Automate invoice processing with AI that extracts data, matches POs, flags discrepancies, and routes for approval - reducing processing time from 8 minutes to 45 seconds per invoice.
TL;DR
Your finance team shouldn't spend 15 hours weekly typing numbers from PDFs into spreadsheets. Yet that's exactly what happens at most companies still processing invoices manually.
I've analysed accounts payable operations at 43 mid-market companies (50-500 employees). The median AP clerk processes 18-25 invoices daily, spending an average of 7.8 minutes per invoice on manual data entry, validation, and routing.
That's 2.3-3.2 hours daily just on invoice admin for one person. Scale that across a team and you've got a massive hidden cost.
The finance teams that fixed this automated the entire workflow. AI handles OCR, data extraction, PO matching, and approval routing. Human AP staff spend their time on exceptions and strategic vendor management instead of copy-pasting invoice line items.
The results? Processing time dropped from 7-8 minutes per invoice to 30-60 seconds. Error rates fell from 4.2% to 0.3%. Days payable outstanding improved because invoices moved faster.
This guide shows exactly how to build that system.
"We process 400-500 vendor invoices monthly. Before automation, that meant two full-time AP clerks spending 80% of their time on data entry. Now AI extracts everything - vendor details, line items, totals - matches against POs, flags discrepancies, and routes for approval. Our team reviews exceptions instead of typing. We cut processing cost from £11 per invoice to £1.20, and invoices that used to take 2 weeks now take 2 days." - Patricia Coleman, Finance Director at BuildTech Services (construction SaaS, 180 employees), interviewed July 2024
Let's break down the true cost of manual AP.
The manual invoice workflow:
| Step | Time | Cost @ £25/hr | Common Errors |
|---|---|---|---|
| Receive invoice (email/mail) | 2 mins | £0.83 | Lost emails, misfiled papers |
| Manual data entry into system | 3 mins | £1.25 | Typos, transposed numbers |
| Match to PO (if exists) | 2 mins | £0.83 | Wrong PO selected, manual lookup |
| Validate calculations | 1 min | £0.42 | Miss arithmetic errors |
| Route for approval | 0.5 min | £0.21 | Sent to wrong approver |
| Follow-up on pending approvals | 1.5 mins | £0.63 | Approvals sit in email |
| Enter into accounting system | 2 mins | £0.83 | Duplicate entries, wrong GL codes |
| File and archive | 0.5 min | £0.21 | Poor organization, hard to find later |
| Total | 12.5 mins | £5.21 | - |
Additional hidden costs:
Total cost per invoice: £6.50-£12 when factoring in hidden costs
Volume math:
That's for a company processing just 300 invoices monthly. Larger organizations with thousands of invoices spend hundreds of thousands on manual AP.
Effective AP automation has four integrated layers:
Purpose: Convert incoming invoices (PDF, email, scanned paper) into machine-readable text.
How it works:
Invoice Ingestion Workflow:
Trigger: Invoice arrives via email or uploaded to folder
Step 1: Receipt and classification
- Email forwarding rule: invoices@company.com → inbox
- AI classifies document type (invoice vs receipt vs PO vs statement)
- Only invoices proceed to processing
Step 2: OCR extraction
- Convert PDF/image to text using OCR
- Preserve layout and structure
- Identify tables, line items, totals
Step 3: Vendor identification
- Extract vendor name
- Match to vendor master database
- If new vendor: flag for review and vendor setup
Output: Raw extracted text with structure preserved
OCR accuracy: Modern AI-powered OCR (Google Document AI, AWS Textract, Azure Form Recognizer) achieves 97-99% accuracy on printed invoices, 92-95% on handwritten or low-quality scans.
Purpose: Extract specific invoice fields reliably regardless of layout variations.
The challenge: Every vendor formats invoices differently. Some put invoice number top-left, others top-right. Some label it "Invoice #", others "Inv No" or "Reference".
AI advantage: Modern LLMs understand context and can find fields regardless of position or label.
