What We Learned Building AI Into a Real Product

Everyone's building AI into everything. Most of it is nonsense — a chatbot bolted onto a product that didn't need one, so someone can stick "AI-powered" on the landing page.

We've shipped AI features into three products now. Estimait, Solora, Cynk. Each one taught us something about what actually works versus what sounds good in a pitch deck.

The Hype vs The Reality

The hype: AI will transform everything.

The reality: AI is good at specific things. Pattern matching at scale. Summarising large amounts of text. Generating first drafts. If your use case fits those strengths, it's genuinely transformative. If it doesn't, you're forcing it.

Estimait works because estimation is pattern matching. You're looking at a brief and comparing it to hundreds of past projects. That's exactly what LLMs are good at. We're not asking the AI to be creative — we're asking it to find patterns faster than a human can.

Solora works because story generation is a first draft problem. The AI produces something, a parent reads it to their kid, the kid doesn't care that it's not Pulitzer material. The bar is "good enough to enjoy at bedtime," not "publishable."

Where We Got It Wrong

Early versions of Estimait tried to do too much. We wanted the AI to understand context it couldn't possibly know — client relationships, team dynamics, the politics of why a project might run over.

It couldn't. And we wasted weeks trying to make it.

The fix was simple: let AI do what AI does, let humans do the rest. Estimait generates a baseline. The human applies context. Together you get something better than either alone.

The best AI features disappear into the product. You don't notice the AI — you just notice that the thing works.

What We'd Tell Other Builders

Start with the constraint, not the technology. What specific problem could AI solve better than the current approach? If you can't answer that clearly, you're building a demo, not a product.

Build the non-AI version first. Understand the workflow without AI, then figure out where AI actually helps. Otherwise you're guessing.

Expect to throw away your first attempt. Your assumptions about what AI can do will be wrong. That's fine. Ship, learn, rebuild.