The question comes up in every second conversation: “Which AI model do you recommend?”
The honest answer is uncomfortable: that is the wrong question. And it comes two steps too early.
The real problem sits in the folder, not in the model
After twenty years, a mid-sized development lab for sealants and adhesives has typically run a few thousand batches. That is a treasure. It just sits like this:
- Formulations in Excel – every file built a little differently, grown across three generations of staff.
- Measurements in the lab notebook, handwritten, partly scanned as PDF.
- Raw material data in the ERP, under article numbers nobody links to the formulation.
- And the most important part: in people’s heads. “Watch out with that filler, it pulls moisture.” That is written down nowhere. It retires when the colleague retires.
No model in the world learns from a collection of inconsistent Excel files. Not because it is too stupid – but because those files lack the context that turns data into knowledge in the first place. What was the question? Which batch? Which climate? Which trial was the predecessor?
A data point without context is not information. It is a number.
What happens once data is structured
As soon as formulation, raw material, process and measurement hang together in the same system, the picture flips. Not because magic sets in, but because mundane things suddenly work:
- Search instead of remember. “Did we ever have a system with this filler and over 400 % elongation at break?” – a query instead of a week in the archive.
- See relationships. Multi-dimensional analysis shows interactions that stay invisible when each factor is viewed alone. Two variables together behave differently than either one on its own.
- Predict instead of guess. With enough structured history, a model estimates the outcome of a batch before it is run. Not perfectly – but well enough to sort out the hopeless trials beforehand.
The last point is the lever. AI does not find the one brilliant experiment. It saves the ninety you could have spared yourself.
The platform we work with reports figures of up to 70 % fewer experiments and up to 5× faster product development. Those are the platform provider’s numbers, not a measurement from your own house – and that is exactly why the sequence that follows matters so much.
What AI in the lab does not do
So that expectations are right, this belongs on the same table:
- It does not replace the chemist. A model has no sense of plausibility. It will happily propose combinations any formulator discards within two seconds. The judgement stays with the human.
- It does not replace testing. A prediction is not a test report. For CE marking and the Declaration of Performance, the measured test to the standard counts – nothing else.
- It does not repair bad data. If the input is dirty, so are the predictions. Only faster, and with more decimal places.
- It is not a one-quarter project. The effort sits at the front, in structuring. The benefit comes afterwards.
Where the entry point actually is
Not with the software. With three questions:
- Which data do you really have? Not which you believe you have. The difference is regularly sobering – and the most honest starting point.
- Which question should be answered? “We want to do something with AI” is not a goal. “We want to substitute the plasticiser without losing low-temperature flexibility” is one.
- What happens when the colleague leaves? If the answer is “then we no longer know”, you do not have an AI problem. You have a knowledge problem – and that is where structure pays off immediately.
Anyone who can answer these three questions needs no grand decision for the next step. The rest is craft.
The uncomfortable core
Digitalisation in the lab rarely fails on the technology. It fails because nobody gets the time to bring yesterday’s data into a shape that still carries tomorrow.
That is not an IT task. That is a lab task – and it needs someone who speaks both languages.