Many pages, little control: what goes unnoticed in freight invoices

Freight invoices can run to many pages. Whether every line matches the order and the agreed conditions is nonetheless something hardly anyone checks in full. A typical case shows why a small historical review can make more sense than the next big AI project.

One freight invoice covers a single day and runs to several dozen pages. In months with high shipping volumes, invoices like this keep arriving. Realistically, nobody can read them all the way through.

Take a mid-sized company with an ordinary invoice workflow. Documents are coded, approved and posted. For normal overheads that works. Freight invoices are harder. Logistics looks at them, checks whether the overall picture seems plausible, and passes them on.

Line-by-line auditing rarely happens. For a long time, nobody notices. Only when freight costs have risen unusually sharply at year end does someone pull samples. Then it sometimes turns out that a wrong condition was applied, or a surcharge was billed that cannot be found in any of the existing agreements.

That is not an AI problem at first. It is a simple commercial question: are we paying what we agreed?

Auditing by feel

A freight invoice can be formally in order. Supplier, invoice number, date, amount and cost centre are all there. Accounting can process it. Whether the correct weight band, the valid tariff and the agreed diesel surcharge were applied to a particular shipment remains unanswered.

To answer it, you would have to bring together the invoice with all its lines, the conditions valid at the time of shipping, and the freight order. Often these three pieces of information live in three different places: the invoice in the invoice workflow, the conditions in PDF or Excel files, the shipment in the transport system.

A clerk can gather the sources. For every line on every invoice, though, that becomes work that in daily practice only happens on a sampling basis. The carrier knows its billing in detail. The shipper often only knows the total that ends up on a cost centre.

Look first, then build

The obvious reaction today is: then let’s build an AI agent. Copilot Studio and similar tools are available, the in-house developers know the systems, a first prototype would be quick to produce.

I still would not start there.

Anyone who builds interfaces, encodes rules and runs yet another production process should know beforehand whether the problem is big enough. Perhaps the invoices are essentially correct. Perhaps a handful of audit rules would suffice. Or perhaps there is so much money in the deviations that continuous freight invoice auditing is the economically obvious step.

All three outcomes are possible. A sensible analysis has to allow for them.

So I would start smaller: one carrier, a clear period (the last year, say), the existing conditions and, where possible, the associated shipment data. No ERP integration, no new approval process. Just the question of what is hiding in the invoices that have already been paid.

In a typical case, two or three carriers cover a large share of the volume. You can start with one of them and one year of invoices. The documents can be exported from the existing systems. Logistics and IT know how the conditions are meant and what data the transport system can provide.

That is easier to explain internally than the introduction of a new platform. And it delivers an answer sooner.

Not every hit is an error

A historical review must not end in a dashboard. Every finding needs evidence: the invoice, the line concerned, the basis of comparison and the amount of the deviation.

What you can search for: surcharges not provided for in the existing conditions, wrong tariffs or weight bands, duplicate lines, and invoices with no clear reference to a freight order. Differences in currency, unit of measure or validity period also stand out as soon as the information is in one common structure.

Not everything flagged in the process is automatically wrong. A surcharge may sit in a side agreement missing from the first data package. A shipment number may be spelled differently in the transport system. An invoice may bundle several orders without showing the allocation cleanly.

So the sensible sequence is not: the AI finds an error and the invoice gets short-paid. The system lays the deviation open, together with its source position and the comparison. The customer decides whether and how to reconcile it.

The line items do the work

Reading a freight invoice sounds like classic document recognition. In practice, the pieces of information that belong together are often far apart. A shipment is named on one page, the costs follow later. Surcharges refer to several lines. Every carrier uses different terms and groupings.

Whether this reconciliation succeeds reliably enough has to be measured on a bounded set of invoices. A technical hit rate alone is not enough. What matters is whether line, order and condition are correctly brought together and whether a deviation stays traceable to its source. Even a good technical metric does not yet answer whether the specific case is economically worthwhile.

Small analysis, clear decision

What should stand at the end is not a stack of slides but a decision.

If deviations are rare and small, the existing process can stay. Perhaps a simple pre-check will do. If certain surcharges or tariff errors keep recurring, continuous auditing can be set up for exactly those. If a broad and regularly recurring problem shows up, freight bill auditing can later be built into the invoice process all the way to approval and posting.

I like this entry point because it spares both sides a big claim. The customer does not have to believe in a possible saving. We do not have to sell a platform before we have seen the data. We look.

Processing can run on our infrastructure in Austria, in a data centre in Switzerland, or inside the company’s own environment. For freight invoices that is rarely the reason to start an analysis. But it is a framework condition that is better settled at the beginning.

Afterwards both sides know more. Perhaps the problem is small. Perhaps it is big. Either would be a usable answer.

Anyone with the same question can likewise begin with one carrier and a bounded period. A historical review can be discussed here.