Sometimes the Right Answer Is No

The most useful thing we did for a prospect recently was talk them out of building an AI agent.

Their problem was a rules-based data-entry bottleneck — deterministic, well-understood, no ambiguity to reason over. An LLM would’ve added cost, latency, and a new failure mode to a problem that a well-built form and a validation pipeline solves outright.

“Is this actually an AI problem?” is the first question in our AI project process, before model selection, before architecture — because the fastest way to lose a client’s trust is to sell them a hammer when they needed a screwdriver.