Invoice Automation for SMEs: OCR → Validation → Posting

From PDFs to posted entries without chaos. This guide shows how SMEs can automate invoice intake with OCR, structured validation, and safe human approvals before posting to the accounting system. We keep two tracks: Manager Mode for decisions and ROI, Builder Mode with pipelines, schemas, and prompts.

Manager Mode – no-code playbook for SMEs

This section is decision-first. Use it to scope a 30-day pilot, define approvals, and pick tools by budget.

Scope your pilot

• Volume: 300-1,500 invoices per month works well for a first sprint.
• Sources: accounts@ inbox, mobile scans, supplier portals, e-invoice PDFs.
• Systems: your DMS for archiving, accounting/ERP for posting, chat/email for approvals.
• People: AP lead as owner, 1 backup approver, 1 ops builder.

Before you commit, estimate value and risk. The table below gives a lightweight ROI model you can adapt in a spreadsheet.

Approval policy – simple and auditable

• Auto-post: vendors on whitelist, PO match within tolerance, total under €1,000, VAT valid.
• Needs approval: new vendor, PO mismatch, total ≥ €1,000, missing VAT or bank change.
• Dual approval: spend ≥ €5,000 or sensitive cost centers.
• Escalation: no response in 24 hours – ping backup approver.

Success metrics

• Straight-through processing rate: target 60-80 percent in month 1.
• Exception aging: median under 1 business day.
• Posting accuracy: 99 percent fields correct after approval.
• Cycle time: inbox to posted under 24 hours for most invoices.

Tooling picks by budget

• Starter: shared inbox rules, hosted OCR, low-code automation, chat approvals.
• Pro: vendor master sync, PO match, LLM extraction with validator rules, audit logs.
• Enterprise: e-invoicing, SSO, fine-grained RBAC, encryption keys, custom evaluations.

Builder Mode – pipeline, schemas, prompts

This appendix is for the person wiring tools. Copy what you need and adapt to your stack.

Reference pipeline

1. Intake: watch accounts@ mailbox, vendor portal folder, and scanner uploads.
2. OCR: convert to machine text + layout coordinates.
3. Extract: LLM parses fields into JSON using a strict schema.
4. Validate: check vendor ID, bank IBAN, VAT, currency, PO lines and totals.
5. Approve: if validations pass and policy allows, auto-post. Else, send to exception queue.
6. Post: create vendor bill with line items, tax, cost centers, and attachments.
7. Archive + audit: store source PDF, JSON, hashes, and approval trail.

Use a strict schema so your extraction and validations are predictable. Here is a compact JSON schema you can mirror in your automation tool.

{
  "invoice_id":"string",
  "invoice_number":"string",
  "invoice_date":"YYYY-MM-DD",
  "due_date":"YYYY-MM-DD",
  "currency":"EUR|USD|DKK|...",
  "vendor":{
    "name":"string",
    "vendor_id":"string",
    "vat_id":"string",
    "iban":"string",
    "email":"string"
  },
  "buyer":{
    "company":"string",
    "cost_center":"string",
    "po_number":"string"
  },
  "totals":{
    "subtotal": "number",
    "vat": "number",
    "grand_total": "number"
  },
  "lines":[
    {"sku":"string","description":"string","qty":"number","unit_price":"number","vat_rate":"number","total":"number"}
  ],
  "hashes":{"pdf_sha256":"string","json_sha256":"string"},
  "attachments":[{"name":"string","url":"string"}],
  "source":{"channel":"email|scan|portal","filename":"string"},
  "validation":{
    "vendor_match": "pass|fail",
    "bank_match": "pass|fail",
    "vat_valid": "pass|fail",
    "po_match": "pass|warn|fail",
    "totals_check": "pass|fail",
    "policy":"auto|needs_approval|dual_approval",
    "messages":["string","string"]
  },
  "approvals":[{"by":"user@domain.com","at":"ISO-8601","note":"string"}]
}

Validator rules – minimal set

• Vendor match: vendor_id must exist in your master vendor table.
• Bank match: IBAN equals last approved IBAN for vendor. If changed, force approval.
• VAT valid: country format check and checksum where applicable.
• PO match: each line maps to a PO line, price within tolerance (for example ±2 percent).
• Totals check: sum(lines.total) + VAT equals grand_total within 0.01.
• Policy engine: compute auto vs approval vs dual approval based on thresholds.

