AI Invoice Processing: How Agents Pre-Code Invoices While Keeping Dual Control

By Dennis Weidner · 10. August 2026

In many mid-sized finance departments the daily reality still looks the same: some invoices arrive by email, others by post, and a growing but still small share as structured e-invoices. Each one has to be read, checked, coded and approved before it enters DATEV or an ERP system. This is exactly where AI agents come in, drafting the accounting entry and flagging anomalies before a human ever looks at the invoice. The important part: dual control does not disappear, it simply becomes more efficient rather than being replaced.

The current picture: three channels, one bottleneck

How fragmented invoice receipt still is in Germany becomes clear from the Bitkom Company Survey on E-Invoicing 2024: 96 percent of companies receive invoices by email, 93 percent still by letter post, while only 45 percent actually use structured e-invoice formats such as XRechnung or ZUGFeRD. In other words, most invoices still arrive as PDFs, scans or paper and need to be made machine-readable before any automated checking can even begin. With Germany's phased B2B e-invoicing mandate taking effect from 2025, this balance will shift, but until then finance teams have to handle all three channels in parallel.

What an agent actually does during pre-coding

A well-configured invoice agent takes over the steps that cost time today but require little professional judgement:

• Capturing invoices regardless of channel, including text recognition for scans and PDFs, and direct processing of structured e-invoices
• Extracting header and line-item data: vendor, invoice number, amount, tax rate, service period, purchase order reference
• Matching against purchase orders and delivery notes where available in the ERP system (three-way match)
• Proposing account coding, cost centre and tax code based on historical booking patterns
• Handing the draft entry over to DATEV or the relevant ERP system as a proposal, not as a posted transaction

The agent works as a pre-checker, not a decision-maker. It reduces the manual data entry that consumes most of the time in many finance teams and frees up capacity for the actual professional review.

Catching discrepancies before they become a problem

The real value lies less in speed than in consistency. An agent applies the same rules to every invoice, regardless of workload or time of day. Typical patterns that get flagged include:

• Price or quantity deviations from the purchase order or framework agreement
• Duplicate invoices from the same vendor with identical or slightly altered invoice numbers
• Missing mandatory information required under VAT law
• Recently changed bank details, a classic pattern in payment fraud
• Invoices without a matching purchase order or outside agreed framework terms

These cases are not posted automatically but routed to the responsible person for clarification. Processing time drops mainly where every invoice used to receive the same level of scrutiny, regardless of whether it was actually unusual.

Dual control remains mandatory, not optional

Especially in the public sector and audited mid-sized companies, automation is often mistaken for a relaxation of controls. The two have nothing to do with each other. The German GoBD requirements for traceability, immutability and documentation apply unchanged, even when an agent is involved in pre-checking.

An agent proposes, a human decides. Reverse that order and you lose not just control, but potentially audit compliance as well.

In practice this means the approval workflow inside DATEV or the ERP system stays in place, including role-based permissions and complete logging of who reviewed and approved which draft entry and when. The agent supplies the groundwork and the reasoning behind flagged discrepancies. Final approval, particularly above defined thresholds, remains bound to at least two people. This is not a technical limitation but a deliberate organisational choice that preserves audit integrity toward auditors and supervisory bodies.

What it is actually worth

How much effort is genuinely saved depends heavily on invoice volume, channel mix and the existing ERP landscape. Blanket promises are not particularly useful here. A more reliable approach is to run the numbers through our German-language savings calculator (available in German only), which compares invoice volume, error rates and current staffing effort. For a broader view of where similar agents make sense across finance and administration, take a look at our services and insights.

From pilot to production

A realistic starting point is not the entire accounts payable department but a clearly scoped segment, such as a high-volume vendor or a single cost centre. This makes it possible to test how reliable the pre-coding actually is and where the checking logic needs adjustment before scaling further.

If you want to understand what an agent like this would mean for your invoice volume in concrete terms, get in touch through our contact page. We can show you how pre-coding, discrepancy detection and dual control fit together cleanly within your existing DATEV or ERP setup.