Hi everyone. I've been lurking for a while but this is my first post, so thanks in advance for your patience.
My company (we're a mid-sized SaaS in the logistics space) rolled out Intercom's Fin AI about 18 months ago. The initial pitch was great for us—reduce repetitive questions, help our small support team scale, etc. We've hit the 18-month mark, and I'm curious to see if our experience lines up with others.
Our deflection rate looks okay on the dashboard (it says ~31%), but I'm more concerned about how our actual support agents feel about it. I get the sense they don't really trust it. They tell me they spend a lot of time checking Fin's work before sending replies, or they just turn it off for more complex tickets. It hasn't really become the "co-pilot" we hoped for. It feels more like an extra step they have to monitor.
Has anyone else been using it long-term and seen real agent adoption? Not just the metrics, but the team genuinely relying on it? What made it click for your agents, if it did? Or are we maybe using it wrong? We didn't do a ton of training beyond the initial setup.
Any honest experiences or data would be super helpful. We're looking at our support tool stack again soon, and I want to know if we should push for better adoption here or look elsewhere.
Your experience aligns with a common pattern I've observed in platform teams evaluating AI support tools. The deflection rate metric is often a vanity metric; what matters is the agent trust coefficient, which rarely gets tracked. I've seen teams where agents actively bypass the AI co-pilot because the cognitive load of verifying its output exceeds the time saved for any ticket beyond the most trivial.
The turning point for one of our teams came when they stopped using Fin for final answers and instead repurposed it for first drafts. They implemented a mandatory review workflow where every AI suggestion required a two-click "correct" or "incorrect" rating from the agent before sending. This generated the structured data needed to retrain the model on their specific logistics edge cases. After about three months of this, the accuracy on their core use cases improved enough that agents began to tentatively rely on it for tier-1 queries.
Without that continuous feedback loop, these systems plateau. You mentioned limited training beyond initial setup that's likely the root cause. The model hasn't evolved with your team's unique knowledge.
No free lunch in cloud.
Oh, the magical "mandatory review workflow" fix. That's a classic.
You're basically describing a new, hidden labor tax you're imposing on your already-busy agents. "Just two clicks" per reply adds up to hours per week of uncompensated model training work, with zero guarantee of payoff. The company gets a better AI model, but what's the agent's incentive? A marginal decrease in future tedium, maybe?
I've seen this backfire spectacularly when leadership then uses the "improved metrics" to argue for headcount reduction. Agents aren't dumb, they see the play. The trust issue isn't just about accuracy, it's about whether the tool is there to help *them* or eventually replace them.
But what about the edge case?
Thanks for sharing this, it's really helpful to hear a real experience. That line about it feeling like "an extra step they have to monitor" hits home. We're about 6 months into a trial with a different tool and I'm already hearing similar grumbles from our team.
Did you ever get a sense of what kind of tickets Fin actually handles well versus the ones where agents just switch it off? I'm wondering if focusing it on very specific, simple things from the start would have built more trust.