Agent Adds Value When:

CriterionExample
Workflow involves multiple distinct specialtiesMonth-end close (modeling, validation, risk, narrative)
Tasks benefit from isolated contextScenario analysis shouldn't be cluttered with reconciliation details
Quality depends on structured handoffsValidator must complete before Narrative Generator starts
You need audit trail of which agent did whatCompliance requirements for financial controls
Tasks can run in parallelBoard prep with independent workstreams

Single-Agent Simplicity Wins When:

CriterionExample
Task is straightforwardClassifying 10 bank transactions
Context from one step informs the nextIterative modeling where you refine as you go
You need conversational back-and-forthExploring a complex accounting treatment
Setup overhead exceeds time savedOne-time analysis that doesn't repeat
Team is smallSolo finance person doing everything

The Complexity Threshold Question

Ask yourself: "Would I delegate this to different people on my team, or would one person handle it start to finish?"

If you'd delegate to specialists, multi-agent orchestration mirrors that workflow.

If one person would handle it, a single Claude Code conversation is probably better.


Practical Considerations

You Don't Need to Build All Five Agents

Start with the one or two that address your biggest pain points:

  • Most time spent on manual validation? Start with a Validator subagent.
  • Board always asks "what if" questions? Start with a Scenario Analyst.
  • Struggling to explain variances? Start with a Narrative Generator.

Add more subagents as you get comfortable with the orchestration pattern.

Subagents Suggest, You Decide

Remember the governance model from Lesson 9: AI reasons and suggests, you approve and execute. This applies to subagent outputs too.

The Financial Modeler creates a forecast; you review the assumptions. The Risk Assessor flags items; you decide what requires action. The Narrative Generator drafts the CFO summary; you edit and send.

Human-in-the-loop governance applies at every step.

Context Handoffs Matter

When orchestrating multiple subagents, be explicit about what information flows between them:

  • Validator findings should inform Risk Assessor priorities
  • Financial Modeler outputs should feed Scenario Analyst comparisons
  • All quantitative outputs should reach Narrative Generator

Design your orchestration so each subagent has the context it needs.


Try With AI

Apply the finance subagent pattern to your own workflows.

Design a Variance Analysis Subagent

Design a subagent for variance analysis. Think through:

1. What should its specialty be? What expertise does it need?
2. What information does it need as input? (actuals, plan, prior periods?)
3. What format should it output? (table, narrative, both?)
4. When should it be invoked? What triggers the need for variance analysis?

Give me a complete specification I could use to create this subagent.

What you're learning: How to define a subagent's scope, inputs, and outputs. The specification you create becomes the template for building the actual subagent using the /agents workflow from Chapter 5.

Map Your Month-End to Subagents

Our month-end close involves these steps:
1. Data validation (checking all transactions posted)
2. Variance analysis (comparing actuals to plan)
3. Forecast updates (revising projections based on actuals)
4. Risk flagging (identifying items needing attention)
5. Board reporting (creating executive summary)

Map each step to a specialized subagent. Show:
- Which subagent handles each step
- What information flows between them
- Which steps can run in parallel vs. must be sequential
- Where human review is required before proceeding

What you're learning: The orchestration design process. You're translating a business workflow into a multi-agent architecture, identifying dependencies and decision points.

Evaluate Single vs. Multi-Agent

I'm debating whether to use one general Claude Code conversation or
multiple specialized subagents for my quarterly reporting process.

Context: We're a mid-size company with 3 people in finance.
Our quarterly process includes: variance analysis, forecast updates,
risk assessment, and board narrative creation.

Help me evaluate the trade-offs:
- What's the setup cost of multi-agent vs. single conversation?
- Where would isolated context help vs. hurt?
- Given our team size, is the orchestration overhead worth it?
- What's your recommendation and why?

What you're learning: The decision framework for when multi-agent complexity adds value. Not every workflow needs subagents; understanding when to use them is as important as knowing how.

Safety Note: When designing finance subagents, remember that outputs are suggestions requiring human validation. Never automate posting to ledgers or sending communications without review. The subagent pattern enhances your workflow; it doesn't replace your judgment on financial decisions.