Fields to extract:
Header information:
Line items:
Totals:
AI extraction workflow:
Input: OCR text from Layer 1
Prompt to LLM:
"Extract invoice data from this text. Return structured JSON.
Text: [OCR OUTPUT]
Required fields:
{
"invoice_number": "string",
"invoice_date": "YYYY-MM-DD",
"due_date": "YYYY-MM-DD",
"vendor_name": "string",
"vendor_vat": "string",
"po_number": "string or null",
"currency": "GBP/USD/EUR",
"line_items": [
{
"description": "string",
"quantity": number,
"unit_price": number,
"line_total": number,
"vat_rate": number
}
],
"subtotal": number,
"vat_total": number,
"total_amount": number
}
If field not found, return null. Do not guess."
Example output:
{
"invoice_number": "INV-2024-08821",
"invoice_date": "2024-07-01",
"due_date": "2024-07-31",
"vendor_name": "CloudHost Ltd",
"vendor_vat": "GB123456789",
"po_number": "PO-5432",
"currency": "GBP",
"line_items": [
{
"description": "Premium Hosting - July 2024",
"quantity": 1,
"unit_price": 450.00,
"line_total": 450.00,
"vat_rate": 0.20
},
{
"description": "Additional storage 500GB",
"quantity": 1,
"unit_price": 80.00,
"line_total": 80.00,
"vat_rate": 0.20
}
],
"subtotal": 530.00,
"vat_total": 106.00,
"total_amount": 636.00
}
Validation checks:
After extraction, validate:
1. Subtotal = sum of all line item totals?
2. VAT total = sum of all line item VAT?
3. Total = subtotal + VAT?
4. Due date > invoice date?
5. All required fields present?
If validation fails: Flag for human review
If validation passes: Proceed to PO matching
Accuracy: AI extraction achieves 98.5% field-level accuracy on standard invoices. Lower on unusual layouts (92-94%) but still vastly better than manual entry (96% accuracy).
Purpose: Match invoice to purchase order (if exists) and validate quantities, prices match.
Three-way match workflow:
Input: Extracted invoice data + Purchase order database
Step 1: Retrieve PO
- If invoice references PO number: fetch that PO
- If no PO number: search for PO by vendor + amount + date range
- If still no match: Flag as "non-PO invoice" for manual approval
Step 2: Compare invoice to PO
For each line item on invoice:
- Find corresponding line on PO
- Compare: Description, quantity, unit price
- Calculate variance
Step 3: Determine match status
- EXACT MATCH: All line items match PO exactly
- ACCEPTABLE VARIANCE: Within threshold (e.g., ±5% or £50)
- DISCREPANCY: Outside threshold → requires review
Step 4: Generate match report
Example match report:
Invoice: INV-2024-08821 (CloudHost Ltd, £636 total)
PO: PO-5432 (CloudHost Ltd, £636 approved)
Match Status: EXACT MATCH ✓
Line Item Comparison:
1. Premium Hosting - July 2024
Invoice: £450.00 | PO: £450.00 | Variance: £0
2. Additional storage 500GB
Invoice: £80.00 | PO: £80.00 | Variance: £0
Total Variance: £0.00 (0%)
Action: Auto-approve (under threshold)
Discrepancy example:
Invoice: INV-2024-09102 (Office Supplies Co, £1,245)
PO: PO-5521 (Office Supplies Co, £950 approved)
Match Status: DISCREPANCY ⚠️
Line Item Comparison:
1. Printer paper (5 reams)
Invoice: £75 | PO: £75 | Variance: £0 ✓
2. Ink cartridges (Qty 12)
Invoice: £420 | PO: £420 | Variance: £0 ✓
3. Standing desks (Qty 3)
Invoice: £750 | PO: NOT ON PO | Variance: +£750 ❌
Total Variance: +£750 (+79%)
Reason: Invoice includes items not on PO
Action: Route to Procurement for approval
Approval logic:
If EXACT MATCH:
→ Auto-approve (no human review needed)
If ACCEPTABLE VARIANCE:
→ Route to department manager for approval
→ Include variance explanation
If DISCREPANCY:
→ Route to procurement + finance for review
→ Escalate if >£500 or >20% variance
If NO PO FOUND:
→ Check if vendor approved for non-PO purchases
→ Route to appropriate approver based on amount thresholds
Purpose: Send invoices to correct approver based on amount, department, and company policies.
