Automating business processes with AI: where it works and where it doesn't

Short answer. AI is suited to automating process steps where someone now reads messy information and transfers it: turning emails into order lines, pulling data from invoices, sorting messages, summarising text. For calculations that must be exact, such as margin and VAT, use regular software. Build AI in as one step in a mapped process, with checks on the exceptions.

AI isn't a tool you install. It's a change you have to lead. Still, I see companies buy an AI tool first and then look for a process to unleash it on. That is the wrong order. Start with the process and you will see at which step AI solves something regular software can't.

This article belongs to process optimisation: map the process first, then the software that fits.

What companies use AI for today

Adoption is growing fast, especially at mid-sized companies. According to Statistics Netherlands (CBS), 45 percent of Dutch companies with 50 to 250 employees used AI technology in 2025, up from 20 percent in 2023. Across all companies with two or more employees it was 17 percent (CBS, 2025).

What do those companies use it for?

Application	Share of AI-using companies (2025)
Marketing or sales	35%
Business administration or management tasks	32%
R&D or innovation	25%
Production or service processes	18%
Accounting, controlling or financial management	17%
ICT security	11%
Logistics	4%

Source: CBS, provisional figures 2025.

AI now mostly sits at the edges of the business, with text and marketing. In production, finance and logistics its use is still small. That is where the most is to be gained, and where you have to be most careful. How use differs by company size is in how many Dutch companies use AI and automation.

Which processes can you automate with AI?

AI is strong where the input is messy and a person now reads, interprets and re-types. These are the applications I see working most often in practice.

• Order intake from emails. Suppliers and customers email orders as free text, in tables or as attachments. AI pulls out items, quantities and prices and prepares them in your system. For resellers and wholesalers this is one of the biggest time sinks; see also reading orders from a supplier portal automatically.
• Pulling data from documents. Invoices, packing slips, receipts and customs papers, even when every supplier uses a different format. See reading invoices and receipts automatically.
• Sorting customer questions, return requests and notifications and routing them to the right person or process.
• Matching items that have ten different names at ten suppliers, where fixed rules keep missing an exception.
• Summarising and drafting: a weekly report from your numbers, a draft reply to a customer question, the gist of a long email thread.

Where AI is not the answer

Just as important is where you do not use AI.

Calculations that must be exact belong in regular, auditable software. Think margin, VAT, stock and prices. A language model gives an answer that sounds plausible, and it doesn't guarantee that it is right. For the real margin per product, channel and supplier the calculation belongs in a system.

Decisions with consequences shouldn't be left entirely to a model either: a credit limit, a rejection, a payment. Let AI prepare and a person decide on the cases that matter.

And then the data. If your sources contradict each other, AI doesn't make that better. It makes it wrong faster. First it has to be clear which source is leading. Automating an unnecessary step with AI gains nothing either. It stays an unnecessary step, only more expensive.

How I build AI into a process

AI becomes reliable when you treat it as one step in a process, with clear boundaries around it.

1. Map the process first. At which step is time lost because someone reads messy information and transfers it? That is where AI belongs.
2. Let AI do only that step. For example: read the email and propose a structured order. Creating the order, the price and the stock stay in your regular systems.
3. Check the outcome automatically. Does the item number exist? Does the quantity match the unit? Does the price deviate no more than agreed? Whatever fails the check goes to a person.
4. People keep looking, but only at what deviates. That is the gain.
5. Measure and adjust. How much goes right automatically, how much lands with a person, and why? That is how it gets better every month.

An example of what that delivers. If forty order emails arrive per day and each takes six minutes to transfer, that is four hours a day. If 80 percent passes the automatic check, eight emails remain for a person: just under an hour. The other three hours go to work that matters. How high that percentage gets for you depends on how messy your input is. You only get the honest answer after a trial with your own emails.

Whether you use an existing tool for this or have something built depends on how standard your input is. That trade-off is in off-the-shelf, integrate or custom. To get the process sharp first, one of the process optimisation models helps.