AI Agent Cuts Median Ticket Resolution Time From 22 Hours to 22 Minutes
An AI agent in a support service cut the median time to actual ticket resolution from 21 hours 52 minutes to 22 minutes. But formal closure still took 49 hours, and after an update two out of ten responses became faster and worse. The main lesson of the case: speed and a "resolved" status cannot be taken for quality without checking the substance of the response.
AI-processed from Habr AI; edited by Hamidun News
The figure looks almost indecently good: after the AI agent launched, the median time to actual resolution of a request fell from 21 hours 52 minutes to 22 minutes, roughly 59 times faster. But these are exactly the results that are most dangerous to accept without scrutiny.
The team studied 1,876 closed requests before launch and 3,078 after it. This is not a controlled experiment: topics, traffic volume and the number of mass incidents differed. So the case does not prove that the agent produced the entire effect. It does, however, show well how to measure working automation and where attractive metrics start to lie.
A closed ticket does not mean a solved problem
In the CRM, a request could remain open for two days after the user had received a working instruction within 12 minutes. The opposite also happened: the ticket status changed even though there had been no substantive answer yet.
So the team defined the resolution point as the last useful response from the agent or an engineer before the technical tail. A user's thanks, automatic closure, a status change, a log request and an internal comment did not count as resolution on their own.
The difference turned out to be huge. After launch, the median actual resolution time was 22 minutes, while formal closure took 49 hours 2 minutes. The long tail did not disappear: 90% of requests were resolved within 76 hours 36 minutes. The median alone would easily have hidden the hard cases.
Of the 3,078 closed tickets, the agent resolved 1,647, handed another 743 to engineers, and the remaining 688 did not go through the agent pipeline. However, the 53.5% share should not be read as "the agent replaced the engineer in half of cases." It was given topics in advance that had clear diagnostics, available data and a safe scenario. Complex cases more often stayed with humans.
Faster sometimes means worse
After one update, the team re-ran ten old tickets without sending responses to customers. In five cases the new version answered better, in two about the same, in one the result was mixed, and in two it got worse. Both degraded answers appeared faster than the previous ones.
In one example, the agent stopped correcting obvious typos in object codes. Strict search protected against false matches, but did not understand the difference between the digit 0, the Latin O and the Cyrillic O. Formally the behavior became safer and more reproducible, but for the user it became less useful.
The fix was a bounded search over permissible substitutions. The system checks only predefined variants and accepts a correction if exactly one match is found. If there are several matches, the agent asks for clarification. This is a practical compromise between the model's guesswork and useless strictness.
What to check in production
The first rule is to separate a tool's result from its status. An empty response can mean there is no record, but also a timeout, service unavailability, insufficient permissions or incomplete data. The agent must not turn a failed check into a confident diagnosis.
The second is not to let the model pass off a plausible technical detail as an established fact. The name of a table, field or API may be included in a response only when it has been found in current data, documentation or a verified schema.
The third is to maintain a regression set of real requests. It should be expanded after each new class of error and run whenever the prompt, tools, search or order of sources change. The correctness of the service action and the quality of the response to the person should be evaluated separately.
The practical conclusion is simple: deploying an agent does not end with the release and a high share of automatic resolutions. You need the median and 90th percentile of time to a useful answer, analysis of escalations, explicit tool statuses and ongoing regression checks. Otherwise the team risks not eliminating the error, but merely teaching the system to make it faster.
- Nov 12, 2026Claude in Chrome requires approval for every click on category 3 sites
- Nov 12, 2026Norwest raises record $4 billion fund, betting on agents, voice and physical AI
- Nov 12, 2026AWS shows how 100% retrieval precision can hide an AI agent's failure
- Nov 12, 2026Garage Eight dropped peer approvals for 69 MRs in an internal service using AI review
Key Questions Answered
What are the key details on A closed ticket does not mean a solved problem?
In the CRM, a request could remain open for two days after the user had received a working instruction within 12 minutes. The opposite also happened: the ticket status changed even though there had been no substantive answer yet. So the team defined the resolution point as the last useful response from the agent or an engineer before the technical tail. A user's thanks, automatic closure, a status change, a log request and an internal comment did not count as resolution on their own.
What are the key details on Faster sometimes means worse?
After one update, the team re-ran ten old tickets without sending responses to customers. In five cases the new version answered better, in two about the same, in one the result was mixed, and in two it got worse. Both degraded answers appeared faster than the previous ones. In one example, the agent stopped correcting obvious typos in object codes. Strict search protected against false matches, but did not understand the difference between the digit 0, the Latin O and the Cyrillic O.
What to check in production?
The first rule is to separate a tool's result from its status. An empty response can mean there is no record, but also a timeout, service unavailability, insufficient permissions or incomplete data. The agent must not turn a failed check into a confident diagnosis. The second is not to let the model pass off a plausible technical detail as an established fact. The name of a table, field or API may be included in a response only when it has been found in current data, documentation or a verified schema.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.