A case published on x.ai describes 204 tickets handled in about ten minutes after human approval. The decisive detail lies in the limits imposed on the approved batch.
On 20 October, Chris Bowie and Isaac Hopkinson describe on x.ai two agents used for support. IncidentWatch spots clusters of similar tickets; IncidentSweep prepares their handling once the incident is resolved. The authors report that the latter handled 204 tickets in about ten minutes during an incident on 18 September, with review and approval by a support engineer.
This is an internal account of experience. The figure measures neither customer satisfaction nor the error rate of the service as a whole. Its interest lies in the way of working the authors detail.
The approved scope is not widened afterwards
The engineer specifies the tickets to consider, the exclusions and the expected message. The agent presents the proposed batch and the exact text before sending. It sets aside the cases it cannot confidently tie to the incident. According to the article, it adds no ticket after approval and stops in case of failure, reporting what happened.
This limit gives precise content to the expression “human approval”. The person approves a defined population and a defined reply. They do not grant a general permission to continue as the tool’s matches see fit.
A lead for customer-relations incidents
For a house facing a delivery incident or a digital outage, this partitioning offers a design lead. Requests that look identical can conceal a personal situation, a prior commitment or a complaint requiring separate handling.
A relevant test should therefore examine the excluded cases as much as the handled ones. Speed of execution makes sense when the team can retrieve the message sent, its recipient and the approval that authorised it. These are the proofs that would allow an assessment of an adaptation to a house’s service.
Sources — editorial verification
Using Grok Bot to detect and respond to incidents — 20 October 2026. Accessed 26 October 2026.
Read also on Luxsure: Team Bots: deciding what a team can pool · Claude Sonnet 5.5: cost also depends on the work requested · Mistral in Munich: bringing research closer to industrial know-how · GPT-6.1 Sol: what the price of reused context changes
Cette publication est également disponible en :
