AI for wholesale and logistics: where do you start? A six-step plan

How far along are trade and logistics with AI?

For its 2025 study of AI in trade and logistics, evofenedex surveyed almost three hundred logistics professionals. Sixty percent of the participating companies are fairly to very interested. Seven percent have developed a specific AI policy. Almost two thirds cite a lack of knowledge and skills as a barrier. For this step-by-step plan, one figure matters most: 42 percent of the companies have their internal data largely in order, and for a third that is the case only to a limited extent. The participants cite digital assistants, inventory management and demand forecasting as the most promising.

CBS measures actual use. According to provisional figures for 2025, published on 12 December 2025, 12 percent of companies in the trade sector use AI, and 7 percent in transport and storage. Across all sectors, 27 percent of companies with 10 to 50 persons employed use AI, and 45 percent of companies with 50 to 250. Of the companies that considered AI but do not use it, 73 percent give lack of experience as the reason.

A survey by software vendor Descartes, published on 6 May 2026, arrives at much higher figures. The survey was conducted in 2025 among 300 senior decision-makers at European shippers and logistics service providers, 75 of them from Belgium and the Netherlands. Only 3 percent say they do not use AI in modernising transport processes. Of all respondents, 41 percent use AI to automate data entry and 37 percent for route and load optimisation. That is a different group, surveyed by a company that itself sells logistics software, and therefore cannot be compared with the CBS figures.

The common thread in the figures from evofenedex and CBS: the barrier is rarely the technology, but knowledge and data. That is why the steps below do not start with a tool.

Which processes lend themselves to AI first?

Start where there is a lot of repetition, the rules are clear and an error quickly becomes visible. In wholesale and logistics, that usually means five processes:

• Order intake: orders from email, PDF or a portal that are now copied into the ERP by hand.
• Planning: trip and load planning, and replanning when something drops out or is delayed.
• Inventory: demand forecasting and order recommendations per item or product group.
• Customer service: status questions, questions about delivery times and first responses to complaints.
• Documents: CMR consignment notes, packing slips, purchase invoices and certificates that are read or filled in.

The step-by-step plan: six steps for every process

The steps are the same for every process. Do not choose the process with the most exceptions, but the process where your people lose the most time retyping, searching or repeating. Exactly what you check and measure in steps 2 to 5 varies by process and is set out in the sections that follow.

• 1. Choose one process and one owner. Describe in two sentences what currently goes wrong and for whom. The owner comes from the line organisation and runs the process every day, such as a planner, buyer or team leader.
• 2. Carry out the data check. Check whether the data the process runs on is correct and complete, and who maintains it.
• 3. Record the baseline measurement. In an ordinary week, measure how often the task occurs, how much time it takes and how often something has to be corrected.
• 4. Determine the autonomy level for each step. We use four levels: proposing, preparing, executing after approval and executing independently. Almost always start with proposing or preparing.
• 5. Test for 30 days alongside the current work. Week 1 is for learning. In weeks 2 to 4, you measure in the same way as for the baseline measurement, including the time for checking and correction.
• 6. Decide on the basis of the figures and record the decision. Continue, adjust or stop. If you continue, train everyone who runs the process, for example with in-company AI training, and record the working agreements.

Order intake: from email and PDF to a draft order

In order intake, the gain lies in the retyping. A system reads the order, matches customer and items and prepares a draft. Whether that works depends mainly on your master data. Microsoft writes of the Sales Order Agent in Business Central that the quality of the product information affects how well the agent finds items. That agent recognises customers only by the email address on the contact card. What such an agent may and may not do on its own beyond that is set out in AI agents in operations.

• Data check: does each customer have the right email addresses on file? Are customers' and suppliers' item numbers linked to your own numbers? Are units and packaging correct?
• Baseline measurement: orders per week per channel, minutes per order from receipt to confirmation, and the number of corrections or credit notes caused by entry errors.
• Autonomy level: preparing. The system creates a draft order, an employee checks the discrepancies and passes it on. The confirmation to the customer only goes out after approval. Microsoft shows the same caution with its agent: it always creates a quote first, because a quote has no effect on planning, reservations and availability, and it first submits all outgoing messages to an employee.
• First 30 days: one channel, for example email orders from regular customers. For each order, note whether the draft could go through unchanged, and which fields went wrong.

Planning: AI proposes, the planner decides

In planning, do not expect a completely different route. Look at replanning when something drops out or is delayed, and at travel times estimated from your own trip data. The rules on driving and rest times continue to apply and belong in every proposal as a hard limit. What this does and does not deliver is set out in our article on AI route optimisation in transport.

• Data check: are addresses, loading and unloading times and time windows correct? Are the actual driving times from the on-board computer or TMS available, alongside the planned ones? Are vehicle characteristics such as refrigeration or a tail lift in the system?
• Baseline measurement: planning hours per day, the number of replans after a drop-out or delay, empty kilometres or load factor, and the share of deliveries within the time window.
• Autonomy level: proposing. The system makes a proposal, the planner decides. The planner knows what is not in the system, such as an agreement with a customer or the refrigerated capacity that is really free.
• First 30 days: have the system propose a plan for the next day for one depot or region, alongside the planner's own. Compare kilometres and trips, and note each time why the planner deviates.