Prompts should be deterministic and constrained. Use XML or JSON with explicit field lists and return nothing extra.

Extraction prompt template

System:
You are an extraction engine. Return ONLY valid JSON matching the provided schema. Do not invent fields.
User:
Extract the following fields from the invoice text and layout.
- Required: invoice_number, invoice_date, due_date, currency, vendor.name, vendor.vat_id, vendor.iban, buyer.po_number, totals.*, lines[*].description, lines[*].qty, lines[*].unit_price, lines[*].vat_rate, lines[*].total
- Return strictly in the given JSON schema. If a field is missing, use null and add a message under validation.messages.
Invoice text:
{{OCR_TEXT}}
Layout hints (x,y,section):
{{LAYOUT_HINTS}}
Schema:
{{JSON_SCHEMA}}

Approval message template

Subject: Approve invoice {{invoice_number}} - {{vendor.name}} - {{totals.grand_total}} {{currency}}
Body:
Please review. Auto-checks: vendor={{vendor_match}}, bank={{bank_match}}, VAT={{vat_valid}}, PO={{po_match}}, totals={{totals_check}}.
Approve: ✅ Reply "approve {{invoice_id}}"
Reject: ❌ Reply "reject {{invoice_id}} reason: ..."
Open PDF: {{pdf_url}} • JSON: {{json_url}}
SLA: 24h, escalates after 1 reminder.

Automation outline – Zapier/Make style

1. Trigger: new email in accounts@ with PDF or new file in folder.
2. OCR step: call OCR API, store text + layout.
3. LLM extract: call extraction prompt with schema, get JSON.
4. Validate: lookups to vendor master, PO, VAT, and bank registry.
5. Branch: if policy=auto and all checks pass, post – else send approval message.
6. Post: create vendor bill in accounting app with lines and attachment.
7. Archive: save PDF, JSON, and audit record to DMS with hashes.
8. Metrics: log result, timings, and exception reason for dashboard.

Error handling and retries

• Idempotency: compute pdf_sha256 and skip duplicates.
• Retry policy: exponential backoff for OCR/LLM up to 3 times.
• Dead-letter: if repeated failures, push to manual queue with reason.
• Security: redact PII in logs, rotate API keys, use least privilege.

Evaluation checks for quality

• Field coverage: percent of required fields extracted per invoice.
• Numeric consistency: totals and VAT recomputation equals payload.
• PO alignment: percentage of lines with exact or within-tolerance matches.
• Human corrections: fields adjusted by reviewers per 100 invoices.

FAQ – invoice automation and OCR

Do we need human approvals?
Yes for exceptions, new vendors, bank changes, or large totals. For routine invoices that pass checks, auto-post is fine with periodic audits.

Is OCR enough without an LLM?
OCR reads text. LLM-guided extraction helps structure messy layouts into reliable JSON for validation and posting.

What if vendors use different layouts?
Schema-first extraction normalizes outputs, so validations and posting remain consistent regardless of layout.

How do we stop fraud from bank account changes?
Compare IBAN against the last approved record. Any mismatch forces human approval with an out-of-band vendor verification.

Can we handle multi-currency and VAT?
Yes. Include currency and VAT rules in validators and map GL accounts per cost center.

What metrics should we track?
Straight-through rate, exception aging, posting accuracy, and time saved vs baseline.

Where do we store proofs?
Archive the source PDF, extracted JSON, validation results, and approvals with hashes for auditability.

What about data privacy?
Redact sensitive fields in logs, apply least-privilege scopes, and keep a vendor DPA on file with data region and retention notes.

Further reading

• AI Agents Explained – planners, tools, memory, approvals.
• Automation Workflows – connect email, CRM, docs with approvals.
• AI Data Privacy 101 – redaction, retention, vendor responsibilities.

Final thoughts

Start with a 30-day pilot across your top vendors, lock in a schema and validator rules, then scale. Most value comes from consistent approvals, clean archives, and tight exception handling. When ready, extend the same pattern to expenses, receipts, and purchase orders.