Where does intelligence belong?
· 6 min read · AI UX, Experience architecture, Product Design
Every AI brief I have seen in the last two years arrives in the same shape: add an assistant. A chat window, a sparkle icon, a box to type in. It is a reasonable request from a team that has watched the demos. It is also the wrong first question.
Treat AI as a feature and it behaves like one. Each team adds its own piece to its own part of the product: an AI summary here, an “Ask AI” button there, a recommendation box somewhere else, and a chatbot floating over everything. Each one behaves slightly differently. That is not intelligence. It is a product coming apart, with a gradient on top.
The brief says chatbot. The job is to decide where intelligence belongs.
So the real question is where AI should help, and how much. That question has only a few parts. My example is the AI-assisted data import I designed for KNIME, a data analytics platform. The brief was exactly this: help beginners get their data in, without taking control away from experts. The same rules work in any complex product.
1. Start from the things people work with
Every product has a small set of things its users work with and talk about. In a banking app, that is accounts, cards and payments. In KNIME, it is workflows, nodes (the individual steps of an analysis), datasets and columns. Designers call these objects. Users do not think about AI models or capabilities, and they should never have to. So my first question was not what the assistant should say. It was: what does the AI create that the product does not have yet? The answer was four things: the input file the user brings, the suggestion the AI makes, the draft step it fills in, and the explanation behind each choice.
Treating each of these as a real thing in the product, not as text in a chat, pays off right away. A suggestion can be pending, accepted, changed or rejected. That gives you undo, history, trust and analytics without extra work. A new kind of data source becomes a new kind of input file, not a new feature. And “explain this setting” works better than “ask the assistant”, because the setting already knows its context.
2. Use the smallest help that solves the problem
A conversation is the most demanding thing a product can ask of someone. Most needs are smaller. In KNIME, a beginner who drags in a spreadsheet file needs three things: the first rows of their own data, a suggested step with its settings, and an Apply button. No paragraphs of text. An expert who pastes a database address needs even less: a filled-in draft to skim, a couple of fields to adjust, and Apply. The same four things serve both, so nobody has to learn two systems.
The general rule: the home page shows what needs attention, search finds the right place, the item’s own page explains it, documents show the source, and chat handles what nothing else could. Chat comes last, not first. When it does open, it should already know what the user was looking at.
3. One consistent format for every AI proposal
Once a system can make proposals on its own, the hard part is not producing them. It is keeping them easy to read. If every AI feature designs its own card, the interface turns into a feed of whatever the latest feature said. In KNIME, every proposal is a suggestion with the same layout: what the AI saw, what it proposes, and an Apply button next to an Info button, every time. Suggested settings are checkboxes the user can untick, not magic. The same format covers a single setting and a whole drafted step.
The same rule works at a larger scale. On a big platform, many AI features can raise alerts on their own. Put one layer between them and the user. It checks whether an alert matters right now, removes duplicates, ranks what is left, and shows it in one shared format. Four types have been enough everywhere I have used this: something needs action, something needs your attention, something changed, and something could help. New features use those four types instead of inventing a new card.
4. Show where the answer comes from
A confident answer with nothing behind it is a risk wherever mistakes cost time or money. A disclaimer does not fix that. Showing the evidence does. In KNIME, the evidence is the user’s own data: the input file card shows the first rows of the real file before any suggestion appears. Each setting explains itself in a side panel: why it was chosen, what happens if you skip it, and when to pick something else. People can check the basis of a suggestion instead of being asked to trust it. In testing, that is where trust came from, more than from how well the AI wrote.
5. The AI proposes, the user decides
Some things a system should never decide alone: an unclear data structure, a connection it cannot verify, a setting with effects further down the workflow. The rule that made KNIME’s design work is simple: only the user applies. Nothing changes in the workflow until they click Apply, every decision is logged, and anything applied can be changed or removed later.
In products with human support, the same rule decides how a handover works. When a person takes over, they join the same conversation, with the item, the history and the evidence already there. The user does not start again. Handover is part of the design, not an exit from it, and it is where most of the trust in an AI product is won or lost.
What this means for the design system
None of this needed a separate AI design kit. KNIME already had an assistant panel, and the data import lives inside it, built from four components that share one layout: a part you can take in at a glance, supporting details, and actions. No new navigation and no new patterns to learn. The product still looks like itself. The intelligence is in how it behaves, not in how it looks.
AI does not need to look like AI to create value.
If you take one thing from this: before anyone draws a chat window, write down five things. The things your users work with, the places where help can appear, one format for AI proposals, how evidence is shown, and the rule that the user decides. That page is the product. The chat window is just one detail of it.
More writing
Assistant usage is not the metric. Resolution is.
Teams measure AI features by how often people use them. That rewards the wrong design. Five behaviours that tell you whether an intelligence layer is doing its job, and how to instrument them from the first release.
AI UXMetricsProduct DesignSmart results vs guided next steps
Some users need smarter results; others need guided next steps. Designing AI that adapts to a user's domain expertise – without becoming unpredictable.
AI UXProduct DesignUX