Building Iterative Pilots

In the previous unit, you mapped the employee journey and built personas that told you where AI intervention would outperform headcount — and where human judgment had to stay. Now comes the part where many teams stall: actually building and shipping. As Iris McQuillan-Grace walked through in the session, Oliver didn't spend months perfecting a bot behind closed doors. They assembled what they already had, gave it a persona and guardrails, tested it in rapid stages with stakeholders who were invited to break it, and launched it the same day as the process it was designed to support. The results — 19,000 feedback requests against an expectation of 3,000 — proved that speed and iteration beat waiting for perfect.

One of the most practical takeaways from Iris's session was how un-glamorous the bot's foundation was. Oliver's "Spybot" wasn't trained on proprietary datasets or custom-built models. Iris described feeding it "I would say, like 150-ish artifacts" — the organization's "mission, our vision, our values, all of our own IP around training, how to run a one-on-one, how to write" feedback — plus external data. As she put it, "We basically threw the kitchen sink at it." They chose the SBII feedback model as the bot's behavioral backbone (hence the name "Spybot") and then layered on "a series of constraints and considerations" to govern how it responded. The bot had a defined persona — not just instructions, but a character that shaped tone and boundaries. This matters because the artifacts you already have — your behavioral frameworks, training decks, values documents — are the raw material. You don't need new IP. You need to curate what exists, choose a model that aligns with how your organization already talks about the skill you're coaching, and set explicit constraints so the bot stays in its lane.

Building on the artifact assembly, Iris described a four-stage pilot that moved at a pace most HR teams would find uncomfortable — and that was the point. Stage one: Iris tested the bot "in my own sandbox." Stage two: she brought in roughly 30 stakeholders from the earlier co-creation workshops — people who "were already engaged... already invested... and they wanted this to work" — and gave them two weeks with one instruction: "Try to break this." Stage three: revisions based on that feedback, then nine days for the C-suite to review. Stage four: launch "the same day as performance management launched." The results exceeded every projection. Oliver expected 3,000 feedback requests from 3,000 eligible employees; they got 19,000 requests with a 97% completion rate. Global performance review completion jumped from roughly 35% to 77% year over year — a 40% increase. HRBP feedback-related tickets dropped immediately. And perhaps most telling, approximately 2,000 employees who had never touched Oliver's AI sandbox before started engaging with it, organically building "basic prompt engineering" skills just by interacting with the bot. That last metric matters because it shows the ripple effect — a well-placed AI tool in a high-stakes moment doesn't just solve its own problem; it becomes the on-ramp for broader AI fluency across the workforce.