Every week we get at least one pitch for an "AI feature" that, on paper, does something a well-written SQL query could already do. Sometimes we build it anyway. Usually, we refuse — and explain why.
Here's the filter we use.
Three questions
Does it do work a human was going to do?
If the answer is no, it's not AI earning its keep — it's an AI demo. The moment the novelty fades, the feature collapses into a cost line on the P&L.
Can we measure the value per run?
Voice agents book appointments. RAG search saves analyst hours. A chatbot that "helps" users without a measurable saved-minute is just an expensive marketing gesture.
What's the failure mode?
Some failures are trivial (misfiled support ticket). Some are catastrophic (wrong dosage, wrong price, wrong instruction to a contractor). We build features where the worst case is embarrassing, not dangerous — and we design the guardrails first.
What we refuse
- "Let's add AI to our dashboard" — unless the dashboard is already being ignored for a reason AI won't fix.
- "Can it write our marketing copy?" — unless you want to sound like every other company that asked the same thing.
- "We need a copilot" — unless you can tell us, precisely, which task it's replacing.
What we love
Voice agents that book real appointments. RAG over a messy data room that analysts used to pick through by hand. Custom evals that turn a flaky model into a production tool. Anything where the business math is already obvious, and we just need to build the thing well.
That's the job. AI doesn't need to be magic. It just needs to earn its keep.