In this series, AI Start — Understanding AI without a technical background, we explore step by step what AI is, how it works in everyday work processes and how professionals can use it consciously and responsibly.
The conclusion in the previous part was simple: for someone using software, one question matters — is there AI inside the product? But that is not the whole story. “AI in the product” is not one single thing, and even if you know what the AI does, you do not yet know whether you can trust it. This part covers both.
See
Three steps
AI does not do the same thing in every piece of software. Broadly, there are three steps, each giving the system more room of its own.
- Interpreting. The software derives meaning from something raw — a photo, a text, a series of measurements. Unclear input becomes something useful. The decision remains yours.
- Deciding. The software draws a conclusion itself and acts on it. Not “here is the information”, but “I have done this”. A decision that used to be yours now rests with the system.
- Learning. The software adapts itself based on what it encounters. Its rules are no longer fixed. They change.
Each step gives the system more autonomy and moves something that previously rested with you.
Confidence is not correctness
A second layer sits apart from this ladder and makes things harder: AI sounds confident. A wrong answer can be just as fluent and convincing as a correct one. Its tone does not reveal doubt — at any of the three steps.
This is no coincidence. AI is built to produce plausible language and plausible outcomes. Whether they are correct is a separate question. The two often coincide, but not always — and when they diverge, the tone does not tell you.
Do
The same app, three steps, three ways to get it wrong
Back to the app that assesses swimming-pool water.
Interpreting. You photograph a test strip. The app recognises the colours and converts them into values — saving you manual entry. But in the evening sun, it misreads a colour and reports: “chlorine level fine”. As confidently as ever. Nothing reveals that the measurement is wrong. You check the input, and only notice if you question what you see.
Deciding. The app is connected to the dosing system. It measures, sees that chlorine is low and adds more itself — using a standard dose. Sounds logical. But your pool is in full sun and loses chlorine faster than average; the advice ignores that. Correct in general, wrong for your situation. Afterwards, you receive a message: “chlorine adjusted”. You no longer check the input, but the outcome: was its action right?
Learning. The app tracks the water throughout the season and adjusts its own target values. It neatly explains why — the explanation sounds conclusive. But it may be a tidy story wrapped around an outcome, rather than its true basis. The standard it uses in August is no longer May’s standard, and you no longer have a fixed benchmark to compare it with.
In all three cases, the presentation is perfect. The substance need not be. And the presentation alone does not show the difference.
Own
What shifts, and what you need to check
With each step, responsibility shifts from you to the system — and at each step, trust the basis, not the tone.
With interpreting, you stay in control. The app provides information; you decide. If something goes wrong, you see it when making your own decision — provided you do not blindly accept what the app supplies.
With deciding, you hand over some control. The question becomes: can you see what the app based its decision on? An app that adds chemicals without showing why asks for blind trust.
With learning, even the standard shifts. This is the hardest step: you can no longer assess the app against a fixed standard, because the standard changes. The question is then not only “is the outcome right?” but “is its basis still right?”
At all three steps, the same two questions help calibrate your trust:
- How serious would it be if this were wrong?
- Can you check it independently — with a second measurement, your own eyes or someone who knows?
If a check is inexpensive, do it, especially for decisions that matter. The higher the app sits on the ladder, the less the outcome alone tells you, and the more important those two questions become.
Summary
“AI in the product” has three steps — interpreting, deciding and learning — and each shifts responsibility from you towards the system. What makes that shift difficult is that AI can sound equally confident at all three steps, even when wrong. A confident tone therefore says nothing about correctness. The skill you need is not distrusting AI, but calibrating trust: the higher the step and the more serious the consequences of an error, the more checking an answer deserves — and the less sufficient it is simply to ask whether it sounds convincing.