Early prototypes of our AI copilot were, frankly, annoying. It interrupted every few seconds, over-explained obvious UI elements, and used a tone that felt more like a pop-up ad than a helpful colleague.
We found this out the unglamorous way: watching session recordings back-to-back and literally counting interruptions per minute. The first prototype averaged one copilot message every 8 seconds during active exploration. We tagged each message by whether the buyer's next action suggested it had helped (they engaged with what it pointed at) or annoyed them (they clicked away, closed the panel, or ignored it entirely for the rest of the session). Roughly two-thirds of messages fell into the second bucket.
The redesign started with a simple rule: the copilot only speaks when it can say something the buyer couldn't figure out faster on their own. That meant cutting explanatory messages by roughly 70% and replacing generic tooltips with contextual, single-sentence nudges tied to actual buyer behavior.
A concrete example: the old copilot greeted every session with "Welcome! I'm here to help you explore our platform — click anywhere to get started, and I'll guide you through the key features!" Nobody needed that sentence; it just delayed the moment a buyer could start clicking. The new copilot says nothing at all until a buyer either pauses for more than 10 seconds on one screen or triggers something worth explaining — at which point it says one sentence, tied to exactly what's on screen, and then goes quiet again.
We also gave the copilot a consistent, low-key personality — direct, technically fluent, comfortable saying 'I don't know, here's who to ask.' Buyers trust an assistant that admits its limits far more than one that pretends to have all the answers.
Tone had to work for very different audiences without becoming generic. A security engineer stress-testing SSO config and a marketing lead looking for reporting dashboards are having completely different conversations with the same copilot, and neither should feel like the personality was designed for the other one. We solved this less by writing more personas and more by writing fewer, sharper rules: short sentences, no exclamation points, no manufactured enthusiasm, and never explain something the buyer's own click just demonstrated they already understood.
The result, Agent Sarah, now drives measurable engagement without becoming the thing people try to click away from.
We're not done — the two-thirds-annoying number from the first prototype is now closer to one-fifth, based on the same next-action tagging method, and closing that remaining gap is this quarter's design goal. The honest long-term test isn't a satisfaction score; it's whether buyers keep letting the copilot talk at all.