Most AI support agents can talk. Very few know when to talk first.
That's the gap Teachable ran into with Fin, its AI support agent. Fin could answer a ticket the moment someone opened one. What it couldn't do was notice a user quietly struggling on the Payment Settings page, or stuck mid-setup on Course Pricing, and say something before that user gave up and left. Teachable's own numbers made the size of the problem clear: 33% of new users were writing in specifically from the Course Pricing page, and another 30% were getting stuck on Payment Settings, often abandoning the flow entirely rather than filing a ticket at all.
That's the part reactive support can't see. A ticket only exists once someone decides to ask for help. Everything that happens before that decision — the rage clicks, the page abandoned mid-form, the second attempt at the same failed step — was invisible to Fin.
The fix: wire behavioral data straight into the agent
Teachable connected Fin to Pendo's Agent Toolkit (ATK), which exposes in-app behavioral signals — page visits, drop-off points, rage clicks, error clicks — to an agent through webhooks. The flow is straightforward: Pendo detects friction on a tagged page, fires a webhook with the page context attached, and Fin opens the conversation itself. Instead of "How can I help?" Fin can say "I noticed you're having trouble with X — want help?"
That's a meaningful architectural shift, not just a UX one. It's the difference between an agent that responds to queries and one that's watching for state changes and deciding when a change crosses a threshold worth acting on. That's the same problem a recent arXiv paper on context graphs for enterprise agents tackles directly: most agents are reactive by design, and closing that gap requires a delta detection engine that tracks state changes and a proactivity scorer that ranks which of those changes are worth surfacing. The paper's broader point — that reactive AI is an architectural limitation, not a model capability limitation — is exactly what Teachable ran into. Fin didn't need a smarter model. It needed a wire into the product's behavioral data.
What changed, with numbers attached
Teachable's before/after comparison, drawn from six weeks of live data:
- With Pendo context: 70% resolution rate, 17% escalation rate, across 1,100 conversations
- Without Pendo context: 62% resolution rate, 24% escalation rate, across 2,200 conversations
That's roughly an 8-point resolution lift and a 7-point drop in escalations just from adding behavioral context to the same agent. No larger model, no new intents — same agent, more context.
The proactive-outreach numbers add another layer: a 38% open rate on Fin's first proactive message, well above the response rate Teachable's old onboarding pop-up got. A 6% reply rate sounds modest until you remember the baseline is zero — these are conversations that wouldn't have happened at all under the old reactive model.
Where this fits the bigger pattern
Teachable's case isn't an outlier — it lines up with what the broader proactive-support data has been showing. Self-service triggered by behavioral signals is running 40 to 60% higher resolution rates than portals customers have to find on their own, and one industry benchmark this year put deeply integrated, action-taking agents at 70 to 85% resolution on well-scoped use cases — right in the range Teachable is now hitting.
The mechanism matters more than the vendor here. Pendo's approach — webhooks pushing real-time behavioral events into an agent — is one implementation of a pattern that's showing up across the MCP ecosystem generally: giving agents standardized, structured access to product and usage data instead of building one-off integrations per data source. Whether that access comes through a webhook, an MCP server, or a direct API call, the underlying bet is the same: agent quality is now as much a data-plumbing problem as a model problem.
For teams building or buying support agents, that's the actual takeaway. Before asking whether your agent needs a better model, ask whether it can see what your users are actually doing. Teachable's Fin didn't get smarter between the "without" and "with" numbers above. It just stopped being blind.
