“We tried ChatGPT, but it did not stick”

We hear this in almost every intake. According to the Dutch Chamber of Commerce in June 2025, 30% of business owners use ChatGPT or Copilot. A large share slid back to the old way of working after a few enthusiastic weeks. That is not the people’s fault, and not the technology’s. It comes down to what was missing.

Why a chat window is not a process

An employee opens ChatGPT, pastes in a client email, and asks for a polished reply. It works. The second time too. But:

• They have to copy the email in and paste the answer back themselves.
• The AI does not know what this client ordered last week or what the delivery status is.
• The answer sounds good, but whether it is right they have to look up themselves.
• Their colleague does it slightly differently, or not at all.

After a month it takes as much time as before, only with an extra step. So they stop. A chat window is a tool for one person, at one moment, with no memory of your business. A process is something that runs the same way every day, for everyone, with the right data.

Three missing links: source, integration, and approval

These are the three things a chat tool does not have and a business solution does.

Without these three, every AI attempt is an experiment. With these three, it is a working part of your business.

From answering questions to completing tasks

The difference between a chat window and an agent is in the verb. A chat answers. An agent does.

An example from logistics. A client emails: “Where is my shipment with reference 8842?”

• Chat window: the planner copies the question, asks for a polished wording, looks up the status in the TMS, pastes it in, and copies the answer back.
• Agent: the agent recognizes the status question, looks up reference 8842 in the TMS (read-only), drafts a reply in the company’s tone with the latest status, and sets the draft ready. The planner reads it and clicks approve. Done in thirty seconds.

As a diagram: Client email → Agent reads and looks up in TMS → Draft for approval (human) → Reply sent. The highlighted block is the approval: that is where the person is, and that is where the trust is.

What an agent may and may not do

An agent gets exactly the permissions it needs, and no more. This is the list we record for each agent:

May:

• Read agreed sources (for example TMS status, a client file, a knowledge base).
• Draft something: an email, an order, a summary, a case file.
• Set up a task in a system and leave it on hold.
• Log what it did and why.

May not:

• Send an email to a client on its own.
• Book an order, change a price, or make a payment on its own.
• Pull data from outside the agreed sources, or process it outside the EU region.
• Make decisions about people (applicants, employees, credit).

As trust grows, you can relax the approval per task type: standard status emails automatic, for example, with unusual cases still going past the planner. But that is a choice you make later, with the numbers in hand.

How to keep the second pilot focused

The first attempt failed because it was too broad: “let’s start using AI.” The second attempt succeeds when it is narrow. This is how:

1. Pick one question or task that comes up dozens of times a week. Status questions, standard client questions, requests in a shared inbox.
2. Decide the source. Where is the right answer now? Usually in one system.
3. Decide the approval. Who looks at it before it goes out?
4. Build a prototype for that one task, on a business account in an EU region, with logging.
5. Measure for four weeks. How many drafts, how many approved without changes, how much time per case.
6. Only then decide on the next task.

This approach fits in 10 business days and asks about two hours a day from one process owner on your side in the first week.

Example

A logistics provider with 35 employees and five planners got 40 status questions a day, in between the planning work. After setting up a status agent with read-only access to the TMS and approval by the planner, 85% of status questions are answered within 15 minutes with an approved draft, and the planners spend 17 fewer hours a week on email. The figures are illustrative; the mechanism is exactly what is described above. Read the example case on the planning inbox and status requests.