Intelligent invoice processing should turn documents into verifiable proposals, never into blind postings or blind payments. A reliable architecture keeps the original file, the extraction, the rules, the corrections and the approval; validates supplier, tax ID, amounts, taxes and IBAN; detects duplicates; and separates capture, posting and payment through distinct permissions.
End-to-end workflow
- Receipt through an authorised channel.
- Antivirus and format screening.
- Classification.
- OCR and extraction.
- Deterministic validations.
- Cross-check against master data and purchase orders.
- Exception handling.
- Approval.
- Posting.
- Archiving and traceability.
Each stage has a defined status and owner: a document must never disappear inside a "black box".
Fields and risk
| Field | Risk | Control |
|---|---|---|
| Supplier/tax ID | Wrong posting or fraud | Master data and tax validation |
| Number/date | Duplicate or wrong period | Composite key and calendar |
| Base/rate/tax | Tax error | Recalculation |
| Total | Incorrect payment | Independent sum check |
| IBAN | Fraud | Master data and dual control |
| Order/delivery note | Invoice not received | 2- or 3-way matching |
| Due date | Cash flow | Contractual rule |
Overall accuracy is not a useful metric: it must be measured field by field, weighted by impact.
Original document and evidence
The original is kept with an identifier, a hash, a date, a channel and a link to the extraction. Corrections never overwrite the document without leaving a history.
At any point the system must be able to reconstruct:
- What document arrived.
- What the model read.
- Which rules failed.
- Who corrected it.
- Who approved it.
- What was posted.
- What was paid.
OCR, model and rules
OCR turns the image into text; a model can interpret the structure; rules validate the facts. The model should never be asked to decide whether a sum adds up when that calculation can be resolved deterministically.
The system must abstain when it encounters low-quality scans, handwriting, ambiguous fields or unfamiliar formats.
Image quality
Before extracting the data, it is worth checking:
- Resolution.
- Orientation.
- Complete pages.
- Contrast.
- Protected PDF.
- Duplicates.
- Attachments.
An unreadable image is returned to the supplier or routed to manual review. Values are never invented.
Suppliers and master data
Registering or changing a supplier sits outside the ordinary automated flow. Identity, bank account, terms and contact details are validated through an independent channel.
An IBAN change detected on an invoice never updates the master record automatically: it triggers reinforced verification.
Duplicates
Duplicate detection combines several signals:
- Tax ID.
- Number.
- Date.
- Amount.
- File hash.
- Similarity.
- Order.
- Bank account.
Likely duplicates are blocked and a person decides. The system also detects invoices resent under a different name.
Matching
For purchases backed by an order, three elements are matched:
- Approved order.
- Goods receipt or service.
- Invoice.
Tolerances are set for quantity, price and tax, and exceptions require a reason and approval.
For expenses without an order, category, cost centre, limit and owner apply.
Tax compliance
Rates, taxable bases, tax due, exemptions, reverse charge, dates and territory are validated against the applicable rules. AI can suggest, but tax classification needs rules and competent oversight.
The workflow is coordinated with the ERP, VAT ledgers, the issuer-side SIF/VERI*FACTU system where applicable, and B2B e-invoicing. These are separate layers.
Data protection
Invoices can contain names, addresses, accounts and sensitive line items. The data processor, sub-processors, transfers, retention and any use in training are all reviewed.
The model receives only the minimum necessary. Logs never copy full documents without need, and access is restricted by company, site and role.
Security
An invoice can contain text designed to manipulate the AI. That content is always treated as data, never as an instruction.
Controls include:
- Antivirus and sandboxing.
- Allowed file types.
- Field validation.
- Least-privilege access.
- Segregation between capture, approval and payment.
- MFA.
- Supplier allowlists.
- IBAN change alerts.
- Amount limits.
- Operation logging.
Human review
The interface shows the original and the extracted field side by side, together with confidence, applied rules and detected discrepancies. Review focuses on exceptions, not on mechanically rubber-stamping everything.
Approvers cannot modify rules or master data without a separate, additional permission.
Metrics
- Accuracy per field.
- Percentage requiring no intervention.
- Exceptions by cause.
- Duplicates detected.
- Cycle time.
- Later corrections.
- Blocked payments.
- Cost per valid invoice.
- Privacy and security incidents.
Segment by supplier and format to fix the root cause at source.
Test set
The test set must include:
- Standard invoices.
- Multi-page documents.
- Credit notes.
- Different tax rates.
- Poor image quality.
- Duplicates.
- Changed IBAN.
- New supplier.
- Extreme amounts.
- Malicious instructions.
- Documents that are not invoices.
90-day plan
Days 1-30
Baseline, fields, rules, master data and test set.
Days 31-60
Integration, validations, permissions and audit trails.
Days 61-90
Pilot, full review, metrics and gradual automation.
Common mistakes
- Measuring only OCR.
- Not keeping the original.
- Updating the IBAN automatically.
- Posting without rules.
- Not detecting duplicates.
- Mixing companies together.
- Giving the agent payment access.
- Not testing exceptions.
- Logging full documents.
- Automating before cleaning up master data.
Checklist
- Original and hash.
- Accuracy per field.
- Mathematical and tax rules.
- Suppliers and IBAN under control.
- Duplicates.
- Matching.
- Segregated permissions.
- Document security.
- Exception review.
- Metrics and regression testing.
Frequently asked questions
Can it post invoices without human intervention?
Only in low-risk scenarios, and only after sufficient evidence, deterministic rules, limits and sampling.
Are OCR and AI the same thing?
No. OCR transcribes; AI can classify and interpret. Both need validation.
Can the IBAN change automatically?
No. It must be verified through an independent channel with dual control.
At Summum AI we design the extraction, validation, integration and evaluation of workflows like this one, with a complete audit trail. Our intelligent document processing service builds on AI-powered document classification and AI process automation to cover every stage, from receipt through to posting.