AI features actually worth building (and ones that aren't)

AI is having a moment, and that brings a lot of pressure to "add AI" to your product. Some of it is genuinely worth it. A lot of it is a shiny feature nobody asked for that quietly costs you money every month. Here is how we tell the difference before writing a line of code.

Start with the problem, not the technology

The wrong question is "where can we put AI?" The right one is "what is slow, repetitive or annoying for our users, and could a machine do it well?" AI is a tool, not a goal. If a feature only exists so you can say you have AI, users feel it, and it ages badly. The features that stick solve a real, boring problem people already have.

Features usually worth building

• Search that understands meaning. Letting people find things by describing them in their own words, instead of guessing the right keyword, is a genuine upgrade almost everywhere.
• Summarising long stuff. Turning a wall of text, a call, or a document into a short, useful summary saves real time.
• Answering questions from your own content. A support assistant that answers from your real docs, so people get an answer instead of a ticket.
• Sorting and tagging. Quietly categorising, routing or flagging things in the background, where a small mistake is cheap and a human still has the final say.

Features usually not worth it yet

• AI as the last word where being wrong is expensive, with no human checking. Think money, legal or medical decisions made automatically.
• A chatbot bolted onto a product that did not need one, added only because a competitor has one.
• "Generate anything" features with no clear job, that look magical in a demo and get used twice.

The question that saves you money

Before building any AI feature, ask: what happens when it is wrong? Good AI features are ones where a mistake is cheap and easy to catch, and where a person can step in. AI is brilliant at being roughly right, fast, across huge volumes. It is risky as the final say on something that has to be exactly right. Build for the first case and keep a human on the second.

How we approach it

We start small and real: one feature, on your actual data, that solves one clear problem, shipped behind a simple check so you can see it working before you trust it. AI features are easy to demo and hard to make dependable, so the value is in the unglamorous parts, the guardrails, the testing, the "what if it is wrong" handling. That is the difference between a headline and something your users quietly rely on.