Recently, I wrote about giving RB2B's AI support agent a name and a face. The results felt significant but were still anecdotal. The team at Intercom were apparently interested enough in the observations that they offered to run a proper analysis themselves. Here's what they found.
The setup
The rename happened in early March. Intercom compared August–October 2025 (before) against January–March 2026 (after), and restricted the analysis to conversations where Fin was involved — which controls for the fact that Fin's scope also expanded in January, from handling around 70% of conversations to about 87%. Raw numbers would have been misleading; these aren't.
What changed
The frustration finding is the one I keep coming back to. Words like "frustrated," "broken," "useless," and "not working" appeared 37% less often — and this was while Arbi was handling 19% more conversations. More volume, less friction.
Resolution improved too. Soft resolution — where the user left without needing a human — climbed from 63% to 69%. Routing to human agents fell from 23% to 20%. In other words: Arbi got better at actually helping people, not just at seeming friendlier.
What didn't change
Politeness held flat. "Thank you" rates: 11.6% before, 11.7% after. "Please" rates: 8.1% before, 8.2% after. The raw numbers initially appeared to drop, but that was entirely a composition effect — Arbi was now talking to a larger, different-shaped pool of conversations. Controlling for that, people were exactly as polite as before. That's actually reassuring: the experiment didn't make people perform politeness, it just made them less likely to express frustration.
Intercom's analysis also corrects something from the original post. I wrote that users were sending longer, more descriptive messages — that they were explaining what they'd tried, what they were after. That felt true qualitatively. The data disagrees: messages were about 4% shorter on average after the rename, not longer. That may be a scope expansion effect, or it may mean the interactions got more efficient rather than more elaborate. Either way, the claim didn't hold up and it's worth saying so.
What this probably means
There's no control group here — we can't run a parallel universe where RB2B didn't rename Fin and compare outcomes. The before-and-after design holds up well enough given the methodological controls, but it's still a before-and-after. Something else might have contributed.
That caveat noted: the pattern is coherent. Users who feel like they're talking to someone — rather than submitting a ticket to a system — bring different energy to the interaction. Less adversarial. More willing to engage. And when the interaction is more cooperative, the AI has more to work with and does better work. Frustration drops, resolutions go up, escalations fall.
It's not a trick. Arbi still announces itself as an AI agent in every message. But a name and a face changed the psychological frame, and the psychological frame changed the outcome. That's a lesson that generalises well beyond AI support agents.
Analysis by Intercom, covering August–October 2025 vs January–March 2026. Population restricted to Fin-involved conversations to control for simultaneous scope expansion.
