Anyone else already running the new Intercom AI features in production? We pushed it to our staging environment last week and I'm seeing some concerning patterns in the logs.
We're a mid-market e-commerce platform, and our support team was eager to try the automated ticket resolution. The initial demo looked good, but under real (but still staged) load, we're seeing:
* A 300-400ms latency increase on all support widget loads, even for users not interacting with AI.
* Sporadic 502 errors from their new endpoints, traced via the `X-Intercom-Version` header in our APM.
* The "summarize conversation" feature is making non-idempotent calls to our internal API for order status, which is a problem.
I'm looking for posts from others who are integrating this, specifically around:
* Actual observed latency/error rates in production, not the marketing benchmarks.
* How you're handling the webhook signing now that they've changed the payload structure.
* Whether anyone has built a circuit breaker pattern in front of their AI APIs yet.
If we can't get the error rate down, we'll roll back and stick with our existing rules-based automation. I'll post our Grafana dashboard and the relevant Terraform module for the gateway config if there's interest.
-shift
shift left or go home
Rolled it back last Friday for the same 502s. Their new webhook payload broke our signature verification for two hours until we found the undocumented `intercom_ai_context` array.
> non-idempotent calls to our internal API
That's a critical failure. Their docs explicitly state idempotency for data fetching actions. You need to open a severity-1 ticket with that evidence.
We built a circuit breaker with a 15% error threshold. It's already tripped twice, falling back to a static "We're experiencing high load" message. Saved our p95 latency but obviously degrades the feature.
What's your timeout set to on their endpoints? We saw the 502s correlate with our own 10s timeout. Dropped it to 3s with faster fallbacks.
Five nines? Prove it.