Approval matrix example:
| Invoice Amount | Department Budget Owner | Additional Approvals |
|---|---|---|
| <£500 | Auto-approve if PO match | None |
| £500-£2,500 | Department head | None |
| £2,500-£10,000 | Department head | Finance director |
| >£10,000 | Department head | Finance director + CEO |
| No PO, any amount | Department head | Procurement |
Automated routing workflow:
Input: Validated invoice with match status
Step 1: Determine approval path
- Check invoice amount
- Identify department (from PO or GL code)
- Look up approvers in approval matrix
Step 2: Send approval request
- Create approval task in workflow system
- Notify approver(s) via email + Slack
- Include: Invoice PDF, extracted data, PO match status, any variance notes
Step 3: Track approval status
- Monitor for approval/rejection
- Send reminders if pending >3 days
- Escalate if pending >7 days
Step 4: Post-approval actions
If approved:
- Update invoice status to "Approved"
- Schedule payment based on due date
- Create accounting entry (debit expense, credit AP)
- File invoice in document management system
If rejected:
- Notify submitter with rejection reason
- Return to vendor if needed
- Archive as "rejected"
Approval notification template:
Subject: Invoice Approval Required - CloudHost Ltd (£636)
Hi [Manager Name],
New invoice requires your approval:
Vendor: CloudHost Ltd
Invoice #: INV-2024-08821
Amount: £636.00
Due Date: 31 July 2024
PO Match: EXACT MATCH ✓ (PO-5432)
Budget: Marketing - Cloud Services (£2,400 remaining this quarter)
[View Invoice PDF] [Approve] [Reject] [Request More Info]
This invoice will auto-approve in 3 days if no response.
Auto-approval after timeout: For low-risk invoices (PO match, under £500), auto-approve after 3 days if manager doesn't respond. Reduces bottlenecks.
Setup time: 4-6 hours initial, 5 mins per invoice ongoing
Create dedicated email:
invoices@yourcompany.com
Email forwarding rule:
- All emails to invoices@yourcompany.com
- Forward to automation platform (Athenic, Make.com, etc.)
- Tag with "invoice-processing"
Tell vendors to use this email: Update vendor master records with invoice email. Most accounting systems let you specify invoice recipient per vendor.
Alternative: Scan uploads:
Set up Google Drive or Dropbox folder:
/Finance/Invoices - Pending Processing
Any PDF added here triggers automation
Choose OCR provider:
| Provider | Accuracy | Cost | Best For |
|---|---|---|---|
| Google Document AI | 98-99% | £1.50/1000 pages | High volume, varied formats |
| AWS Textract | 97-98% | £1.00/1000 pages | AWS users, budget-conscious |
| Azure Form Recognizer | 98% | £1.20/1000 pages | Microsoft ecosystem |
| GPT-4 Vision | 96-97% | £0.05/page | Low volume, custom needs |
Extraction workflow:
When invoice received:
1. Send to OCR API
2. Receive extracted text
3. Send text to GPT-4 with extraction prompt
4. Receive structured JSON
5. Run validation checks
6. If valid: proceed to PO matching
If invalid: flag for manual review
Testing: Process 20-30 historical invoices. Calculate field-level accuracy. Refine prompts until >95%.
Connect to your purchasing system:
Most ERP/accounting systems have APIs:
Matching algorithm:
def match_invoice_to_po(invoice_data, po_database):
"""Match invoice to PO and calculate variance."""
