What AI actually does when it does something
“AI handles customer service” tells you nothing. Here are 11 situations taken apart: what the machine computes, what is left to people, and where it typically goes wrong.
Note that only 3 of the 11 involve a language model at all. Most of the AI in use never talks to anyone.
Work and everyday life
In these, AI handles text, images or transaction records. A mistake means a wrong answer that somebody has time to correct.
01 · Online retail
Customer service answers a question about return conditions
A customer asks in chat whether an opened product can be returned. The answer comes in seconds, with a link to the right clause.
What kind of AI: A language model with a search over the company's own documentation behind it.
What the machine does
- Finds the passages closest to the question in the company's own guidance texts.
- Writes the passages it found into a single answer shaped to the customer's question.
- Attaches the source the information came from.
- Hands the conversation to a person when the question does not match the guidance.
What the person does
- Writes and keeps up to date the guidance the search draws on.
- Decides which matters the chat may finish on its own and which always go to a person.
- Goes through the cases where the customer was not satisfied with the answer.
02 · Finance
Details are lifted off a purchase invoice into the accounts
A PDF invoice arrives by email. The system reads the amounts and dates off it and proposes how to post it.
What kind of AI: No conversational AI at all. Image recognition and a small classification model trained for one narrow task.
What the machine does
- Turns the image of the invoice into text.
- Identifies the fields in that text: supplier, invoice number, due date, amount, VAT.
- Proposes a posting based on how comparable invoices were posted before.
- Marks the uncertain parts for checking instead of guessing.
What the person does
- Approves or corrects the proposal, and the model learns from the corrections for next time.
- Handles the exceptions: a new supplier, an odd invoice layout, a disputed invoice.
03 · Everyday life
A service suggests what to watch next
The front page shows a list assembled for you. This is probably the AI you use every day without noticing.
What kind of AI: A recommendation model, not a language model. It does not understand the content; it spots regularities in what similar users have watched.
What the machine does
- Compares your viewing history with the histories of millions of others.
- Finds users whose choices resemble yours.
- Suggests what they watched and you have not.
- Measures continuously whether the suggestion was clicked, and adjusts the next list accordingly.
What the person does
- Decides what is being optimised: watch time, satisfaction, or discovering something new. These lead to different results.
- Sets the limits on what may not be recommended.
04 · Office work
A long report is turned into a summary for the board
A thirty-page review has to come down to one page. The model does it in a minute.
What kind of AI: A language model directly, with no separate retrieval. Everything it needs is given to it at once.
What the machine does
- Reads the text as tokens and produces new text that is a typical summary of this particular text.
- Adjusts the style as asked: shorter, more formal, bullet points.
- Renders the same content in another language if asked.
What the person does
- Chooses what matters, because the model summarises what recurs, not what is important.
- Checks the figures and the names. Those are exactly the places where fluent text can be wrong.
- Answers for the result. Responsibility does not transfer to the machine.
05 · Software development
An assistant fixes a bug and runs the tests
A developer describes a fault. The assistant finds the place in the code, makes the change, runs the tests and reports whether they passed.
What kind of AI: An agent: a language model given tools (reading files, editing them, running tests) and permission to use them in a loop.
What the machine does
- Finds the part of the code that matches the description.
- Proposes a change and writes it to the file.
- Runs the tests and reads the result.
- If the tests fail, corrects and tries again. That repetition is what makes it an agent rather than just an answer.
What the person does
- Decides what may be changed and what may not, and reads the change before approving it.
- Sets the limits: how many attempts, which tools, what it has no access to.
- Answers for what goes into production.
The physical world
In these, the machine moves, grips or measures. A mistake is not a wrong sentence but a dropped object, a wrong route or a fault that went unnoticed, and there is no undo button.
01 · Industry
The machine says a bearing will fail next week
Sensors on a production line send measurements continuously. The system notices a change in the vibration before a person hears anything.
What kind of AI: Not a language model at all. A prediction model trained on measurement data, of a kind industry was using long before ChatGPT.
What the machine does
- Compares the current vibration and temperature data with what the data looked like before previous failures.
