What happened
An internal configuration change restricted which AI models the Synthflow platform was permitted to call. The restriction was applied more broadly than intended and blocked the models used by our post-call processing pipeline.
As a result, AI-generated post-call outputs failed to produce calls completed during the window. The configuration was reverted at 11:10 UTC and all functionality returned to normal immediately.
What was affected
Calls placed and received during this window connected, ran, and completed normally. The failures were confined to AI processing within the post call webhook that runs after a call ends.
The following were affected:
Info extractors: Extracted variables not populated
Call summary: Not generated
LLM-as-a-judge: Did not run
Custom evaluations: Did not run
Post-call SMS actions: Not triggered or sent
If you consume the post-call webhook, the webhook is still fired and delivered but these fields were empty or missing in the payload for affected calls.
What was not affected
Inbound and outbound call connectivity
Live conversation with your agents during the call
Call recordings and audio
Transcripts
Post-call webhook delivery itself (the webhook fired as normal)
Call logs, billing records, and the dashboard
What you should do
If your workflows depend on extracted variables, summaries, or post-call SMS for calls completed between 08:20 and 11:10 UTC on 7 August, treat those records as incomplete. Recordings and transcripts for those calls are intact.
What we're changing
We've completed an internal review of this incident and are making improvements in three areas:
Detection — strengthening monitoring so issues of this type are surfaced and acted on faster.
Change controls — tightening the review and approval process for platform configuration changes.
Impact review — improving how we assess the downstream reach of a change before it is applied.
Resolved
What happened
An internal configuration change restricted which AI models the Synthflow platform was permitted to call. The restriction was applied more broadly than intended and blocked the models used by our post-call processing pipeline.
As a result, AI-generated post-call outputs failed to produce calls completed during the window. The configuration was reverted at 11:10 UTC and all functionality returned to normal immediately.
What was affected
Calls placed and received during this window connected, ran, and completed normally. The failures were confined to AI processing within the post call webhook that runs after a call ends.
The following were affected:
Info extractors: Extracted variables not populated
Call summary: Not generated
LLM-as-a-judge: Did not run
Custom evaluations: Did not run
Post-call SMS actions: Not triggered or sent
If you consume the post-call webhook, the webhook is still fired and delivered but these fields were empty or missing in the payload for affected calls.
What was not affected
Inbound and outbound call connectivity
Live conversation with your agents during the call
Call recordings and audio
Transcripts
Post-call webhook delivery itself (the webhook fired as normal)
Call logs, billing records, and the dashboard
What you should do
If your workflows depend on extracted variables, summaries, or post-call SMS for calls completed between 08:20 and 11:10 UTC on 7 August, treat those records as incomplete. Recordings and transcripts for those calls are intact.
What we're changing
We've completed an internal review of this incident and are making improvements in three areas:
Detection — strengthening monitoring so issues of this type are surfaced and acted on faster.
Change controls — tightening the review and approval process for platform configuration changes.
Impact review — improving how we assess the downstream reach of a change before it is applied.