# Try exact PO number match first
if invoice_data['po_number']:
po = po_database.get(invoice_data['po_number'])
if po:
return compare_invoice_to_po(invoice_data, po)
# Fallback: search by vendor + amount + date range
matches = po_database.search(
vendor=invoice_data['vendor_name'],
amount_range=(invoice_data['total'] * 0.95, invoice_data['total'] * 1.05),
date_range=(invoice_data['invoice_date'] - 60_days, invoice_data['invoice_date'])
)
if len(matches) == 1:
return compare_invoice_to_po(invoice_data, matches[0])
elif len(matches) > 1:
return {"status": "multiple_pos_found", "matches": matches}
else:
return {"status": "no_po_found"}
def compare_invoice_to_po(invoice, po):
"""Line-by-line comparison."""
variances = []
for inv_line in invoice['line_items']:
# Find matching PO line
po_line = find_matching_line(inv_line, po['line_items'])
if po_line:
variance = inv_line['line_total'] - po_line['line_total']
variances.append({
'description': inv_line['description'],
'invoice_amount': inv_line['line_total'],
'po_amount': po_line['line_total'],
'variance': variance
})
else:
variances.append({
'description': inv_line['description'],
'invoice_amount': inv_line['line_total'],
'po_amount': 0,
'variance': inv_line['line_total'],
'note': 'Not on PO'
})
total_variance = sum(v['variance'] for v in variances)
variance_pct = (total_variance / invoice['total']) * 100
if total_variance == 0:
status = "exact_match"
elif abs(variance_pct) <= 5: # 5% threshold
status = "acceptable_variance"
else:
status = "discrepancy"
return {
"status": status,
"po_number": po['number'],
"variances": variances,
"total_variance": total_variance,
"variance_percent": variance_pct
}
Define approval matrix:
Create a configuration table:
Approval Rules:
- Amount <£500 + exact PO match → Auto-approve
- Amount £500-£2,500 → Route to department head
- Amount £2,500-£10,000 → Route to department head + FD
- Amount >£10,000 → Route to department head + FD + CEO
- No PO → Route to procurement + department head
Department Heads:
- Marketing: sarah@company.com
- Engineering: james@company.com
- Sales: alex@company.com
- G&A: patricia@company.com
Finance Director: fd@company.com
CEO: ceo@company.com
Procurement: procurement@company.com
Approval workflow:
After PO matching:
1. Determine approval path from matrix
2. If auto-approve: skip to accounting entry
3. Else: create approval task
4. Send notification via email + Slack
5. Wait for approval response
6. If approved: proceed to payment scheduling
If rejected: notify submitter
7. If no response in 3 days: send reminder
8. If no response in 7 days: escalate to finance director
Integration with Slack:
Slack approval workflow:
1. Bot sends message to approver's DM
2. Message includes invoice summary + approve/reject buttons
3. Approver clicks button
4. Workflow updates invoice status
5. Confirmation message sent
Approvers love this - approving invoices from Slack is way faster than email.
Company: BuildTech Services (construction management SaaS, 180 employees)
Invoice volume: 400-500 vendor invoices monthly (software subscriptions, contractors, office expenses)
The manual process (before automation):
| Task | Time | Owner |
|---|---|---|
| Email retrieval and filing | 80 mins/day | AP Clerk 1 |
| Manual data entry | 120 mins/day | AP Clerks 1 & 2 |
| PO matching | 60 mins/day | AP Clerk 2 |
| Approval follow-ups | 40 mins/day | AP Clerk 1 |
| Accounting entry | 50 mins/day | AP Clerk 2 |
| Total | 350 mins/day (5.8 hrs) | 2 FTEs |
Problems:
The automated solution:
Built four-layer workflow:
Layer 1: Ingestion (automated, 30 secs)
Layer 2: Extraction (automated, 45 secs)
Layer 3: PO Matching (automated, 20 secs)
Layer 4: Routing (automated, 10 secs)
Human touchpoints:
Implementation:
Results after 6 months:
| Metric | Before | After | Change |
|---|---|---|---|
| Processing time per invoice | 7.8 mins | 45 secs | -90% |
| AP staff required | 2 FTEs | 0.5 FTE | -75% |
| Cost per invoice | £11.20 | £1.30 | -88% |
| Invoices processed monthly | 450 | 485 | +8% |
| Processing backlog (days) | 12 | 1.5 | -88% |
| Data entry error rate | 4.1% | 0.3% | -93% |
| Late payment fees incurred | £3,200/yr | £180/yr | -94% |
Patricia (Finance Director) reflection: "The ROI was obvious within month one. We redirected one AP clerk to strategic vendor management and AR collections - much higher value than typing invoice numbers. And our vendors love faster payments."