- Calculates the probability that the equipment fails within a given period.
- Raises an alert when that probability crosses a threshold.
What the person does
- Sets the alert threshold: too sensitive and it cries wolf, too blunt and it is late.
- Goes and looks at the equipment and makes the maintenance decision.
- Records the failures that actually happened, so the model can be measured and improved.
03 · Logistics
A robot picks one item out of a heap
A crate holds a jumble of products. The robot lifts the right one, even though it has never seen this particular heap.
What kind of AI: A vision model and a grasping model, trained largely in simulation before the robot touches anything real.
What the machine does
- Separates individual objects out of an overlapping heap in the camera image.
- Judges where on the object to grip so the hold will keep.
- Adjusts the grip during the movement: if the object slips, the hold changes.
- Tries again from a different angle if the first attempt fails.
What the person does
- Defines what may be handled and with how much force.
- Handles the exceptions: a broken package, the wrong product, a collapsed heap.
- Answers for safety at work when people are in the same space.
Security and monitoring
In these, the machine detects and classifies but does not decide. Its job is to cut thousands of observations down to the few a person should look at. The decision, and the responsibility, stay with the person.
01 · National security
A vessel switches off its identification at sea
Hundreds of vessels are moving in a sea area. The system surfaces the few whose movement does not match what they say they are doing.
What kind of AI: Not a language model. Fusing several observation sources and detecting anomalies.
What the machine does
- Combines radar observations, vessels' own identification signals and satellite imagery into one picture.
- Links observations of the same vessel from different sources into a single track.
- Compares each vessel's movement with how comparable vessels usually move in the same place.
- Surfaces the anomalies: an identification signal going dark, a route weaving for no reason, a vessel stopping where there is no reason to stop.
What the person does
- Decides what counts as an anomaly: the same movement means something different for a fishing boat and a cargo ship.
- Confirms the alert against other sources before drawing a conclusion.
- Makes the decision and answers for it. A machine does not make an official decision.
02 · Emergency response
An uncrewed aircraft searches woodland for a missing person
The search area is hundreds of hectares and darkness is coming. The aircraft films the terrain with a thermal camera and marks the spots worth checking.
What kind of AI: An image recognition model running on the thermal and ordinary camera feeds. No conversation and no reasoning, pattern recognition and nothing else.
What the machine does
- Flies a predetermined search pattern and films the terrain continuously.
- Goes through thousands of frames and marks the ones with a person-sized heat source or an unusual shape.
- Screens out known false positives: animals, sun-warmed rocks, vehicles.
- Presents a person with a list of places to check, not a decision about who or what is there.
What the person does
- Bounds the search area and decides the altitude and the search pattern.
- Looks at the marked spots and decides where to send a patrol.
- Runs the search and answers for it.
03 · Surveillance
A camera detects a person, but identifying one is a different matter
Something is moving at night inside a fenced site. The system tells a person from a deer and alerts only on the first.
What kind of AI: Image recognition. Detection (“there is a person in the frame”) and identification (“this particular person is in the frame”) are different tasks, and different rules apply to them.
What the machine does
- Detects movement and tells a person from an animal, a vehicle and a branch swaying.
- Tracks the detected subject from frame to frame, so the same passer-by is not counted many times.
- Marks the anomalies: movement in an area nobody should enter, travel in the wrong direction.
- Screens out the great majority of the pointless alerts a plain motion sensor would produce.
What the person does
- Decides what is filmed, how long recordings are kept and who may look at them.
- Looks at the alerts and makes the decision.
- Answers for the use being lawful. The EU AI Act sharply restricts biometric identification in public spaces, and identifying a person is legally a wholly different matter from detecting one.
What all 11 have in common
The machine does a narrow, measurable task. A person decides what is being aimed at, sets the limits and answers for the result. When an AI project fails, the cause is rarely the model and more often that those three were never settled.
The more physical and the more critical the use, the more one question matters: what happens when the machine is uncertain? In a well-designed system the uncertainty is visible and leads to a person. In a badly designed one it is not visible at all.
More about the application layer →