Artificial Intelligence

AI accelerates a well-defined task; it does not define an undefined one. Most of the failures we see in projects stem not from model choice but from the task never being clear in the first place.

So in the first meeting we talk not about models but about which job improves, and by what measure.

Where it works

• Document processing — extracting structured data from invoices, waybills, forms and contracts
• Request routing — classifying incoming messages and requests to the right team
• Customer communication — first responses to frequently repeated questions, alongside the bot side
• Drafting — first drafts of proposals, descriptions and standard texts
• Search and summarization — meaningful search across your own document archive

Where it does not

This needs saying too. Processes that produce binding decisions on their own, are hard to reverse, and whose errors reach the customer directly are not suitable for a first project. Likewise, in an area where data is scattered and inconsistent, AI does not fix that scattering — it builds on it and amplifies the error.

In that situation we recommend fixing the data side first. That shrinks the work on offer, but it is the correct order.

Human approval sits inside the architecture

In the first version, the system's output does not go straight to the customer. A human approval step sits in between; it lowers the cost of errors and shows you where the model gets things wrong. Removing an approval step later is easy — recovering from never having had one is not.

The legal framework is set upfront

In AI projects the legal side is part of the design, not a clause added at the end:

• If the content fed to the model contains personal data, that processing falls under data protection law — does your privacy notice cover it?
• If your provider's servers sit abroad, there is a cross-border transfer, requiring its own legal basis
• Does the NDA you signed with your client permit sending that client's data to a third-party service?
• Who bears liability when the system produces a wrong output? This ties directly to where human approval sits

These questions usually end not in "you cannot" but in "you can, under these conditions." Knowing the conditions upfront is cheaper than discovering them after the build.

We compare those conditions technically against what the system actually does; we describe how in a separate note. Legal assessment of the conditions belongs to your own lawyer.

How we work

We start with a narrow pilot and write the success measure in advance. "Improve efficiency" is not measurable; "cut first response time from 4 hours to 15 minutes" is.

If the pilot holds, we expand it. If it does not, we report why. That is also a result, and learning it early is cheap.

For the wider picture, see our article on AI for business.