From 4,500 to 175,000 invoices a year: what these two extremes teach us about invoice processing using AI

A consultancy firm with 4,500 purchase invoices a year. A Benelux retailer with 175,000. Two completely different organisations, yet the same conclusion: the key to success lies in the design of the audit, and only then in the technical aspects.

Two organisations, a 40-fold difference

On paper, these two clients have little in common. One is a consultancy firm with over 150 employees, where invoices from subcontractors and freelancers need to be matched against recorded hours. The other is a retailer with several divisions across the Benelux, where some 175,000 goods invoices are reconciled against purchase orders and delivery notes each year.

The difference in volume is a factor of 40. The bottleneck is the same in both cases: a accounts payable process in which qualified staff spend most of their time on work that is predictable. Opening an invoice, entering the details, locating the source, comparing the figures, and processing the payment. Knowledge work that has become routine processing.

What is more interesting than the similarity is where the value comes from in each case. It turns out that this depends on the volume.

Case 1: the consultancy firm, where complexity is the cost driver

With 4,500 invoices a year, pure processing time is rarely the real problem. The real problem is the nature of the match: invoices from subcontractors must tally with the hours recorded in the project accounts. This is a three-way match against a dynamic source, as hours are adjusted, projects continue to run and rates vary depending on the agreement.

Two AI colleagues now carry out this work end-to-end: reading the invoice, selecting the correct supplier and the correct project, matching amounts against recorded hours, and preparing a booking proposal. The organisation deliberately opted for 100% human approval. Every invoice is approved by a member of staff before anything is posted to the accounts.

That sounds like semi-automation. The opposite is true. The employee assesses a fully prepared file rather than having to do the research themselves. The result: approximately 0.7 FTE shifted from administrative work to billable consultancy work. In an organisation that relies on billable hours, that is a more direct business case than any cost saving whatsoever.

Case 2: the retailer, where volume is the cost driver

With 175,000 goods invoices a year, the maths changes. Here, every second spent processing each invoice represents a significant cost, and 100% human approval is practically unfeasible. The question shifts from “how do we process each invoice” to “which invoices actually warrant human attention”.

The solution is a validation layer on top of the existing ERP system. The AI colleague compares every incoming invoice with the purchase order and the delivery note, checks amounts and tolerances, and processes those that are correct. Approximately 85% are processed fully automatically. The remaining 15% is sent to a review queue with full context: this differs, for this reason, and this is the proposal.

The least obvious figure is perhaps the most important: approximately 50% less rework. Errors that used to only come to light during the payment run or month-end closing are now flagged as soon as they are received. Correcting errors retrospectively in an accounts payable process is many times more expensive than carrying out checks in advance.

The pattern: control is the design; automation follows

Put the two cases side by side and a clear picture emerges:

The technology is the same in both cases: an AI colleague that operates natively within the ERP system, understands documents and escalates issues with context. The difference lies in the control design. At the consultancy firm, full human approval is a deliberate choice that is appropriate to the nature of the match and the volume. At the retailer, exception-driven control is the only workable approach. Both are valid. The level of control is a parameter that you set for each process, based on risk, volume and maturity, and which you can adjust over time as confidence grows.

That is precisely why automating invoice processing is, to half an extent, a finance issue. What tolerances do you apply? Where is the line between automatic processing and escalation? How does this fit in with your internal controls and segregation of duties? What does your accountant think? These are not software-related questions.

Why AgentsLab and Improven are working together on this

This is where the partnership with Improven comes into play. The division of roles is straightforward: AgentsLab develops the AI colleagues and the ERP integration, whilst Improven provides the business and finance expertise that determines what the control framework should look like.

Improven’s consultants have an inside-out understanding of accounts payable processes: the preliminary process analysis, the controls, the approval structure and how to present the data in a way that meets an auditor’s requirements. They ensure that the AI colleague fits in with the process as it should run, and not the other way round. AgentsLab then ensures that it actually works: live within weeks, measurable from day one, with the same audit trail as an employee.

The actual figures across both cases: a reduction of approximately 90% in processing time per invoice, a lead time reduced from days to less than two hours, and a conservatively estimated payback period of 8 to 11 months.

Want to see how this works? Come along to Data Expo 2026

AgentsLab and Improven will be exhibiting together at Data Expo 2026. At our stand, we’ll be demonstrating live how an AI colleague can take an invoice from your inbox and turn it into a booking proposal, and we’ll work through a business case on the spot based on your volume and ERP system.

Would you rather get started sooner? Kick off the process. Demonstrate the value in around 8 weeks.

Book a Quick Scan.