AI agents in production: lessons from real NZ deployments
AI agents look effortless in a demo. Production is a different environment entirely — real customers, real edge cases, and real consequences when something goes wrong. Here's what we've actually learned putting agents into live use for NZ businesses.
What goes wrong
The failures are rarely dramatic. An agent doesn't usually go rogue — it quietly misreads an ambiguous request, gives a confident answer to a question it should have escalated, or gets stuck in a loop when a tool call returns something unexpected. The risk isn't Hollywood AI going haywire, it's a small, plausible-sounding mistake that nobody notices until a customer flags it.
The second common failure is scope creep: an agent built to answer FAQ-style questions gets asked something it was never designed for, and instead of saying "I don't know", it improvises. Improvising is exactly the wrong behaviour for anything touching bookings, money, or commitments made on the business's behalf.
Designing for graceful failure
The fix isn't a smarter model — it's tighter boundaries. Every agent we deploy has an explicit list of what it's allowed to do and a hard rule to escalate to a human the moment it's outside that list, rather than guessing. Every action that changes real-world state — sending an email, confirming a booking, updating a record — goes through a defined tool call with validation, not free-form text generation. That turns "the AI might say something wrong" into "the AI can only do things we've explicitly allowed it to do".
Logging matters just as much as the guardrails. Every conversation, every tool call, and every escalation gets recorded, so when something does go sideways, it's a five-minute fix instead of a mystery.
What we've learned deploying in NZ
Local context matters more than people expect. An AI receptionist needs to understand Kiwi phrasing, local business hours conventions, and the fact that callers will happily talk over a scripted greeting. Generic, overseas-trained assumptions about how a conversation should flow don't always hold here, and the agents that perform best are the ones tuned against real local call transcripts, not a generic template.
The businesses getting the most value aren't the ones chasing the most ambitious agent — they're the ones that scoped a narrow, well-defined job, built proper guardrails around it, and let it run reliably before expanding what it's trusted to do.