Prompting across the audit

Multi agent systems

Thus far, we've explored prompting as a "flat" process: a single instruction given to a single, powerful AI model. This approach is transformative, but it doesn't mirror the complex, multi-layered nature of an audit. The frontier, which we are actively researching and developing today, lies in Multi-Agent Systems, AI systems that think like audit teams.

From single prompts to orchestrated intelligence

Traditional prompting is like asking a brilliant junior auditor to perform every task from vouching a single invoice to assessing entity-level control weaknesses. While they may excel at granular tasks, they lack overarching strategic context. Multi-Agent Systems solve this by deploying specialized AI agents at different levels of abstraction, mirroring how real audit teams operate.

🎯 Level 3: Director Agent

Strategic oversight, risk synthesis, report generation

• Integrates findings across all audit areas
• Identifies systemic control weaknesses
• Drafts executive summaries and opinions
• Escalates critical issues

📊 Level 2: Manager Agents

Assertion-level analysis and synthesis

Revenue Manager Occurrence, Accuracy, Cutoff

Inventory Manager Existence, Valuation, Rights

Payables Manager Completeness, Obligations

Controls Manager Design, Implementation, Operating Effectiveness

⚙️ Level 1: Worker Agents

Specialized task execution

ContractReader InvoiceAnalyzer ConfirmationProcessor JournalEntryScanner VarianceCalculator ControlTester CutoffAnalyzer ReconciliationChecker RatioComputer

A practical example: Auditing accounts receivable

Let's trace how a multi-agent system would audit a company's $127M accounts receivable balance, showing the actual prompts and data flows:

1 Director Agent initiates AR audit

PROMPT: "Initiate comprehensive audit of Accounts Receivable ($127M). 
Materiality: $2.5M. Risk assessment: HIGH due to new credit policy changes."

2 AR Manager Agent activates and plans

PROMPT: "Design audit procedures for AR balance of $127M comprising:
- Trade receivables: $98M (2,847 customers)
- Related party receivables: $29M (3 entities)
Focus on Existence and Valuation assertions."

RESPONSE: Deploying 4 worker agents:
→ ConfirmationAnalyzer (sample 67 accounts, $47M coverage)
→ AgingAnalyzer (full population analysis)
→ CreditLimitChecker (accounts > $1M)
→ SubsequentCashAnalyzer (collections through Feb 2025)

3 Worker Agents execute specialized tasks

ConfirmationAnalyzer:

Processed 67 confirmations:
- 52 confirmed without exception
- 8 timing differences totaling $1.2M
- 4 disputes totaling $780K
- 3 non-responses totaling $4.1M

AgingAnalyzer:

Aging buckets analyzed:
- Current: $89M (70%)
- 31-60 days: $23M (18%)
- 61-90 days: $9M (7%)
- >90 days: $6M (5%)
RED FLAG: $3.2M from bankrupt customer in >90

4 AR Manager synthesizes findings

SYNTHESIS: AR testing reveals potential misstatement of $4.0M:
- Bankrupt customer exposure: $3.2M (100% write-off needed)
- Disputed amounts likely uncollectible: $500K
- Timing differences cleared: $0
- Non-response alternative procedures pending: $300K potential

RECOMMENDATION: Increase allowance by $3.7M minimum

5 Director Agent integrates into audit opinion

INTEGRATED FINDING: Material weakness in credit monitoring controls.
- AR adjustment required: $3.7M (exceeds performance materiality)
- Pattern detected: Similar credit issues in Revenue cycle testing
- Management letter point: Enhance credit monitoring procedures
- Impact on opinion: Potential qualification if unadjusted

Implementation roadmap

Moving from traditional prompting to multi-agent systems requires a phased approach:

Phase 1: Foundation (Months 1-3)

Map current audit procedures to potential agent responsibilities
Identify high-volume, rule-based tasks for Worker Agents
Develop prompt templates for common procedures
Test with non-sensitive data

Phase 2: Pilot (Months 4-6)

Build first Worker Agents for specific tasks (confirmations, ratios)
Create simple Manager Agent for one audit area
Implement feedback loops and quality checks
Measure efficiency gains and accuracy

Phase 3: Scale (Months 7-12)

Expand to multiple audit areas
Develop Director Agent for synthesis
Integrate with audit software platforms
Train audit teams on orchestration skills

Technical architecture considerations 🏗️

Building multi-agent systems requires careful technical planning:

Model Selection: Use smaller, specialized models for Workers (faster, cheaper) and larger models for Directors (better reasoning)
Memory Management: Implement vector databases to maintain context across agent interactions
Error Handling: Build redundancy—if one Worker fails, Manager can reassign or escalate
Audit Trail: Log every agent decision, prompt, and output for review
Security: Ensure end-to-end encryption and role-based access controls

Real-world applications in production

Here are a few fictional examples of how multi-agent systems are transforming audit workflows. The path to get there requires a lot of raw engineering work but I hope it gives us all inspiration for where we can go.

🏦 Financial Services Audit

A firm deployed a multi-agent system for loan portfolio testing:

Worker Agents reviewed 10,000+ loan documents in 3 hours
Manager Agent identified 127 documentation exceptions
Director Agent correlated exceptions with credit risk ratings
Result: 75% reduction in testing time, 2x more exceptions found

🏭 Manufacturing Inventory Audit

Internal audit team used multi-agent system for perpetual inventory testing:

Worker Agents processed daily cycle counts for 12 months
Manager Agent identified patterns in count variances
Director Agent linked variances to specific warehouse locations
Result: Identified $2.3M in systematic inventory shrinkage

The evolving role of the auditor 🎭

In the multi-agent system paradigm, auditors transition from task executors to system architects and orchestrators. Key skills for the future include:

System Design: Structuring agent hierarchies for specific engagement needs
Prompt Engineering: Crafting precise instructions for each agent level
Quality Assurance: Validating agent outputs and identifying failure patterns
Strategic Judgment: Focusing on complex issues that require human insight
Client Communication: Translating AI findings into business insights

The AI handles complexity and volume; the auditor provides wisdom and judgment. This isn't replacement—it's elevation.

Learn