AI integration

Drafting text

A quotation, a rejection, a reply to a complaint: the draft appears in seconds, and it is checked and signed as before. The time is not saved on typing — it is saved because nobody sits in front of an empty page.

Searching your own documents

Old quotations, assembly instructions, standards, test records. The question is asked in plain sentences and the answer comes with the source. Without the source the answer is worthless — we build for that.

Dictating instead of typing

The site manager speaks the report into the phone and the text arrives sorted in the system. On site, wearing gloves, that is the difference between recorded and not recorded.

Paper into figures

Reading delivery notes and incoming invoices, taking over items and amounts. Saves the retyping and the transposed digits that come with it.

Sorting the post

Recognising incoming enquiries and routing them to the right person. Useful from the point where enquiries sit unanswered because nobody feels responsible.

Translating

Instructions, safety briefings, correspondence. Good enough for internal use; not for contracts.

• When the data is wrong. A model invents the order that is missing from the data — and words it convincingly.
• When the task comes up twice a month. Setting it up and looking after it costs more than the saving.
• Forecasts from a handful of numbers. The workload of a twenty-person company is not statistics, it is a handful of dates.
• Decisions that have to be justified — prices, staff, dismissals. Leave those to a machine and you have nothing in your hand in court.
• Anything where a mistake goes unnoticed. If nobody checks the result, using it is dangerous, not economical.
• A chatbot on the website, as long as the telephone goes unanswered.

Collecting and sorting the use cases

Everything that comes up often and involves text, speech or images. Sorted by frequency times duration times the cost of errors. What ends up on top goes into the pilot — usually one thing, at most two.

Settling data protection first

Which data would leave the building, to whom, on what legal basis. That question is answered before the tool is chosen, not after.

A measured pilot

Four weeks, one department. Before it starts we measure how long the task takes, and afterwards we measure again. Without those two figures anything can be claimed later.

Building in a check

Who checks the result, how do they recognise a mistake, what happens when they find one. A model that is occasionally wrong is usable; one whose mistakes nobody notices is not.

Instruction

People need to know when not to believe the thing. That is not a courtesy — since the EU AI Act it is an obligation.

Public service

Fast and cheap. Without a data processing agreement no customer data, no staff data and no costings belong there. For general drafts with no link to a person it is enough.

A provider with a contract and servers in the EU

Data processing agreement, an assured storage location, no use of your input for training. The usual route for anything that touches real business.

A model in your own building

Open models run on a machine at your site; nothing leaves the network. Needs a suitable graphics card and somebody to look after it. Worth it when documents are constantly involved that are nobody else's business.

What belongs nowhere

Health data, job applications, costings, drawings, anything under a confidentiality agreement. Not because of the technology, but because a breach costs more than any time saved.

Your people have to be able to handle it

Anyone using AI in a company must make sure the staff understand what the tool can do and where it goes wrong. An hour of instruction with examples does it; a circular does not.

It has to be recognisable

Anyone writing to a machine must be able to tell. This mainly concerns chat functions on the website and automatic replies.

Not every application is harmless

Most applications in the trades count as low risk. Two areas do not: selecting applicants, and monitoring the performance or conduct of staff. Stricter rules apply there, and the works council (Betriebsrat) has to be involved.

The liability stays with you

For a wrong quotation, a wrong statement, a wrong dimension, you are liable — not the provider of the model. "The AI said so" is not a defence.

This is not legal advice. We tell you at which point you need a lawyer, and write down what they need to know.

Of what is currently written about AI, the smaller part is usable for a company with ten to fifty employees. Sorting that out is our job, not yours.

• What almost always works: searching your own documents, dictating instead of typing, drafting text.
• What rarely works: training your own models, forecasts, a chatbot for customers.
• What we make of it: one or two use cases, a four-week pilot, and afterwards a figure instead of an opinion.

If there is no saving at the end of the pilot, we say so and build nothing further. A tool that only saves in the brochure costs twice over in the workshop.