
Planning an AI feature people will actually use
“Our competitors all have AI now. We need some too.” If you haven't sat in a meeting where someone said that yet, give it time. Boards want AI on the roadmap, investors keep asking about it and a competitor seems to announce something new every week. So teams end up with a shiny piece of technology attached to a project and no actual problem for it to solve. The usual response is a chat bubble in the corner of the screen. I get the appeal. It's easy to ship, it demos beautifully in a board meeting and nobody could mistake it for anything other than AI. Unfortunately, it also sits outside the product, knows next to nothing about what the user is doing and just waits for them to think of a question. It's a bit like a waiter who hovers by the kitchen door and waits for you to wave. And users are not exactly clamoring for more of it. In a June 2025 Pew Research survey, half of US adults said the growing use of AI in daily life makes them more concerned than excited. Back in 2021, that figure was 37%. More AI has not made people any fonder of it. People don't want more AI. They want AI that works. I have been preparing a workshop on designing AI-powered products for Master.dev , and much of it leans on Josh Clark and Veronika Kindred's excellent book Sentient Design . You'll see their names a lot below, because they did a lot of the hard thinking. But the bit I keep coming back to is the planning. If you get that wrong, no amount of interface polish will save you. So here's how I'd get from “we need AI” to a feature that's worth building. To keep this grounded, I'll use the made-up example from the workshop. Ledgerline is an expenses app for small teams. Employees log receipts, managers approve claims and finance pays everyone back at the end of the month. Nobody enjoys any part of this, which makes it rather good material. Start with where people get stuck Most teams start by asking “where can we use AI?” The better question is “where are people getting stuck?” The first question gets you features that show off the technology, and then users have to work out why they should care. The second gets you features that quietly remove hassle. People notice the result rather than the clever stuff behind it. When I go looking for friction, I find it usually comes in four flavors: Too slow. The task takes longer than it should. Too repetitive. People do the same dull steps over and over. Too confusing. People don't know what to do or what the rules are. Too much. There are too many options or too much information to wade through. The good news is that you probably have the evidence already. Dig through your support tickets for the questions people keep asking. Check your analytics for where people drop out or slow down, and watch a few session recordings. Ask people which tasks they put off or quietly work around, and which mistakes somebody keeps fixing by hand. When we did this for Ledgerline, users said things like “ I never know which category to pick ,” “ My claims get rejected and nobody tells me why ” and “ Writing the business purpose for every coffee is soul-destroying. ” Not one of them asked for a chatbot. Funny, that. If you can't name the friction, don't build the feature. Make sure it's actually a job for AI Not every bit of friction needs AI. There is a useful line between tasks with one right answer and tasks with lots of acceptable ones. A tax calculation has one right answer. So does a bank routing number, or a currency conversion at the official rate. There's nothing to interpret and mistakes are serious, so that's a job for ordinary rule-based code in most cases. Boring is a compliment here. A sentence, a summary, a category or a suggestion is different. Plenty of answers are good enough, and a person can easily tweak them. That's where AI earns its keep. So in Ledgerline, reimbursement totals and sales tax stay firmly with the rules. Suggesting a category, drafting the business purpose and spotting unusual spending are fair game for AI. The value versus risk grid Next, I plot each idea on two axes. How valuable is it, and how forgiving can you afford to be when it gets things wrong? High value, forgiving of mistakes. This is the sweet spot, because mistakes are cheap and easy to fix. High value, unforgiving of mistakes. It's worth doing, but a person makes the final call, so design for review. Low value, forgiving of mistakes. This is the fun zone, where delight matters more than return. Low value, unforgiving of mistakes. This is the “why bother” zone, so keep it simple and rule-based. One idea on Ledgerline's list was letting AI approve small claims automatically. The friction is real enough, because managers admit they approve claims on their phones between meetings without really reading them. But approvals involve money and accountability, which puts the idea firmly in the high value, unforgiving corner. AI can summarize the claim, but a manager still decides. Five awkward questions Before committing, I also run through five questions adapted from Sentient Design: How critical is absolute accuracy for this task? How accurate is AI at doing it today? Is there something more accurate, including a person? Would a person and AI working together do better than either on their own? Have you actually tested any of these assumptions? That last one tends to produce an awkward silence in meetings, which is rather the point. Work out what kind of help it needs Once you know the friction, you can figure out what kind of help it needs. Clark and Kindred boil what AI can do down to five things and I think they are spot on: Recommend. It surfaces the right option from too many. Predict. It uses past behavior to guess the next step. Classify. It sorts things into categories you've already defined. Cluster. It finds groupings nobody defined in advance. Generate. It creates new content. Generation hogs the limelight, but many of the most valuable