AI explained · system design for people who are not engineers
How AI systems at work actually work.
You do not need to code to make good decisions about AI. You need to know what happens between the question and the answer, and where people have to stay in charge. Here are the four systems you will meet most, drawn step by step. Click any step to see where it goes wrong.
01
An AI assistant, like ChatGPT or Copilot
The simplest system, and the one most people use. It feels like a search engine or a colleague. It is neither.
-
Whatever you paste may leave your organisation, depending on the tool and its settings. This is how employee data ends up in unapproved tools.
-
It sounds just as confident when it is wrong. Researchers call this the jagged frontier: tasks that look equally hard can sit inside or outside what AI does well.
-
Fluent is not the same as correct. Sent on unchecked, this is how workslop is born.
-
This step is the one that disappears under time pressure. Make it a rule, not a hope.
The lessonAn assistant is a fast first draft, not an answer. The human check is part of the system, not an extra.
02
AI connected to your company's documents
Many workplace tools now let the AI search your own files first, for example the HR policy library. The technical name is retrieval augmented generation.
-
People ask about personal situations. The answer has to be right for them, not just in general.
-
If the library holds an old policy, the AI finds the old policy. Outdated documents become wrong answers. Access rights matter too: can the AI see files this employee may not?
-
It can still mix things up or fill gaps with invented detail, especially when the passages do not fully answer the question.
-
Without sources nobody can check. Ask for this before you buy.
-
Automating these conversations saves minutes and costs trust.
The lessonAI with company data is only as good as the documents behind it. Cleaning up your policy library is part of the rollout.
03
An AI agent that takes actions
Agents do not just answer, they act: they plan steps and use tools such as email, calendars or HR systems. This is where the newest products are heading.
-
A vague goal leaves the agent to guess what you meant.
-
Small misunderstandings early in the plan grow with every step.
-
Every tool it can use is something it can also get wrong, at speed and at scale. Give it the fewest permissions that do the job.
-
Skipping approval to save time is how a wrong email reaches 50 candidates.
-
Keep a log, so you can see what the agent did and why.
The lessonDecide in advance which steps an agent may take alone and which need a human click. Start strict and loosen later.
04
An AI tool that ranks job applicants
Under the EU AI Act this is a high risk system. It shows best how a system can be technically correct and still unfair.
-
It learns the past, including its biases. Amazon stopped an experimental recruiting tool after it learned to downgrade CVs that contained the word “women's”, because past hires were mostly men (Reuters, 2018).
-
Candidates often do not know AI is involved. Telling them is a matter of fairness, and in the EU increasingly of law.
-
A score looks objective. It hides which features drove it, and some features, like gaps in a CV or a postcode, can stand in for age, gender or background.
-
When the list looks precise, people tend to follow it. Real oversight means looking below the cut line too, and being allowed to disagree.
-
Without this check, bias is invisible until someone complains, or a regulator asks.
The lessonNever let a ranking decide alone. Ask vendors what the model was trained on, how bias was tested, and how a person can override it.
05
Design an AI system that helps people
Every system above has the same five layers. Before you switch anything on, answer one question per layer. If one answer is missing, that is where the system will break.
Start with one boring, frequent task, not with the tool. Write down what good output looks like.
Sort data into green, amber and red before anyone pastes anything. Personal data is red until checked.
Name the person who reviews the output and the person who owns the decision. They can be the same, but never nobody.
Agree it upfront: focus, learning or going home. Otherwise it fills up with new work.
After a month, ask the people using it: what got better, what got worse, what did we not expect?
The lessonThe technology is usually the easy part. The questions that decide whether AI helps or harms are about people, data and time.
Connected topics
Sources
- Dastin, J. (2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters, 10 October 2018.
- Dell'Acqua, F., and colleagues (2023). Navigating the jagged technological frontier. Harvard Business School Working Paper 24-013. source
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex III, point 4, as amended by Regulation (EU) 2026/1744. source