Integration is not a checklist. It's the connective tissue between your systems and your AI agents. Get it wrong and agents create manual workarounds. Get it right and you get 95%+ automation on day 1.

Phase 1: System Audit (Week 1)

What to document:

ERP System: SAP, NetSuite, Oracle, etc. API available? Real-time or batch? Custom fields? GL structure?
Banking: Which banks? API support? SWIFT/ISO 20022? File formats? Reconciliation timing?
Current Data Flows: How do transactions flow today? Manual? File-based? API? Reconciliation method?
Custom Fields: Any custom GL accounts? Custom cost centers? Custom dimensions in ERP?

Create a one-page data flow diagram. Who touches what. When. In what format. This becomes your integration specification.

Access protocols changed the integration work

Connecting a model to finance systems used to mean a custom build per source. That part is getting standardised.

The announcement of MCP, Model Context Protocol, has a means for these LLMs, these AIs to access and understand large data sets across multiple systems.

Craig Sullivan, on the AgentCFO podcast

During the audit, note which of your systems already expose a documented API and which will need a file drop instead.

Phase 2: API Connectivity (Week 2)

ERP API Setup

Step 1: Enable ERP API access. Create service account with read-only permissions.

Step 2: Test connectivity. Query 100 GL accounts. Query 10 vendors. Verify data completeness.

Step 3: Document required fields. Every transaction must have: Date, Amount, GL Account, Cost Center, Document Type, Description.

Common Issue: Custom fields missing from API responses. Solution: Add custom fields to API view before going live.

Banking API Setup

Step 1: Get banking API credentials from each bank. Test connectivity in sandbox environment.

Step 2: Verify statement delivery. Daily? Real-time? Format: SWIFT MT940, ISO 20022, CSV, or proprietary?

Step 3: Validate reconciliation fields. Account number, amount, date, description must match GL posting.

Common Issue: Bank descriptions don't match GL vendor names. Solution: Create lookup table for bank description to vendor mapping.

Phase 3: GL Mapping (Week 3)

Chart of Accounts Mapping

Create a mapping document: Source Account Code → Target GL Account → Account Type (Asset, Liability, Revenue, Expense).

Example: Bank CSV "2000" → SAP GL 1010100 (Cash) → Asset
Example: Invoice "COGS" → SAP GL 6010000 (Cost of Goods Sold) → Expense

Vendor Master Standardization: All vendors must have: ID, Name, Address, Tax ID, Email. No duplicates. No nulls.

Customer Master Standardization: All customers must have: ID, Name, Address, Tax ID. No duplicates.

Phase 4: Data Quality (Week 4)

This is the hardest week. But it saves 4 months of problems downstream.

GL Account Cleanup

  • Remove inactive accounts from chart of accounts
  • Fix GL accounts with wrong balance type (Asset coded as Expense)
  • Consolidate duplicate accounts (e.g., two "Office Supplies" accounts)
  • Add missing GL accounts for new cost centers

Vendor Master Cleanup

  • Merge duplicate vendors (e.g., "Acme Inc" and "Acme Inc.")
  • Remove inactive vendors (not used in past 12 months)
  • Add missing phone/email for vendors that need approval
  • Verify tax IDs are correct (matches actual vendor documentation)

Bank Master Setup

Create a mapping of every bank to every GL cash account. Map bank description patterns to vendors/customers.

Example: If bank says "ACH FROM ACME INC" → Match to vendor "Acme Inc" → Post to GL 2030000 (Accounts Payable)

Phase 5: Pilot Integration (Week 5)

Test 100 Real Transactions

Invoice Processing: Take 50 actual invoices. Run through AI agent. Verify: GL posting correct, amount correct, vendor matched correctly, cost center correct.

Bank Reconciliation: Take 50 actual bank transactions. Run through reconciliation agent. Verify: Match found for 48/50, unmatched 2 are legitimate outliers.

Adjustment Entries: Create 10 manual adjustments in AI system. Verify they post to correct GL accounts and have audit trail.

Document Exceptions

The 2 unmatched bank transactions? Don't ignore them. Understand them. Define a rule for them. This becomes part of your governance policy.

Example: "Wire transfers under $500 with no matching GL posting require manual review. Escalate to controller."

Phase 6: Scale & Go-Live (Weeks 6-8)

Production Rollout

Week 6: 50% of daily volume through AI agent. Monitor every transaction. Look for patterns in failures.

Week 7: 75% of daily volume. Fix remaining edge cases. Update vendor master, GL codes, bank mappings as needed.

Week 8: 100% of daily volume. Full automation. Your team monitors exceptions only.

The Integration Success Formula

Time Spent on Integration = 1 - (Manual Workarounds % After Go-Live)

Spend 4 weeks on integration properly: you get 95% automation day 1, zero manual workarounds. Spend 1 week on integration: you get 60% automation, 40% manual workarounds for 6 months. Which team wins?

Need an integration checklist? Download our 50-point integration audit to ensure nothing falls through the cracks.