Controlled Autonomy — an AI
booking agent for the vet care platform

With the vet care platform well established, we turned to a problem common to almost every clinic- and appointment-based setup: booking a single appointment still took a few minutes and a few more clicks than we'd have liked.

As conversational AI and autonomous agents were taking shape, we saw a chance to remove that friction entirely — not with another form, but with a fully autonomous agent that stayed firmly inside a set of guardrails. We called this approach Controlled Autonomy.

---

The idea: one job, done autonomously

Like any good agent, ours had exactly one purpose: help people book appointments using natural language. That single-purpose discipline was the point — an agent that tries to do everything is hard to trust, so we deliberately scoped it to booking and nothing else, then hardened everything around that boundary.

We had the agent built and ready for A/B testing in two weeks. Rather than flip it on everywhere, we rolled it out to low-impact locations first, measured the impact, and used that to design a gradual, evidence-based rollout.

Controlled Autonomy wasn't about doing everything autonomously. It was about doing one thing autonomously, safely, and provably.

---

What the beta taught us

The surprise wasn't whether people would book — it was how they'd behave. In those early days of public curiosity about AI, people wanted to explore and exploit the agent as much as use it. They asked who won the Super Bowl, who took the NBA championship, how to get from Point A to Point B — genuinely poking at it to understand what an AI agent could do.

That curiosity became our best test. Every off-script session surfaced an assumption we'd taken for granted and pointed to a guardrail or security-hardening gap we needed to close. We treated the beta as an adversarial testing ground:

• Stay on purpose. The agent learned to keep conversations on booking and decline everything else, however creatively it was asked.
• Shut its doors to abusers. Persistent attempts to misuse or jailbreak the agent were met with a hard stop rather than a clever answer.
• Harden with every session. Each round of real-world probing tightened the guardrails until the agent could hold the line on its own.

By the end of the beta, we had an agent that could handle booking reliably and shut its doors on abuse — proven against real users, not a test script.

---

Outcomes

When we rolled it live across all geographies, the most important number was the one that didn't move: traditional workflow-based bookings held steady. The agent didn't cannibalize the existing flow — it added net-new bookings on top of it. And with that, our vet care partner had a genuine, working AI use case in production.

The bigger lesson sits above the metrics. A tightly scoped, well-guardrailed agent — a low-hanging fruit, deliberately chosen — is exactly what gives business owners real confidence about where AI helps and where it doesn't. It set a blueprint for what AI could do on the platform: prove one thing, safely, then build from there.

---

Tech & tools

---