AI that checks instead of guessing

Right now you mostly read two kinds of headlines about artificial intelligence. One says AI will soon take over almost all office work. The other says AI makes things up and is therefore useless.

The second headline describes a real problem. But it does not describe AI as such. It describes one particular way of using it. If you want to know whether AI can work reliably in your business, you need to understand that difference.

Why AI makes things up

A chat program has been trained on enormous amounts of text. When you ask it something, it answers from that memory. For well-known things this mostly works. But ask it about the autumn market in your own town, the brass band from the next village or a museum's new opening hours, and it simply does not know.

It is, however, built to always give an answer. So it fills the gap with something that sounds plausible. That is not bad intent, and it is not a flaw that will eventually be trained away. It follows from how the system is built. And that is exactly why it can be solved by how the system is built.

Four building blocks for AI that doesn't guess

A knowledge graph as an organised memory. Instead of relying on whatever a model once read, we store knowledge in an orderly way. Every place, person, event and work exists exactly once, with a fixed identity and links to everything that belongs to it: which piece runs in which venue, who performs in it, which town the venue is in. When a new detail comes in, it is compared with what is already known. Is this a new event or one we already know?

One place for all sources. Information arrives in many ways: from websites, in emails, through forms, through partner interfaces, as a scanned flyer or as a spreadsheet. With us, everything goes through the same check, with the same rules. No source may enter anything without passing that check. So no duplicates or contradictions appear just because something came in through a different route.

Its own research instead of guessing. When the system is not sure, it does not guess. It looks things up itself: on the venue's website, with the organiser, at the ticket office, in the local press. Only with this evidence does it make a decision. "Probably" is not enough.

Source grounding. Specialists call it "grounding". It means the AI only writes what it has found in checked sources, not what it believes it knows. Every decision is backed by the source it is based on, and can be checked. Only what has passed this check is published.

An example from our work

For CIXTRA, our platform for culture and leisure data, we see cases like this every day. A performance is listed on the theatre's website, in a cultural association's newsletter and at the ticket office. The title is spelled slightly differently each time, and the ticket office shows a time half an hour off.

A chat program would pick one of the three, or create all three as separate events. Our system recognises through the knowledge graph that this is the same piece in the same venue on the same evening. It checks the venue's own page, and the venue's information wins. The differing time is not deleted, it is noted.

A second example: a musician from the region who is not in any encyclopaedia. Anyone who only checks the big databases finds nothing and, in the worst case, invents a biography. Our system researches the web, her own website, organisers and the local press, and only writes what it actually found there.

A check in stages

Not every decision needs the same effort. That is why we check in stages.

Clear cases are handled by fixed rules, quickly and cheaply. Fuzzy cases are judged by our own recognition model, trained for exactly this task. Whatever is still open after that is checked again by a stronger model, with all the evidence and its own research.

A person only sees the small remainder that really cannot be resolved. For example when a cinema shows a film with an almost identical title in two screens at the same time, and no source says whether these are two screenings or a typo. The person is not the safety net for everything. That would save no time. The checking is done by the system itself.

What AI does well, and what it doesn't

AI does well today: reading, whether email, PDF, scan, handwritten form or website. Sorting and assigning. Comparing and finding duplicates, even when they are spelled differently. Translating and rewording. And doing all of it a hundred times a day without getting tired.

AI should not take responsibility for decisions about money, contracts or staff. Nor does it know your processes by itself. Which requests go to whom and which rules apply in your business, you have to teach it once. After that it follows them.

Questions to ask any provider

If you are thinking about using AI in your business, you can ask any provider, including us, a few simple questions:

• Does the AI answer from checked sources or from its memory?
• Does the system do its own research when a detail is missing or unclear?
• Is every decision backed by a source you can check?
• Does everything go through the same check, whichever way it comes in?
• What happens to the remainder that really stays unclear?
• What is my data used for, and who may access it?

If a provider has no clear answers to these questions, be careful.

How we do it

At SUPERFREY we build exactly this way: with a knowledge graph as memory, one shared check for all sources, its own research when unsure, and decisions that rest on sources. That is how CIXTRA works with culture and leisure data, and that is how we build solutions for businesses, municipalities and media, whether the work is requests, lists, documents or dates.

If you want to know which work in your business could be taken off your hands: write to us. The first conversation is without obligation and free of jargon.

Arthur Pichl, Co-founder, SUPERFREY FlexCo. Vienna, July 2026.