Somewhere on your roadmap there is an AI feature with a chat box in the mockup. Take it out. The chat box is almost never the product decision. It is what gets shipped when the product decision has not been made yet, and it quietly moves the hardest part of the job, knowing what to ask for, onto the person paying you. The AI features that earn daily use hide the model inside an interface people already understand and never make the user write a prompt. That one call decides whether your feature shows up in the usage numbers next quarter or just demos well in the all-hands.
I am not against conversation on principle, and I shipped one. There is an AI version of me on this site that answers questions about my work, and I will make the case for it near the end, including the part where almost nobody has to type the first message. That is an exception I built deliberately. The default I keep walking into is the opposite: a text field bolted onto a product, pointed at a model, presented to a board as the AI strategy. That is a design failure wearing a pleasant name.
A blank text field is an unfinished product decision
Every other input in your product carries a job. A date picker knows it wants a date. A search field knows what corpus it searches. A chat box is the one field we are willing to ship with no job at all, label it ask me anything, and count as a feature. It is seductive for an honest reason: it is the fastest thing to build. The model already speaks. Wire an input to it and you have an AI feature by Friday.
The cost does not land on your team. It lands on the user, as a series of small taxes. Figuring out what this thing can even do. Guessing at phrasing. Discovering the limits by hitting them. Re-asking in a different voice because the first answer missed. None of that work existed before you shipped, and all of it gets paid by the person you were trying to help.
Jakob Nielsen called generative AI the first new user interface paradigm in 60 years (opens in new tab), and the shift he describes is real. Instead of telling the computer each step, the user states the outcome they want and the system works out the steps. He calls it intent-based outcome specification. Read further in the same piece, though, and he says current AI tools "have deep-rooted usability problems," then puts a sharper number on it: "Based on recent literacy research, I deem it likely that half the population in rich countries is not articulate enough to get good results from one of the current AI bots." The paradigm is a genuine change. The chat box is the crudest possible expression of it, and it is the expression everybody copied.
A chat box is a blank field sitting where a product decision should have been.
The first AI feature I shipped had no prompt box, and that was the design
At Supply Drop I designed the brand, the app, and the web experience for a business built entirely on a prediction. The model watched what a household actually went through and predicted when they would run out of essentials, then shipped the next box before they did. That prediction was not a side feature. It was the whole reason to subscribe. A customer never once described what they wanted to it.
Signup was a short quiz. After that the product did the talking: here is what your home needs, here is when. The hard part was never conversation, it was belief. Nobody hands a recurring charge to a prediction they do not understand, so I built a page called The Science of Supply Drop that explained the engine in plain words, and designed the product to surface what each home needed and when, so the prediction was visible before the box arrived instead of after. I also handed our ML engineers the customer interviews and usage data that made the model better week over week. When I took the paid social launch over end to end, it drove more than a hundred paying subscribers in the first week, recurring revenue rather than one-off sales.
That was a prediction model, not a language model, and I want that to land rather than slide past. The interface question is older than the chatbot. We had a model doing something genuinely useful, and the two real design jobs were deciding what it should do without being asked, and making its reasoning legible enough to trust. A chat box solves neither. Ask our prediction engine anything is a much worse product than paper towels that arrive the week before you run out.
The most generative product I own has no prompt box in it
I founded Story Genie, where AI agents write and illustrate personalized hardcover children’s books. Generative text, generative illustration, an agent pipeline running the entire production. If any product on earth had an excuse for a prompt field, it is that one. There is not one. You answer a few short questions about a child, and about sixty seconds later you see the whole finished book, every page, before you pay a cent.
A prompt box would have been faster to build and worse in three directions at once. It would have made the quality of the book the parent’s fault, because a thin prompt produces a thin story and the customer would never know that was the reason. It would have asked someone buying a gift at eleven at night to be good at describing a child in writing, which is a strange skill to require of a person with a credit card out. And it would have broken the single decision the company rests on, which is that you see the finished book before you pay. I can only put a finished thing in front of you if I control the inputs well enough to make it good every time. Hand the input to the customer and the preview stops being a promise I can keep.
I wrote up how that trust model works in the piece on keeping humans in charge of an agentic product. The part that matters here is a pattern I keep hitting: every time I removed something the customer had to produce before they got value, the product got better. I assumed people would create an account to see their preview, so the account came first. It worked the other way around. People want proof before they give up an email. Now the preview comes first and the account is framed as what it actually is for them, saving their story. A prompt box is that same mistake one level deeper. It asks the customer to produce the hardest input there is, a good description of what they want, before you have shown them anything at all.
The best AI interface of the last five years looks like autocomplete
Look at the most successful AI feature in working software. GitHub ran a controlled experiment with 95 professional developers and found that the group using Copilot completed the task 55 percent faster (opens in new tab), one hour and eleven minutes against two hours and forty-one, writing an HTTP server in JavaScript. The interface that produced that number was not a chat panel. It was grey text appearing in the editor, at the cursor, in the exact file the developer already had open. No context switch, no prompt. Their code was the prompt.
That is the move, and it generalizes. Find the place in the workflow where your user is already expressing intent, and attach the model there. The search they typed. The file they just uploaded. The row they selected. The draft they are halfway through. The filters they already set. Those are prompts. They are prompts the user was going to produce anyway as part of doing their job, which makes them free.