Inventory: order recommendations alongside the buyer

The participants in the evofenedex study cite inventory management and demand forecasting as promising. It is also the process where poor data has the fastest knock-on effect. What was sold out appears in your sales history as a stock limit and not as demand. A detailed step-by-step plan is in our article on AI in inventory management.

• Data check: an item master without duplicates, a sales history in which promotions, lost sales and one-off large orders are flagged, and lead times per supplier as they really are. For fresh and chilled products, also shelf life and waste.
• Baseline measurement: lost sales, average inventory, rush orders and waste, all for one product group.
• Autonomy level: proposing. The model gives an order recommendation, the buyer or planner decides and orders.
• First 30 days: choose a product group of 50 to 200 items and set the recommendation alongside the buyer's every week. Thirty days is short for a forecast: use them to see where model and buyer differ and why, and let the test run on to eight weeks before you judge.

Customer service: draft replies first

In the customer service of a wholesaler or haulier, a lot of time goes into the same questions: where is my shipment, when will it be delivered. AI can prepare those answers from your own systems.

• Data check: is the status in your TMS or WMS up to date at the moment the customer emails or calls? Are delivery terms, standard replies and exception rules recorded in one place?
• Baseline measurement: questions per week per type, the time to first response, minutes per reply and the share of questions passed on to planning or sales.
• Autonomy level: preparing. AI creates a draft, an employee sends it. An independent reply to a simple status question is possible later, if the customer knows they are communicating with AI and can always get through to an employee. Since 2 August 2026, Article 50 of the AI Act has required providers to design systems that communicate directly with people in such a way that people know this. Commitments on delivery time, compensation or terms stay with a human.
• First 30 days: one type of question, for example status questions by email. Count how many drafts could be sent unchanged.

Documents: filling in yes, signing off no

In the Descartes survey, automating data entry is the most frequently mentioned application, at 41 percent. AI can read and fill in documents. Your company remains responsible for what is on a consignment note, customs document or purchase invoice. How to set that up is explained in our article on document extraction for CMR, T1 and eCMR.

A fixed date is also approaching. According to TLN (Transport en Logistiek Nederland, the Dutch transport and logistics trade association), authorities will be obliged from 9 July 2027 to accept freight information electronically under the eFTI Regulation. From that date, companies will have the right to make freight information available electronically during inspections. If you already read and check documents digitally now, you will in any case have that data to hand in digital form by then.

• Data check: which document types come in and in what form: digital, scanned or filled in by hand? Where is the source you check against, such as the order or the trip in the TMS?
• Baseline measurement: documents per week per type, minutes per document, the number of corrections afterwards and the share of exceptions.
• Autonomy level: preparing. AI fills in the document and flags anything doubtful, an employee checks and signs.
• First 30 days: one document type, for example the purchase invoices or packing slips from your five largest suppliers. Measure how many documents come back as an exception. That tells you more than an average accuracy figure.

After day 30: continue, adjust or stop

Set the measurement from weeks 2 to 4 against the baseline measurement, including the time for checking and correction, otherwise the gain comes out too high. The calculator (in Dutch) converts the outcome into an annual amount: an indication based on your own input, not a promise.

Stopping is a valid outcome. If you continue, three things determine whether it keeps running: an owner in the line organisation, someone who maintains the data and a team that knows what the system can and cannot do. Why pilots stall on precisely these points is explained in why AI projects stall after the pilot.

There is also a legal reason why training is part of that. Article 4 of the AI Act has applied since 2 February 2025 and asks organisations that provide or use AI to take measures to support the AI literacy of their staff. Following an amendment to the AI Act in July 2026, that remains an obligation, but without a prescribed level. According to the European Commission, in its questions and answers of 27 July 2026, no certificate is needed for this and an internal record of training is sufficient. So record who has been trained in what and include the AI step in your work instructions, and if you work to a standard such as GMP+, BRC or IFS, do so in the language of that standard.

What does a first step cost?

Software costs vary so much by vendor and licence that we do not give an amount for them. In a trial, ask for the cost per processed order, trip or document and set it against your baseline measurement.

Our own Practice Day costs from €2,500 per group, excluding VAT. On that day, together with the people who run the process, we build one work process from your own operation, at your own location. How that works in logistics is set out on the page AI training for logistics (in Dutch).

The SLIM scheme (the Dutch subsidy scheme for learning and development in SMEs) does not reimburse training costs. An SME can, however, receive a subsidy for activities such as a company review with a training plan or the introduction of a learning method: 60 percent of the eligible costs, up to €24,999 (€20,000 for agricultural businesses). The 2026 application windows have closed and those for 2027 have not yet been published. The current status is on our SLIM page (in Dutch).

All articles on AI in logistics, by process

This step-by-step plan is the overview. The detail on each topic is in these articles:

• Order intake and agents: AI agents in operations: what may an agent do on its own?
• Planning and transport: AI route optimisation and driving times, the Dutch truck toll and AI in trip planning and the driver shortage and AI.
• Inventory and warehouse: AI in inventory management, a step-by-step plan, demand forecasting before the December peak and AI in the warehouse and e-fulfilment.
• Documents and reporting: document extraction for CMR, T1 and eCMR and CO2 per shipment in accordance with ISO 14083.
• People and implementation: why AI projects stall after the pilot and which tasks are changing in logistics and food.