LLM outage - Widespread agent silence
critical- Started
- Oct 14, 6:16 PM EDT
- Resolved
- Oct 14, 7:11 PM EDT
- Duration
- 55 min
Bland AI's updates (6)
- identified · Oct 14, 6:16 PM EDT
The issue has been identified and a fix is being implemented.
- identified · Oct 14, 6:24 PM EDT
An upstream outage with our inference infrastructure provider has taken down our primary dialogue LLM cluster, resulting in agent silence. We have blocked further outbound dispatch and are queueing calls. We are moving workloads over to restore functionality.
- identified · Oct 14, 6:45 PM EDT
LLM workload is scaling up in a healthy cluster, expecting 10 minutes until resolution.
- monitoring · Oct 14, 7:01 PM EDT
Replicas are scaled up in the healthy cluster, monitoring system health and errors now
- resolved · Oct 14, 7:11 PM EDT
Replicas are healthy, opening queue back up. Expect a backlog of requests over the next hour
- postmortem · Oct 14, 8:49 PM EDT
**Summary** On October 14, 2026, from about 3:00 PM to 4:10 PM PT, agents on many calls went silent or took several seconds to respond. Outbound calls requested during the incident were held in a queue and placed late. The queue was fully drained by 4:41 PM PT. **What happened** A data center network outage at our upstream infrastructure provider cut off the cluster that runs the language model our agents use in conversation. The model’s time to first response rose from about 0.3 seconds to more than 7 seconds on average, so agents could not answer callers in time. To keep new outbound calls from connecting into silence, we paused outbound dispatch and queued new requests. The provider moved our model to a healthy cluster, and agents recovered by 4:10 PM PT. We reopened the queue at 4:07 PM PT and released held calls at a controlled rate. **Impact** * Live calls, inbound and outbound, from about 3:00 PM to 4:10 PM PT: agents were silent or slow to respond, and many callers hung up early. * Outbound calls requested during the incident were delayed, typically by 25 to 35 minutes. The backlog cleared by 4:41 PM PT. Calls that were canceled while queued were not placed. * We have found no impact on call records or recordings. **What we’re doing about it** * We’ve deployed a fallback that retries failed conversation requests on a second cloud provider. We’re extending it to switch providers when responses slow down, not only when they fail. * We’re spreading live-call conversation across more than one provider or cluster, so one failure can’t affect every call. * We’re changing agents to recover the turn, for example by asking the caller to repeat, instead of going silent when the model fails. * We’re paging on-call as soon as model response times rise, and slowing outbound dispatch automatically when the model is degraded. We’re sorry for the disruption. If this incident affected your calls, please contact your account team for a closer review.
What our checks recorded
We checked Bland AI 11 times from Bland AI's start time to its resolution, once every five minutes. Every one of them reached Bland AI's site, so this incident did not show up as a failed page load. Our check loads the public site; it cannot see an API error or a broken feature. Median response time was 143 ms, against 129 ms over the 24 hours before. Bland AI posted the incident at Oct 14, 6:16 PM EDT.
Updates are Bland AI's own words from its status page (status: postmortem); we never add an inferred cause. The vendor's incident page ↗