The best AI products hold your hand and nudge you in the right direction instead of overwhelming you with endless possibilities. Building an AI-powered product can create truly magical experiences for users, but so many products today feel like someone decided to throw yet another chat box into their app. 

Which begs the question, when and how should you start baking AI into your product?

Having built AI products at Discord and advised (and invested in) companies navigating our new normal, I think it ultimately comes down to one key question that any product owner should be able to answer. Do you have a clear action you want to nudge your users towards that they wouldn't otherwise take without AI?

It's a fairly common product principle to make sure users know what to do next. Products should (typically) be designed in a way where it's clear what users can do and where to go to do it. Sometimes that can be done with engaging feeds users navigate to by default while other times notifications, tooltips, and other user education can provide guidance.

Where AI can come in is by quickly extracting novel insights that can lead to meaningful actions your users probably want to take under a certain set of conditions. Do you want people to look into recent user reviews trends when sentiment gets negative? Do you want someone to hop into the app when the conversation turns to an interesting topic? Modern AI tools have made it blazingly fast to start a pipeline that generates these insights compared to what we had before, but the trade-off is that you probably get lower accuracy than a traditional ML pipeline. And that might be okay as long as your users have the right expectations and your suggested actions as just that. Suggestions.

On the other hand, knowing when you shouldn't inject AI into your product is just as important to avoid unnecessary product bloat. If you're tempted to add an AI bot into your product workflow, ask what you want your users to do with it. In some cases, conversational flows can be great to get information from a user as part of an onboarding flow, but, a lot of the time, it might be better to invest in your hero use cases and prompt users about them when relevant (like a popup about a potential build break with suggested next steps). Conversely, if you can use simple notifications or basic tooltips for flows that everyone should consider, just do that and don't overcomplicate things.

If you don't know what you want your users to do but just like the idea of them exploring data with your app, please think twice. Your users are probably sophisticated enough to upload their spreadsheet into ChatGPT and ask it some general questions. We don't need to talk to all of our products as LLM chatbots just yet, but I'm pretty sure your users will be thrilled if your product consistently suggests things they want to do with easy next steps.