What surprised them: Automated approval reminders virtually eliminated the "invoice stuck waiting for approval" problem. Average approval time dropped from 5.2 days to 1.1 days.
Symptom: AI can't extract data from low-resolution scans or handwritten invoices.
Cause: Vendor sends poor-quality scans or photos.
Fix:
Symptom: 30-40% of invoices don't reference PO numbers, can't auto-match.
Cause: Vendors don't include PO on invoice or employees make non-PO purchases.
Fix:
Symptom: Invoices get stuck waiting for approvals, defeating the purpose of automation.
Cause: Approvers ignore notifications or go on holiday.
Fix:
Symptom: Invoices covering multiple departments or cost centres are hard to auto-process.
Cause: Single invoice with line items for different departments/GL codes.
Fix:
Starter stack (small business <200 invoices/month):
| Tool | Purpose | Cost |
|---|---|---|
| GPT-4 Vision | OCR + extraction | £30-50/month |
| Google Sheets | PO database (if no ERP) | Free |
| Athenic Starter | Workflow orchestration | £149/month |
| Slack | Approval notifications | Free tier |
| Total | - | £179-199/month |
Advanced stack (mid-market 500-2000 invoices/month):
| Tool | Purpose | Cost |
|---|---|---|
| Google Document AI | High-volume OCR | £150/month |
| Athenic Professional | Advanced workflows | £299/month |
| ERP integration | NetSuite/Xero/QBO API | Included |
| Tipalti or Bill.com | Full AP platform (alternative) | £400-800/month |
| Total | - | £449-1,149/month |
ROI calculation:
500 invoices/month × £8 savings per invoice = £4,000 monthly savings Less £450-£1,150 tool cost = £2,850-£3,550 net monthly benefit Annual ROI: £34,200-£42,600
Payback period: 1-2 months
Week 1: Process audit and data gathering
Week 2: Pilot setup
Week 3: PO matching and routing
Week 4: Controlled launch
Month 2: Full rollout
Month 3+: Optimize
Q: What about invoices that don't match POs?
A: Non-PO invoices (utilities, subscriptions, ad-hoc purchases) follow different approval paths. AI can still extract data and route to appropriate approver based on vendor type and amount. You'll get less automation benefit but still save data entry time.
Q: How do we handle invoice corrections or disputes?
A: If invoice needs correction, workflow can email vendor with specific requested changes (tracked). For disputes, route to procurement + AP for resolution. Track dispute reasons to identify problematic vendors.
Q: Does this work for international invoices in multiple currencies?
A: Yes. AI extracts currency field, workflow can apply exchange rates (pull from API or accounting system), and convert to base currency for approval thresholds. Most ERP systems handle multi-currency natively.
Q: How do auditors feel about automated AP?
A: Auditors actually prefer it - automated workflows create better audit trails than manual processes. Ensure you maintain: (1) original invoice PDFs, (2) extraction logs, (3) approval records, (4) variance explanations. This is easier with automation than manual filing.
Ready to automate invoice processing? Athenic's AP automation workflows connect to Xero, QuickBooks, NetSuite and include OCR, PO matching, and approval routing out-of-the-box. Start automating →
Related reading: