An AI Policy for a Manufacturing Company: What Worked and What to Do First

The starting point

At a manufacturer with two sites, AI in 2025 was everywhere and nowhere at the same time. Some employees used public chatbots for translations and emails, others were afraid of it, and management wanted to know what all this meant for the company. There was no rule, no overview and no shared language.

We turned that into a programme that ran for a good year. Here is what we did, what had an effect, and what we would do first today.

Kickoff: explain, do not sell

The start was a company-wide presentation: fundamentals, fields of application, potential. We deliberately did not design it as a sales event but as an explanation. Two live demos were decisive:

Least squares: an animation of a straight line gradually fitting itself to measurement points. That is not AI in the narrow sense, but it shows the basic principle – a model is fitted to data – in a way everyone in the hall understands.
A small neural network: trained live, with visible weights and visible error. After five minutes the audience had seen that “learning” is not magic but arithmetic.

These two demos created more understanding than all the slides combined. Anyone who has once seen a model learn has less fear and fewer inflated expectations.

The presentation was available as PowerPoint and as an interactive HTML version so it could be revisited afterwards.

The idea form

Right after the kickoff we released a form where anyone could submit an idea: which task costs time, what is repetitive, where the same thing gets looked up again and again. No obligation, no format requirements.

The result was more honest than any survey. The ideas did not come from IT but from logistics, sales, quality assurance. Many of them had nothing to do with AI – they were classic automation tasks. That was an important insight: the word “AI” opened the door, but the solutions behind it were often simpler.

The usage policy

In parallel we wrote a policy for AI use. We kept it short and reduced it to three questions:

Which data may go into which tools? Customer data, prices, drawings – not into public services.
Who checks the output? The person who uses it. AI text is a draft, not an approval.
What has to be labelled? Wherever AI-generated content goes outside.

The policy was read not as a ban but as permission with guardrails. That was intentional. A ban would only have pushed usage underground.

Requirements for an AI knowledge base

From the ideas one wish crystallised: an assistant that answers questions about internal documents – work instructions, machine documentation, processes. Before building or buying anything, we wrote a requirements list:

Which document sources, in which formats, how current
Who may see what – permissions must be inherited from the sources
Where the data runs – locally or in the cloud
How a wrong answer is detected and reported
Who maintains the sources

Alongside this we set up test forms for an assistant to capture users’ expectations concretely: which questions would you ask? What would a good answer look like?

The survey

At the end of the kickoff and again later we ran a short survey: prior knowledge, concerns, expectations, usage. It mattered less for the numbers than for the signals – for instance that concern about jobs was clearly greater in production than in administration. That influenced how we communicated later.

What we would do first

In hindsight we would change the order:

Policy first. It is quick to write and removes the uncertainty that stops people from experimenting at all.
Then the kickoff with demos. Understanding before ideas.
Then the idea form. And take the ideas seriously, even when they do not need AI.
The knowledge base last. It is the largest undertaking and needs the groundwork of the first three steps.

Takeaway

An AI programme in a manufacturing company is first a communication project and only then a technology project. The live demos created understanding, the idea form made the real problems visible, and the policy set the frame. The technology came afterwards – and was the smaller part.