Why we don't build generic AI products

Every week someone asks us to build them an AI chatbot. Our answer is always the same: we don't build chatbots. We build systems that do specific things well.

There's a difference, and it matters more than most people realise when they're writing the brief.

The problem with generic

A chatbot bolted onto a product is almost always a UX regression dressed up as innovation. It replaces unambiguous interactions with ambiguous ones. The user now has to figure out what to ask. The system has to figure out what they meant. Half the time a well-labelled button would have been faster and more reliable for everyone involved.

Generic AI tools are optimised for demo quality, not production quality. They look impressive in a Loom recording. Then real users show up and discover a system that almost does what they need, most of the time. That's actually worse than a system that reliably does one thing, because at least with the latter you know exactly what you're getting.

What bespoke actually means

When we say bespoke, we mean a system shaped around a specific workflow, for specific users, in a specific context. The AI model is one component of that. The interface design, the data pipeline, the integration surface, the human-in-the-loop handoff points: those are the parts that take judgment to get right and can't be templated.

The model is usually the easy part. The hard part is deciding precisely what the system should never do on its own.

The test we use

Before taking on a project, we ask one question: if you replaced the AI in this system with a rules-based system tomorrow, would the product fail? If the answer is no, it's not really an AI project. It's a software project with an AI component. That's fine! But it changes the architecture, the evaluation approach, and what success looks like.

The projects that genuinely excite us are the ones where the test fails hard. Where the reasoning capability, the contextual pattern-matching, the ability to handle inputs nobody anticipated: those things are genuinely load-bearing. That's where the interesting engineering lives.