How to Build an Invoice Processing Agent (Playbook)

Accounts payable is where many organizations should start with agents: volume is high, rules are explicit, the baseline is measurable, and the payoff is visible to finance. This playbook covers building an invoice processing agent that extracts, matches, routes exceptions, and posts under controls, drawing on FISTA's AI agents practice. The functional overview is in ai invoice processing and ai accounts payable automation; the economics are in the Digital FTE economics whitepaper.

What does the agent do end to end?

Stage	Function	Control
Intake	Capture from email, vendor portals, scans, EDI	Deduplication; source logging
Extraction	Header and line fields with confidence	Field thresholds; human review of low confidence
Validation	Vendor master, tax, bank details, duplicates	Hard rules; fraud checks
Matching	Two- or three-way match to PO and receipt	Tolerance rules from the spec
Exception handling	Classify, gather context, draft resolution	Staff decision
Posting	Create ERP entry and schedule payment	Gates by amount, vendor risk, match status
Audit	Full record of extraction, decisions, approvals	Immutable log

Step 1: Establish the baseline and write the spec

Measure current volume by channel, cost per invoice, cycle time, exception rate by type, touchless rate, and error incidents. Then write the specification with finance: fields to extract and their accuracy thresholds, validation rules, matching rules and tolerances, exception categories and owners, posting gates by amount and risk, prohibited actions (never change vendor bank details automatically, never post unmatched invoices), and acceptance criteria. See how to write acceptance criteria for ai.

Step 2: Build intake and deduplication

Consolidate channels into one queue: email parsing for attachments, portal and EDI connectors, and scan ingestion. Deduplicate on vendor, invoice number, amount, and content hash before extraction. Log source and receipt time for audit and for measuring cycle time.

Step 3: Build the extraction pipeline

Combine document AI for layout and OCR with a language model for field interpretation where layouts vary. Extract header fields (vendor, invoice number, dates, totals, tax, currency, payment terms) and line items (description, quantity, unit price, PO reference). Produce a confidence score per field. Normalize vendors against the vendor master and dates and currencies to standard formats. The pipeline pattern is in how to build an ocr pipeline with llms and how to build an ai data extraction pipeline.

Step 4: Implement validation and fraud checks

Deterministic checks run before matching: vendor exists and is active, tax calculations are consistent, bank details match the vendor master (any change is a mandatory human review), invoice number not previously processed, and amounts within vendor norms. Suspicious patterns route to a fraud queue. See ai fraud detection.

Step 5: Implement matching under tolerance rules

Match to purchase orders and receipts using the spec's rules: PO reference, vendor, line quantities and prices within tolerances, and receipt confirmation for three-way matching. Matching is deterministic code driven by spec rules, with the model assisting only in ambiguous line-item reconciliation, where its suggestion is surfaced with evidence rather than applied. Record match status per line.

Step 6: Design exception handling

Exceptions are where staff time goes, so the agent's job is to make them fast:

1. Classify the exception: price variance, quantity variance, missing PO, missing receipt, duplicate suspicion, vendor data mismatch.
2. Gather context: PO, receipt, vendor history, requester.
3. Draft the resolution or the query to the vendor or requester.
4. Present to the AP specialist with evidence and one-click actions.
5. Learn: resolved exceptions and their outcomes feed the evaluation set and, where appropriate, rule refinements approved by finance.

Queue design is in how to build a human review queue.

Step 7: Implement gated posting

Posting rules from the spec: fully matched invoices below an amount threshold from established vendors post automatically with sampling; above threshold, new vendors, or partial matches require approval by an authorized approver; segregation of duties is enforced so the agent cannot both create vendors and post invoices. ERP integration uses the ERP's APIs with idempotency to prevent duplicate postings. Gate design is in what is a human approval gate and integration in how to build an erp ai integration.

Step 8: Evaluate and launch

Build a labeled set of several hundred real invoices across vendors, layouts, and exception types with finance-verified expected fields, match outcomes, and routing decisions. Measure field-level accuracy, match correctness, routing correctness, and fraud-check behavior. Launch in shadow mode alongside the current process, compare, then move to gated posting with sampling, expanding automatic posting thresholds as evidence accumulates. The method is in the AI evaluation and testing whitepaper and the staged path in ai pilot to production.

What controls does finance require?

• Immutable audit log of every extraction, validation, match decision, approval, and posting.
• Segregation of duties preserved in tool scopes.
• Mandatory human review for vendor bank detail changes.
• Sampling of automatically posted invoices by internal control.
• Access control on the agent's ERP credentials with least privilege.
• Change control and re-evaluation for rule, model, or prompt changes.

The governance framework is in the agentic AI governance whitepaper and the audit design in how to build an ai audit trail.

What does it cost to run?

Run cost is dominated by document processing and model calls per invoice and is modest relative to staff time; oversight cost falls as touchless rates rise. Cost per invoice processed against the baseline, including exception handling time, is the figure finance will track. Drivers are in document ai cost and the AI total cost of ownership whitepaper.

What are the common mistakes?

• Trusting vendor extraction benchmarks instead of testing on your own invoices.
• Letting the model decide matches instead of encoding finance's tolerance rules.
• Posting automatically before shadow-mode evidence exists.
• Ignoring bank-detail changes as a fraud vector.
• Building exception detection without exception resolution.
• No idempotency in ERP posting, producing duplicates.

How FISTA Solutions builds invoice processing agents

FISTA Solutions builds invoice processing agents to this playbook: finance-owned specifications with explicit tolerances and gates, confidence-scored extraction, deterministic matching, exception resolution designed for AP staff, gated ERP posting with segregation of duties, and labeled-set evaluation before shadow launch. The AI agents practice delivers the agent, AI enablement the document and integration platform, and forward deployed engineers embed with your finance team to establish the baseline and the rules. The record behind the work is 150+ projects with 47% average efficiency gains.

To scope an invoice processing agent, message FISTA on WhatsApp, or read document processing ai for the broader document automation context.