Conversational AI Use Cases: The Moments Where Talking Beats Tapping

Most lists sort these use cases by industry. Banking. Retail. Insurance. Healthcare. That's the wrong axis. A chatbot in a bank and a chatbot in a clinic fail for the same reason: they answer when they should ask, or talk when a button would be faster. Below are real examples for each moment, two of them products we designed, and two public failures that show where the line sits.

Conversational AI use cases are moments, not industries

Vendors organise this topic by vertical because that's how they sell. Product reality looks different. The same chatbot use cases repeat across every sector: someone needs to hand over context, understand data they already have, get advice mid-task, or make a decision. Industry changes the stakes. It doesn't change the pattern.

The types of chatbots matter less than most teams think. A rule-based FAQ bot, an LLM assistant, and an agent that takes actions sit on one spectrum of autonomy. What separates good conversational AI from bad is whether the conversation is the shortest path to what the user wants. That's a design question before it's a model question.

Four moments keep coming up in the products that work. Each one below comes with real examples, two of them products we designed at The Gradient.

Capturing context: onboarding as a conversation

Forms are good at collecting what you already know to ask. Conversation is good at collecting what you didn't.

Norvana, an AI-native health companion we designed and prototyped for a UK startup, replaced the signup form with voice-first onboarding that covers sleep, diet, activity, and mood. The reasoning: in our experience, people share around ten times more when they talk than when they type. Personalised health advice is only as good as what the AI knows about you, and most health apps skip that step entirely.

That depth is what makes conversational AI feel intelligent later. The same apple scan is a good choice for one user and a warning for a diabetic. The model is identical. The context isn't. A chatbot answers. Conversational AI knows who it's talking to.

Voice was the right input here because onboarding is long, personal, and easier to say than to type. Voice doesn't fit every moment. We mapped where voice earns its place in enterprise products separately.

Claims intake: the same pattern in insurance

Insurance is learning the same lesson at the claim stage. With Hi Marley's conversational first notice of loss, policyholders explain what happened in a natural conversation instead of a claims form. At Ohio Mutual Insurance Group, the conversational option passed 15 percent adoption without any marketing. Nearly 85 percent of claims reported this way arrive with photos or videos attached. The conversation doesn't just collect fields. It invites evidence a form never asked for. If your product's first session feels like paperwork, this is the use case to start with.

Explaining data people already have

Most products show users their data. Few explain it. Revolut is betting its app on that gap. In April 2026, it launched AIR, an in-app AI assistant, to its UK customers first, then expanded it to Ireland and other European markets. Through one conversation, customers get spending insights, track investments, manage subscriptions, or freeze a lost card. Before, each of those tasks lived on a separate screen. Revolut says AIR ends the era of multi-step navigation. It also scoped what the assistant can touch: AIR only accesses data customers already see in the app, and third-party AI partners don't retain it. Drawing that line is the same permissions problem every AI agent faces: scope too wide and trust breaks, scope too narrow and the conversation can't finish the job.

The data was always in the app. What changed is that customers can ask about it in their own words. That's the shift from a banking app to a conversational banking platform.

If your users stare at a dashboard and then call support, conversation belongs on top of that dashboard.

Advice inside the workflow

The fastest way to kill an enterprise AI chatbot is to give it its own tab. People don't go somewhere to ask for help. They ask where they already are.

Happy, an AI team-coaching platform we designed for Happy Companies, taught us this. After testing Happy internally at The Gradient, the lesson was clear: coaching had to live inside the tools people use every day. So the conversational coach moved into Microsoft Teams, Slack, and email, available before a meeting with a manager or colleague, or during the meeting itself. Users ask about a colleague's workstyle or describe a tense situation and get guidance on handling conflict or giving feedback. Monthly active coaching usage reached 78 percent, and 86 percent of users reported improved leadership.

McKinsey's internal chatbot, Lilli, follows the same rule at a much larger scale. About three-quarters of the firm's 43,000 employees use it every month, sending more than 500,000 prompts. For every question, Lilli scans the firm's knowledge base, finds five to seven relevant sources, summarises the key points, links them, and names the internal experts to talk to. Grounded answers plus a route to a person. That combination is why an internal chatbot gets adopted instead of ignored.

The question for your team isn't "Should we build a chatbot?" It's "Where do our people already ask questions, and can the answer live there?"

Shortening the path to a decision

Good conversational commerce compresses a flow that used to take many screens into a few exchanges.

Zalando is the clearest European example of a conversational ecommerce chatbot. Its assistant gives personalised fashion and beauty recommendations in chat. Nearly 10 million customers used it in the first quarter of 2026 alone, against 6 million across all of 2025. Shoppers ask by budget, mood, occasion, or even the weather, and get suggestions from Zalando's own assortment.

Then Zalando kept widening what the conversation can finish. In mid-2026, it added order support inside the same chat: 15 percent of conversations with simple questions like "Where is my order?" are already resolved on the spot. It also added visual discovery: snap a photo of an item you like, and the assistant finds it or similar pieces in the catalogue. Between January and June 2026, high-value interactions with the assistant grew 63 percent. Zalando counts adding to wishlist, adding to cart, and purchasing as high-value.

The pattern is the same everywhere. Filter panels, a separate help centre, and a search box collapse into one place to ask. Every step that used to need its own screen now ends inside the conversation.

Look for flows where users fill in five screens to make one decision. Those are the conversational AI use cases with the clearest return.

Where conversation breaks

Two public failures mark the edges.

Cursor. In April 2025, developers using the AI code editor were logged out whenever they switched devices. A support agent called Sam told them this was expected under a new one-device policy. No such policy existed, and Sam was a bot. The invented policy led to subscription cancellations, and the company now labels AI responses in its email support. It is one of the clearest LLM hallucination examples of recent years, and it failed twice: the bot answered without a source, and users didn't know they were talking to a bot.

DPD. In January 2024, a customer trying to find a missing parcel couldn't get the delivery firm's chatbot to help, or even to give him a customer service number. So he asked it for a poem. It wrote one about how useless DPD was, and when he asked it to swear, it did. DPD blamed a system update and disabled the AI element. The swearing made headlines. The real failure came earlier: a customer with a lost parcel and no way out of the chat.

Both are design failures, not model failures. Cursor's bot answered from the model instead of from the policy. DPD's had no route to a person. Those are the control points where AI agents fail, and teams decide them before launch, not after the screenshot goes viral.

Before shipping, decide two things in writing: where the bot stops, and which source every answer comes from.

The best conversational AI doesn't look like a chat

Look at what worked: Revolut, Zalando, Lilli, Happy, Norvana. None of them is a chat window bolted to the corner of a page.

In Norvana, the "Ask Norvana" bar sits on every screen. It knows which view you're on and what data is in it, so users ask without explaining. The AI decides the format of the reply: sometimes text, sometimes tappable options. Scan a plate of chicken and it asks back: chicken or turkey? Say "my knee hurts" during a workout and it asks follow-up questions, then rewrites that session. The whole app is a conversation. It isn't a chatbot.

That's where conversational AI is heading. The classic chatbot flow diagram (user message, bot reply, repeat) is giving way to products where conversation is one layer alongside UI, voice, and camera, and the AI picks whichever fits the moment. That only works when generative UI runs on your design system, and it's part of a wider shift away from classical UI that product teams can't opt out of.

So when you map conversational AI use cases for your product, skip the industry list. Find the moments where your users hand over context, stare at data, need advice mid-task, or click through five screens to make one choice. Put conversation there. Put buttons everywhere else.