Why adding AI to a product does not automatically make it better

There’s a lot of pressure to put AI into every product right now. Investors ask about it, competitors announce it, and a product without it can feel like it’s already behind.

We build AI features, and some of them are genuinely useful. But adding AI doesn’t automatically make a product better. Sometimes it makes it worse. This is how we think about the difference.

Start with the problem, not the model

The useful question isn’t “where can we add AI?” It’s “what is hard for our users right now?” If the answer involves reading messy text, summarising long documents or finding information scattered across many files, AI may well be the right tool.

If the answer is that the product is slow, confusing or missing a basic feature, AI won’t fix that. It will add a new layer on top of the same problem.

What AI costs, besides money

AI features come with trade-offs that don’t show up in a demo.

• Cost per use. Every request to a model costs something. That’s fine for a feature used occasionally, and painful for one used thousands of times a day.
• Speed. Model responses can take seconds. Where people expect an instant response, that delay is noticeable.
• Unpredictability. The same input doesn’t always give the same output, and a wrong answer can sound just as confident as a right one.
• Data. Information sent to a model provider leaves your system. For some businesses and some data, that’s a serious question.

None of these rule AI out. They’re reasons to use it where the benefit clearly outweighs them.

An example: leaving AI out on purpose

When we built Talk2UsNow, a review platform for local businesses, there was an obvious place for AI. Owners receive private feedback from unhappy customers, and it would have been easy to run each message through a model to detect sentiment and draft a reply.

We didn’t. The customer had already told us how they felt by choosing a star rating. Sorting feedback into topics like waiting time or cleanliness worked well with simple keyword rules. The suggested next step for each topic could be written once, by a person, and reused.

The result is predictable, easy to check and costs nothing to run. For a product priced for small businesses, that mattered. We also labelled the tags as automatic tags rather than calling them AI, because they aren’t.

When AI does earn its place

AI is worth it when the input is genuinely messy and the alternative is a person reading everything. Summarising long documents, pulling details out of varied emails and answering questions across a large set of files are good examples.

Even then, we design for mistakes. Where errors matter, the software suggests and a person decides. We test on real examples from the business rather than a few hand-picked demos, and we plan for what happens when the model is slow or unavailable. There’s more on this on our AI development page.

A simple check

Before adding an AI feature, try describing what it does without using the word AI. If the description is clearly useful, such as “it reads incoming invoices and fills in the supplier, amount and due date”, it’s probably worth exploring. If it sounds vague, such as “it makes the product smarter”, it probably isn’t ready yet.