"What does the AI actually do at your place?" The question comes up in almost every first call. Usually from the managing director, usually after the demo. It is a fair question. BVL surveyed over 200 specialists and executives in the German-speaking region in February 2026. Two thirds want to introduce or expand AI within five years. The obstacles they name: missing expertise, poor data quality, scarce resources. Put differently: the intent is there, the picture is missing.
Here is the picture. Three jobs where AI agents run at our customers today. And one job we deliberately keep away from them.
The most common use. An order arrives as a mail with a PDF. An advice note arrives as a scan. A complaint arrives as running text with three attachments. A language model reads it and fills the fields the TMS needs. Customer, reference, loading point, date, quantity.
What the model does well: handle variants. One customer writes "pickup Thu 11 Sep 8 to 10", another sends a table, a third a handwritten slip. A classic program needs a rule per variant. The model needs an example and a check.
What it must not do: decide alone. Every field it reads is checked against the master. Does the customer exist? Is the postcode plausible for the city? Does the weight fit the pallet count? Only then is the order booked. At a wholesaler with 50 orders a day, 80 percent land in the system this way without a human touch. The rest goes to a review list with a reason. The case study describes the checks.
Customer portals, slot bookings, toll portals, telematics screens. Some have an API, many do not, and the customer's API rarely arrives. An agent operates the portal the way a dispatcher does. Log in, open the list, check the order, confirm.
Here AI is only part of the agent. Logging in and clicking is classic browser automation, which is more reliable and cheaper. The model steps in when the screen changes or an unknown message appears. "Your password expires in three days." "This order was cancelled." The agent reads the message, classifies it and decides whether to continue or to put the case on the review list.
At a contract logistics provider, one such agent has been checking between 900 and 5,300 orders per run for eight weeks. The variation comes from the customer, not the agent: Wednesday is correction day there. (They did not know that before. The protocol showed them.)
Reconciling toll, trip data and invoice per tour is arithmetic; it needs no AI. What AI adds is the step after: explaining the deviation. "Invoice 27 kilometers above toll booking, section A2 charged twice" is something a person can check at once. "Delta 27" is not. The agent reads the lines on both sides and phrases the most likely cause. With the note that it is a guess.
That does not save the fleet manager the decision, but it saves the half hour of searching before it. In our example data set with five vehicles and 1,284 toll bookings in a month (the video runs on our home page), it finds four deviations. All before the invoice gets paid.
The booking itself. Anything that moves money or binds legally is done by deterministic code with tests at our place. The booking into the TMS, the invoice, the confirmation in the portal. A language model does not always answer the same input the same way. For reading, that is an advantage, because it copes with variants. For booking, it is a risk we do not take.
The line is easy to remember: AI reads, classifies and explains. Code checks, books and logs. People decide the review cases.
BVL names data quality as an obstacle, and that matches every project we have done. The model is not the problem. It is customer masters with three spellings of the same company. Article numbers only one person knows. Price matrices with contradicting rules. An agent makes these errors visible because it writes them to the review list every day. That is uncomfortable, and it is the real value of the first weeks.
A caveat: the examples come from companies with 20 to 500 employees and one or two runs a day. For real-time decisions in tour planning, seconds instead of minutes, the split between model and code looks different. I have no operating data of my own on that.
Which of your tasks is reading, which is operating, which is checking? Name one where someone retypes or opens a portal today. I will show you the protocol of an agent doing exactly that. Ask for a process check: 30 minutes, no sales pitch.
