Why Does Your AI Agent Sound Like an Old Answering Machine?

You type a straightforward question in your natural dialect, and back comes a canned reply that has nothing to do with what you asked, followed by a numbered menu: “For account balance, press 1.” Countless services marketed across our region today as “artificial intelligence” behave this way. Some are deliberately muzzled out of fear of legal or financial liability—which is understandable—but many suffer from the exact malady my own team inflicted on our code: not the weakness of the AI model, but the maze of software guardrails built around it that choke the model before it can speak.

I know this labyrinth from the inside because I built one with my own hands. When we started building a WhatsApp bookkeeping agent for small merchants, I assumed prudent engineering meant picking a fast, inexpensive mid-tier model and surrounding it with enough deterministic code rules to keep it on the rails. Financially, the logic seemed sound—a frontier model cost several times more per call, and bookkeeping tolerates zero arithmetic creativity.

Every time the mid-tier model slipped on an edge case, we wrote a regular expression or a pre-classifier to patch the hole. Another edge case appeared, so we added a second conditional branch, then a third. Before we knew it, we had spiraled into a month-long vortex of brittle workarounds—followed by another month just to diagnose what we had done and rip it all out.