What AI can do in an organisation depends on the files it gets to read

The AI models an organisation can use today are the same for everyone. What differs is what they get to read. In my own work, agents run every day that put an incoming question in the right file, write a first version of a text, or search a stack of documents. The quality of what comes out depends on the order in what goes in.

What a model gets to read determines the answer

A model given an ordered file per client answers questions after an hour as if it had worked there for years. The same model on three mailboxes, a shared drive and a spreadsheet with four versions answers just as confidently, and wrongly. A wrong answer that sounds convincing costs more than no answer.

An organisation with its files in one place is ready to start. An organisation with scattered data starts there, and not with the model.

Starting with the work that comes back every week

The first task for AI is work that costs time and asks little judgment: summarising and classifying incoming requests, a draft email for a quote, the weekly overview someone makes by hand every Monday. That work comes back every week, and it goes wrong on the days nobody has time for it.

If the automation works reliably there, the next task is added. If it does not work there, that is known before anything is built.

Limits on what the model sees

It is laid down in advance which data the model gets to see and which stays out, where that data is processed, and what the model's supplier may do with it. If personal data is involved, a data processing agreement belongs with it, and sensitive files wait until it is on paper. Those agreements also lay down who is responsible for what if something goes wrong.

The decision stays with the team

AI prepares. The team decides. A request the model classified is looked at by a person before an answer goes out. A text the model wrote is read before it leaves the building. Where liability is at stake, a person signs.

A chatbot that sits on a website apart from the operation is a gadget. Serious use of AI starts with order in your own data and ends with a person who keeps the last word. What sits in between is work the model takes over, one task at a time.