features quietly recommend, predict or classify. A receipt that files itself is worth more to most people than a chatbot that writes poetry about it. As a rough rule of thumb, slow tasks often suit prediction or generation, and repetitive ones suit classification and prediction. Confusing tasks call for a recommendation or a generated explanation. When there's simply too much, recommendation, clustering and summarizing help. But, treat that as a starting point rather than gospel. For Ledgerline, it played out like this. Picking a category is confusing, so classification reads the receipt and suggests one. Rejected claims with no reason are confusing too. Prediction can warn you that a claim is likely to bounce, and generation can explain why. Writing the business purpose is repetitive, so generation drafts it for you. Rubber-stamp approvals are a “too much” problem. Clustering and summarizing can show a manager which claims deserve a second look. Check the data (and the bill) Every capability depends on data. Classification needs examples of things sorted correctly, prediction needs history and recommendation needs to know what's relevant to whom. A great idea with no data behind it is just a nice sketch. So ask what data you have, and whether you're allowed to use it, before anyone opens Figma. Keep an eye on running costs as well. Every AI call costs money and time, and a feature that delights users but triples your bill will not survive its first budget review. Finance will make sure of that. Decide what it looks like last Only now should you think about what the feature looks like. Friction comes first, then capability, and the form comes last. The same capability can take very different shapes. Ledgerline's category suggestion could be a button that reads a receipt, a chat where you describe the expense, an agent that files everything overnight or a quiet hint while you type. That's one capability and four very different experiences. AI can play four roles: Tools you use. They're precise and focused, and you stay in control. Chat you talk to. It's open-ended and good for exploring, but harder to predict. Agents you delegate to. They work out of sight and report back. Copilots that back you up. They're there when you need them and out of the way when you don't. Ledgerline ended up with a tool for scanning receipts, a copilot that flags policy problems as you fill in a claim and an agent that builds the month-end report but asks before submitting it. Chat survived too, as a side panel for policy questions where it genuinely helps. The floating chatbot did not make the cut, and I can't say I feel sad. A quick word about chatbots Chat is just one shape a feature can take, and it's often a slow one. Typing a request takes longer than clicking a button, and reading paragraphs takes longer than scanning a table. Vitaly Friedman's research into AI interface patterns shows the better products bringing back buttons, sliders and menus, so people can steer the AI without writing an essay first. For plenty of tasks, chat adds work instead of removing it. Plan for when it goes wrong (because it will) AI will get things wrong. It deals in probabilities rather than certainties, so the time to decide how your feature handles mistakes is during planning, not after the first angry support ticket arrives. For each idea, ask what could go wrong and who could get hurt. In Ledgerline, a misread receipt means someone gets overpaid or underpaid. An invented policy rule produces advice that's confident and completely wrong. And a convincing AI-drafted business purpose could help a dodgy claim slip through. Then match the level of human oversight to the stakes: Low stakes. AI acts and the user can undo it, like suggesting a category. Medium stakes. AI flags something and the user confirms, like policy warnings on a claim. High stakes. AI prepares and the user approves, like submitting the month-end report. Critical. Only people do it, like approving and paying claims. This also tells you which rules need to live outside the model. Politely asking a model to follow your rules is not enough. It might, it might not. In Ledgerline, AI never changes an amount, currency conversion always uses the official rate and the agent can prepare a report but never approve or pay one. Write it down before anyone draws a screen The last step is to capture all of this in a short brief before anyone starts designing screens. This is the one-paragraph capability brief I use in the workshop: Friction. What is slowing people down? Who feels it. Which users feel it, and at what moment? Capability. Does it need to recommend, predict, classify, cluster or generate? Data needed. What does the AI need to know to do this well? Risk. Where does it sit on the value versus risk grid? Success. How will we know it worked? For Ledgerline's category suggestion, the success measure was simple. We wanted to halve the number of claims rejected for being in the wrong category. You can test that, and it's a lot more useful than “add AI to expenses.” With the brief in hand, you can prototype the behavior before anyone writes production code. A system prompt, a handful of crumpled real-world receipts and a spreadsheet of results will teach you more than a polished mockup. But that's a newsletter for another day, assuming I remember. The best AI is the kind nobody notices The best AI features don't feel like AI at all. They feel like the product finally understands you. That rarely starts with a chatbot. It starts with someone asking where people get stuck and being honest about whether AI is the right fix. So the next time someone tells you the competition has AI and you need some too, smile, nod and ask them which problem they'd like it to solve. Then enjoy the silence. And if you're wrestling with an AI brief like that right now, hit reply and tell me about it. I might be able to point you in the right direction, or at the very least commiserate.













