Key points in 30 seconds
- Finance teams were manually keying in hundreds of PDF invoices in inconsistent formats.
- An AI agent extracts data with an LLM, normalises it, checks duplicates and vendors, then inserts and archives.
- Result: 100% automated extraction and no duplicate invoice missed.
- The modular architecture means a new vendor or format is a configuration change.
Results in figures
From PDF to database, no re-keying
- PDF receivedIncoming invoices, whatever the vendor's format.Agentic orchestration
- LLM extractionVendor, amounts, dates, line items, tax, PO number.LLM
- NormalisationData standardised for consistency.JavaScript
- Duplicate checkEach invoice compared with history before entry.SQL
- Vendor checkValidation against the approved supplier database.SQL
- Insert and archiveStored in the database, renamed, classified, archived.Database · PowerShell
Five pain points solved
| Issue | Before | With the AI agent |
|---|---|---|
| Manual entry | Hundreds of PDFs keyed in by hand: slow and error-prone. | LLM extraction and automatic insertion, with no intervention. |
| Inconsistent formats | No standard fields from one vendor to the next. | The LLM structures the data, whatever the format. |
| Duplicates and fraud | The same invoice could be processed twice. | SQL check against history, before any payment. |
| Unverified vendors | No systematic validation of vendor details. | Automatic check against approved suppliers. |
| Disorganised archiving | Badly named files: audits and searches were difficult. | Renaming, classification and archiving by PowerShell. |
A finance team freed up and better protected
- Hours given back to finance: no more re-keying: the team focuses on analysis.
- Secure payments: duplicates and vendors checked before any payment.
- A complete audit trail: every invoice is traceable and found instantly.
- Ready to scale: new vendor or format: configuration, not development.
Technology stack
- LLM: reading unstructured PDF invoices and extracting financial data.
- Agentic automation: end-to-end pipeline orchestration and reusable sub-workflows.
- SQL database: duplicate detection, vendor validation and record storage.
- PowerShell: automatic file renaming, classification and archiving.
Why re-keying is costly
Manually entering supplier invoices ties up qualified staff on a task with no added value. It also causes errors, late payments and the risk of paying the same invoice twice. The more suppliers and formats there are, the higher the cost.
How a project of this kind unfolds
- Map the current process: receipt, entry, checks, approval, archiving.
- Build a sample of invoices representative of the formats received.
- Define control rules: duplicates, approved suppliers, amounts, VAT.
- Test in a pilot with systematic human review of the results.
- Go live, then follow up flagged cases and adjust the rules.
Conditions for success
- An up-to-date supplier master file, the basis for all checks.
- Written approval rules shared with finance.
- Human review of the cases flagged by the agent.
- A processing log for audit and continuous improvement.
See also what you can delegate to a machine and AI use cases with fast ROI.
In summary
An AI agent reads PDF invoices in any format, extracts and normalises the data, checks for duplicates and approved suppliers, then records and archives each invoice. In this project, extraction is fully automated and no duplicate invoice has gone undetected. The modular architecture means a new supplier or format can be added through configuration. This type of project falls within our AI and digital transformation support.
Frequently asked questions
Can an AI agent read invoices in different formats?
Yes. The LLM extracts the relevant fields (vendor, amounts, dates, line items, tax, PO number) whatever the vendor's format, then the data is normalised before any check.
How do you avoid paying the same invoice twice?
Each incoming invoice is compared with the history in the database by an SQL rule. Duplicates are flagged before they enter the financial system, so before any payment.
Does a new vendor require new development?
No. The modular sub-workflow architecture means a new vendor, entity or invoice format is added through configuration.
How long does it take to set up an invoice processing agent?
It depends on the number of formats and systems to connect. A pilot on a limited scope validates the approach in a few weeks before extending processing.
Does the AI agent replace human control?
No. It eliminates re-keying and flags anomalies; a person reviews flagged cases and keeps the decision on payments.