Stop Building AI Features Nobody Asked For

Somewhere in the last two years, “add AI” became the default answer to every product question. Users churning? Add AI. Dashboard feels stale? Add AI. Investors asking tough questions? Definitely add AI.

The result is a landscape littered with AI features that nobody uses. Chatbots jammed into apps that didn’t need them. “AI-powered” labels slapped on basic search filters. Summarisation buttons that produce worse output than just reading the original.

This isn’t an AI problem. It’s a product problem. And it’s costing companies real money.

The AI feature graveyard

We work with companies across industries, and the pattern is almost always the same. Someone on the leadership team reads about a competitor launching an AI feature. Panic sets in. The roadmap gets reshuffled. Three months later, there’s a chatbot in the sidebar that 4% of users have tried and 0.3% use regularly.

Here’s why this keeps happening:

• The feature was supply-driven, not demand-driven. Nobody asked for it. No user interview surfaced it. No support ticket hinted at it. It exists because AI exists, not because a problem exists.
• The team built the demo, not the product. AI demos are impressive. They work great on stage and in Loom videos. But production AI that handles edge cases, bad input, and real user behaviour is a completely different animal.
• Success was never defined. What does “good” look like for this feature? If you can’t answer that before building, you definitely can’t answer it after launching.

How to actually think about AI in your product

Start with the workflow, not the technology

The best AI features are invisible. They don’t announce themselves with sparkle emojis and “Powered by AI” badges. They just make an existing workflow faster or less painful.

Look at where your users spend the most time doing repetitive, low-value work. Data entry. Report formatting. Sorting through search results. Triaging support tickets. These are AI opportunities - not because the technology is cool, but because the problem is real.

Ask yourself: if this feature worked perfectly, would the user even notice it’s AI? If yes, you’re probably on the right track.

Build the smallest useful thing

The instinct is always to build the general-purpose version. A chatbot that can answer anything. An assistant that handles every workflow. A magic text box that does it all.

Don’t. Build the version that does one thing well for one specific use case. A feature that automatically categorises incoming support tickets into your existing five categories is more useful than a general-purpose AI assistant that sort of helps with everything.

Constraints are your friend here. The narrower the scope, the better the output quality, and the easier it is to measure whether the thing actually works.

Validate with the ugly version first

Before you build the polished AI feature with the streaming text animation and the beautiful UI, build the ugly version. A script. A spreadsheet. A Slack bot that calls an API.

Put it in front of five users. Do they use it? Do they come back? Do they complain when it breaks? If the answer to all three is yes, you have something worth investing in. If they shrug and go back to their old workflow, you just saved yourself three months of engineering time.

We’ve killed more AI features at the prototype stage than we’ve shipped to production. That’s not failure - that’s discipline.

The three questions that filter out bad AI features

Before committing engineering resources to any AI feature, run it through these:

1. What specific task does this replace or accelerate?

If you can’t name the exact task, with the exact user, in the exact context - stop. “It makes things easier” isn’t an answer. “It auto-generates the weekly client report from the last 7 days of project updates” is.

2. What happens when the AI is wrong?

Because it will be wrong. Sometimes hilariously wrong. The question is whether “wrong” means a minor inconvenience (the summary missed a point, the user edits it) or a serious problem (the AI approved a fraudulent transaction, the chatbot gave medical advice).

Design for failure from day one. The best AI features make it easy to correct mistakes, not hard to notice them.

3. Is this better than a rule-based system?

Honestly? Half the “AI features” we see could be replaced with a well-designed filter, a decision tree, or a simple automation. AI is the right tool when the input is unstructured, the patterns are complex, or the task requires understanding context. For everything else, a few if statements will do.

This isn’t anti-AI. It’s anti-waste. Use the expensive, complex, sometimes-wrong tool where it genuinely shines, and use simpler tools everywhere else.

The real competitive advantage

Here’s what nobody talks about: the companies winning with AI aren’t the ones with the most AI features. They’re the ones with the best data pipelines, the cleanest feedback loops, and the most disciplined product thinking.

AI features improve over time - but only if you’re collecting the right data, measuring the right outcomes, and iterating based on real usage. Shipping an AI feature without a plan to improve it is like launching a product without analytics. You’re flying blind.

The competitive advantage isn’t “we have AI.” Everyone has AI. The advantage is “we know exactly which problems AI solves for our users, and we’re getting better at solving them every month.”

What we tell our clients

When companies come to us wanting to “add AI to their product,” we start by asking what problem they’re trying to solve. About half the time, the best solution doesn’t involve AI at all. The other half, we help them build something focused, measurable, and designed to improve over time.

The goal isn’t to have AI features. The goal is to have a better product. Sometimes AI gets you there. Sometimes it doesn’t. The companies that understand this build products people actually use. The rest build demos.

Stop chasing the technology. Start chasing the problem. The AI part will figure itself out.