When we remember a product, we don't remember its feature list. We remember how it felt. "It made things easier." "I never quite trusted it." People use features, but what they remember is the impression.
I recently read a LongBlack interview with Jeongyeon Yoo, CEO of ScentOn, a Korean scent marketing company. ScentOn designs the scent of spaces like Four Seasons Hotel Seoul, Olive Young N Seongsu and Charlotte Theater. It was supposedly about fragrance, but it read like an essay on AI products. Scent is invisible, everyone's taste is different, and it's hard to prove whether it's good. AI experience has the same problems. Outputs vary every time, and the moment a user thinks "this is actually good" is hard to show on a single screen.
Here are five lessons I took from ScentOn, and how they show up in the work of designing AI products.
1. Speak the user's language, not the machine's
When Yoo took over ScentOn in 2016, the first thing she noticed was the sales deck. Eighteen of its twenty pages described the diffuser. So she took the machine out of the pitch almost entirely.
"Do customers care about the machine? They shouldn't have to."
AI products fall into this trap easily. We lead with the model, the number of agent steps, how sophisticated the pipeline is. For the people building it, that's the part to be proud of. But users want to know one thing: "So how does my work change?"
The difference shows up in the smallest words. Call a feature a "filter" and it feels like a tool for manipulating data. Name it in the units your users actually think in, and it becomes a tool for choosing what they're curious about. The feature hasn't changed. The impression has.
The same goes for stacking AI capabilities. Under the hood, search, summarisation and generation might all be working together. But the user only needs to feel one thing, like "I don't have to switch to another tool." They don't need to know how many machines are running.
2. Let people experience it before you explain it
When ScentOn first pitched scent marketing, people would ask, "You mean the air freshener in the bathroom?" Even overseas case studies and research didn't always convince them. So ScentOn would install a diffuser free for a month.
"Once we take the machine away, they call us. When the scent disappears, that's when they finally notice the difference."
You can't explain a scent into someone's head. They have to smell it. New AI experiences work the same way.
Building AI products, you often hit a moment where nobody knows what a new kind of data or technology could become. That isn't a UI problem yet. It's an imagination problem. Strategy decks and static screens have surprisingly little power here. People leave comments on screens. They don't make decisions from them.
What has worked best for me is quickly building a clickable prototype on real data, so decision-makers can click around and ask their own questions. People who have touched something have a different conversation from people who have only heard about it. The question shifts from "Will this work?" to "How far should we take it?"
Static screens produce comments. Something people can actually touch produces decisions. You can't explain an impression. People have to live through it.
3. Judge by what's right, not what's liked
The hardest part of developing a scent is taste. The marketing manager might love one the CEO hates. ScentOn handles this the way you'd plan an ad campaign. Before creating anything, they define the concept and pull out the key words the scent should express. If the brief is "health," they narrow it down. Vitality? Restful sleep? Youth? Then they present two or three candidates side by side. Smell one on its own and people say "like" or "don't like." Compare several and the feedback gets specific: "Combine these two." "Adjust just this part."
"You might not personally love it, but it can still fit the brand direction we agreed on. That's how you get a standard for judging the scent."
AI design reviews turn into taste contests just as easily. And because AI outputs vary, people are often reacting to different results. So before a review, I try to get two things in place.
First, make sure everyone sees the same reality. Before designing, I get real model outputs and metric definitions from engineering. Screens drawn on clean, fake data look great, then fall apart the moment real outputs arrive. Nobody should be designing around imaginary data.
Second, don't force decisions you can't make yet. Sometimes the same information can be counted two ways, and each number tells a different story. Instead of picking one without evidence, it's often better to keep both and let real usage decide, the way you'd compare scent candidates side by side. And whatever we do decide, I write down at the end of every session. A decision that isn't written down gets made again.
When the criteria come first, a review stops asking "Which one is better?" and starts asking "Which one is closer to the experience we set out to create?"
4. Clear at first, subtle after
Choosing the scent isn't the end of the job. You still have to design when, where and how strongly people will notice it. At a hotel entrance, guests should clearly smell it. Further in, it stays subtle. Some hotels make it come back as you turn a corner.
