Most "AI problems" aren’t AI problems

The fastest way to waste a quarter is to point a language model at a problem that was never about language in the first place.

We start every engagement with a week of discovery, and the most valuable thing we deliver in that week is often a recommendation not to build. Not because the technology can’t help, but because the real bottleneck turns out to be a missing integration, an unowned process, or a decision nobody wants to make.

The audit comes before the build

When a team tells us “we need an AI agent for support,” we don’t start with the agent. We map the actual path a ticket travels — who touches it, where it waits, what information is missing when it arrives. More often than not, half the delay has nothing to do with answering the question. It’s routing, duplicate data entry, or a handoff that drops on the floor.

Automating the answer while leaving the broken plumbing in place just produces a faster version of the same mess.

Leverage lives in a few places, not everywhere

In a typical workflow, a small number of steps account for most of the cost and most of the pain. Those are the places where automation pays for itself. The rest — the long tail of edge cases and one-off exceptions — is usually cheaper to leave to a human.

We score every opportunity on effort versus impact and hand back a short list. The goal isn’t to automate the workflow. It’s to automate the two or three steps that actually move the number.

An honest no is a deliverable

The most useful sentence we say in discovery is sometimes “don’t build this yet.” It costs us a project and saves the client a year. That trade is worth making every time, because the alternative — a system that technically works but solves the wrong problem — is far more expensive than a hard conversation in week one.

If you’re considering an AI build, run the audit first. The roadmap you get back is worth more than the demo you imagined.