I built this shape years before I had a model to put behind it. At Enverus I designed a Deal Finder for NAPE, the largest show in oil and gas, and led the agency that built it. A buyer entered their budget and what they were after. It matched them against the vendors on that floor selling exactly that, then mapped where to find them. It surfaced deals worth millions during the show. Two inputs and a map, hand built, no AI anywhere in it. If I rebuilt it today I would put a model underneath and change almost nothing above it, because the interface was never the limitation. The matching was.
Your product is already full of prompts the user was going to write anyway. Use those instead of asking for a new one.
Four shapes to reach for before you reach for chat
This is the order I work through, and I do not get to the fourth one often.
First, the default that is already correct. The model fills the answer in and the user edits it. A category pre-selected, a title pre-written, a form arriving two-thirds complete, a reply drafted and waiting. Correcting a proposal is a fraction of the work of composing a request, and it has a quieter benefit: the user sees what the model is capable of without having to guess at it, because the capability is sitting right there on the screen already filled in.
Second, the suggestion in place. Inline completion, a ranked shortlist, three options to choose between. Nielsen Norman Group studied this directly, analyzing more than a hundred Midjourney prompts, and named the problem the articulation barrier (opens in new tab): "many users may be able to visualize images in their minds but lack the vocabulary to write the required prompt." Their recommendation is a hybrid interface, prompt suggestions plus a gallery of visual styles, so someone can pick the Jackson Pollock thumbnail instead of needing to know Jackson Pollock. Choosing beats describing, for everyone, every time.
Third, the structured input. A form, a quiz, a set of filters, a selection on a canvas. Teams skip this one because it feels less impressive in a demo, and it is the one that performs. Supply Drop’s signup quiz and Story Genie’s handful of questions about a child are the same shape. You already know the five things the model needs to do good work. Ask for those five things, in plain language, the way you would ask for a shipping address.
Fourth, no interface at all. Some of the best AI work produces nothing for anyone to look at. At the apparel platform where I lead design, writing a launch catalog used to take two writers about a month. I built an agent that brought it to under an hour. The win was not that the writers got a smarter assistant to talk to. The win was that a step came off the calendar. If an agent can do the work against your real systems, do not build it a window. I went through that plumbing in what I actually do with AI as a design leader.
Who each call is for
If you are a product or engineering leader adding AI to a product that already has users, the answer is almost certainly not a chat box. You have the thing a chat box throws away: a workflow with real steps in it, and a user who told you what they are trying to do by being on that screen at all. Attach the model to one step. Measure whether that step got faster, cheaper, or better. That is a result you can defend in a quarterly review, because it moves a number the business already tracks.
If you are building something net new and AI-native, your job is to design the input, not delegate it. Name the fields the model needs. Write them down. If you cannot name them, you do not understand the job yet, and a chat box will hide that from you for roughly two quarters while you tell yourself the model is learning. I use a harder filter before I build anything with AI at all, which I laid out in how I evaluate an AI product idea.
And if the real problem is that nobody in the room can make this call, that is the fork worth naming. Someone has to sit between the model and the customer and decide what the thing does without being asked, which is a judgment problem, not a modeling problem. That is the work I do as an AI design consultant: finding the step in your existing workflow where the model belongs, designing the input so the user never has to perform, and putting a shipped version in front of real people fast enough to learn something. If your team is three weeks into arguing about a chat panel, the argument is the symptom.
When the answer really is a chatbot
There is a chat box on this site, so let me defend it. It exists because the question space is genuinely unbounded. I cannot predict what a hiring manager wants to know about a career’s worth of projects across very different companies, and no navigation I design will ever enumerate it. That is the first honest case for chat: when the useful questions are too many and too varied to put on a screen.
Look at how it opens, though. Nearly every entry point into it is a button that sends the first message for the visitor. On a case study the button says ask about this project, and then it asks. On the about page there is a button for a sixty-second tour. On the 404 page the button builds the question out of the path you tried to reach, so the first thing the assistant answers is where you were actually headed. Chat is the room. Almost nobody has to walk in with a line prepared.
The second case is expert users doing genuinely open-ended work. The articulation barrier is much lower when articulating precisely is the skill the person was hired for. Developers, analysts, lawyers, writers, researchers. They describe things for a living, their vocabulary matches the domain, and a structured interface would only get in their way. Note that even Copilot serves both: inline completion for the common path, a chat panel for the strange one.
The third case is when the conversation is the product rather than the delivery mechanism. Practicing a hard discussion, tutoring, a language partner, rehearsing an interview. The back and forth is the value. Putting a form in front of that would be the same mistake in reverse.
And a caution against the purist version of my own argument. The NN/g recommendation was hybrid, not either or, and hybrid is usually right. Build the structured interface for the eighty percent of cases you can name, and leave a text field for the rest. The failure is not having a text field. The failure is shipping the text field instead of doing the naming, then calling the blank box flexibility.
Here is the test to run in your next review. Write down, word for word, the first message you expect a user to send. If you can write it, build it as a button. If you can write ten, build ten, and you have a product instead of a prompt. If you genuinely cannot produce a single one, that is real information and you may have a chat problem. Most teams find they can write all ten in about four minutes, which tells you what the chat box actually was. Not an interface. The place the thinking stopped.