"At first, the scent should make its presence felt and welcome the guest, enough for them to think, 'Oh, something smells nice.' After that, it should stay subtle so it doesn't get in the way of anything else."
In AI experience, this intensity matters enormously. AI has a lot to say: summaries, recommendations, new findings, all the time. Show everything at the same volume and users get tired fast, until they stop looking at all.
So I try to design what AI surfaces in layers of intensity:
The entrance: the first important change gets announced clearly, like a dismissible banner.
The hallway: after that, only a small marker stays in the list.
The room: "Why did this come up?" sits one layer deeper, for people who want to open it.
Welcome people clearly once at the entrance, then only remind them at the corners that matter. Just like the scent in a hotel lobby.
The same principle tells you where to put evidence. The source or original text behind a pattern the AI found means more right next to that pattern than collected somewhere else. It keeps the path from "what changed" to "why" one step long. A scent has to reach you at the right moment, and so does evidence.
5. One concept, different spaces
Olive Young N Seongsu's scent concept was "Beauty Playground." But each floor smells different. On floors one and two, where anyone can try products, it's fresh herbs and citrus. On floors four and five, where VIPs and partners meet, it's a soft, earthy scent. One concept, expressed differently depending on what the space is for.
ScentOn does this physically too. If a hotel lobby and restaurant share an air duct, the scent drifts everywhere. So their staff open the ceiling access panel and crawl through to find the duct that leads only to the lobby. They call it "zoning."
AI experiences pass through different spaces within one product as well. The screen where a first-time user explores. The dense analysis view full of numbers. The moment you share a result with someone else. The same AI needs a different voice in each: light and friendly when exploring, quiet and precise when analysing, careful when sharing.
This matters even more when you bring an AI experience into another team's product or design system. What helped most wasn't insisting on my own visual style. It was handing over real data, a working prototype and written decisions, then letting that team make it their own. When you walk into someone else's space, don't spray your own scent. Read how the air already moves.
6. An impression isn't built once. It's maintained.
One fact from the interview surprised me most. ScentOn makes more money from supplying and maintaining scent than from developing signature scents. When crowds grow, airflow changes. When a room is warmer, the same amount of scent feels stronger. So they check the machines and refill them, usually once a month.
"We don't stop at creating and installing the scent."
Impressions are especially hard to keep in AI products. Data changes, models change, and the same question can get a different answer. One good first impression doesn't settle anything. I think the most reliable way to keep it is honesty.
Say the AI finds a new pattern but there's no past data to compare it with. Rather than inventing a plausible-looking rate of change, it's better to say "There's no baseline to compare this to yet." When there's no evidence, don't fill the gap with a confident sentence. And let people know up front that running it again may give a different result.
Prototypes work the same way. They don't need to be perfect. They only need to be honest about the one thing they're proving. Show which parts are deliberately faked, and where the gaps are, as clearly as you show the vision. That's how everyone ends up debating the same reality.
The products people trust for a long time aren't the perfect ones. They're the ones that say when they don't know. And that honesty isn't something you ship once. Like refilling the scent every month, you have to keep checking it.
Closing: when the floor drops, impression becomes the ceiling
In my Config 2026 note, I wrote: "When AI lowers the floor, we raise the ceiling." Anyone can now build a decent AI feature quickly. Summaries, chat, agents. Features keep converging. And the more they do, the more users tell products apart not by what they can do, but by what they leave behind.
Near the end of the interview, Yoo describes how her clients' questions have changed. Ten years ago they asked, "How much will scent increase our sales?" Now more of them ask, "What scent feels like our brand?" Those of us building AI products might need to ask the same kind of question. Alongside "How much will this feature move the metric?", we should be asking: "What lingers after someone closes our product?"
Try this yourself
If you want to check the impression your product leaves, these questions might help:
Does your product speak the machine's language, or the user's?
Can people experience it instead of hearing it explained? Do you have a prototype running on real data?
Before reviewing designs, has your team agreed on what the "right" experience is?
Is everything the AI says shown at the same volume? Where are your entrances and corners?
Does your product admit what it doesn't know? Who checks that, and when?
Source: LongBlack, "ScentOn: Branding that lingers, from the scent consultant behind Four Seasons and Olive Young" (4 Nov 2026, Jiheon Kim). Quotes are translated from Korean.
