A bank in Pennsylvania said in October that its agents helped close some business loans in seven days. The usual wait had been 30 to 60 days. That gap is the best example I know of agentic AI cycle time. It is also a warning about how easily the number bends.
Agentic AI cycle time is the clock from a customer’s request to a finished result the customer can use. In lending, that runs from application to decision. In claims, it runs from first notice to payment. It was the first metric in my post on the KPIs that matter, and readers asked for more on it.
Three real cases from the past two years show the same thing. When the clock covers the whole outcome, the gain holds. When it covers only the agent’s slice, the gain can fade, because the waiting simply moves to someone else’s desk.
What the number counts
The clock starts when the customer asks, not when the agent picks up the file. It stops when the outcome lands, not when the agent writes its last line. Everything in between counts, including queues, handoffs and the human who signs off.
That sounds obvious. Still, most dashboards report something smaller. They show how long the model took to draft a memo or read an invoice. That is the part the vendor can see. A three-minute memo that then sits two days in a review queue has not changed the outcome for anyone.
The number also ties to sales in a way savings rarely do. A faster loan decision means fewer buyers who walk to a rival. So the metric belongs to the business owner, not just the tech team. It sits next to the outcome thinking behind my definition of agentic AI: intent in, outcomes out.
Where the clock starts and stops
Each process needs one start event and one stop event, set before the pilot starts. The table shows the pairs I would use for four common jobs.
| Process | Clock starts | Clock stops |
|---|---|---|
| Business lending | Complete application received | Decision sent to the borrower |
| Insurance claims | First notice of loss | Payment issued |
| Invoice processing | Invoice arrives | Invoice posted and approved |
| Customer service | Contact opened | Issue solved, no repeat contact in 7 days |
The last row matters most. A chat that closes fast but comes back two days later never really stopped. So a repeat contact reopens the original clock instead of starting a fresh one.
How to measure it without fooling yourself
First, stamp every handoff, because agentic AI cycle time is only as honest as its timestamps. Each time work moves between the agent, a queue and a person, the system logs a time and an owner. Those stamps belong in the same audit trail that already explains what the agent did and why.
Then split the clock into three parts: agent time, wait time and human review time. Agent time is often the smallest. Wait time is often the largest, and most of the gain hides there. Review time tells you whether people trust the agent’s work on first pass.
Next, report the median and the slowest tenth of cases, never the average alone. A few stuck files drag an average around. Meanwhile the slow tail is exactly where customers give up and leave.
Last, compare it with a baseline from the same product, channel and season. A pilot on the cleanest files will always look fast. That is picking, not speed.
Research from Anthropic shows why no one can skip the review part. A November 2025 study of 100,000 conversations found that Claude cut the time on typical tasks by about 80 percent. Its authors were plain about the limit, though. They said the estimate could not count time spent outside the chat, such as checking the quality of the work. That outside time is exactly what an end-to-end clock catches.
A claims queue that cleared
In July 2025, the Australian arm of a large European insurer put seven agents on food spoilage claims under 500 Australian dollars. These are small claims, like a fridge of food lost when a storm cut the power. According to the insurer’s own account, such claims could take four days or more.
The agents split the work: plan, check cover, check weather, check fraud, price the payout and audit. The whole run takes under five minutes before it reaches a human. Then a claims expert reads the audit summary and decides the payout. The insurer says no payout is ever left to the agents.
The insurer says payment now takes a day, or even hours. It cites an 80 percent cut in time from claim to payment. Those figures are company-reported. Still, the insurer measured the right span, from claim to payment, with the human step inside the clock.
A loan file that waited less
The bank from the top of this post built its agents on its own data and systems. Engineers sat with the bankers who own the work before anything went live. In a second-quarter pilot, select commercial loans closed in seven days, according to reporting on its October earnings.
Its chief executive called it an 85 percent cut in time to readiness to close. That stop event needs a close look. Readiness to close is not funding, and it is not the borrower’s signature. It is still a strong result. But anyone comparing banks should ask where each one stops its clock.
A fast answer that came back
In early 2024, a Swedish payments company said its AI assistant handled 2.3 million chats in its first month. That was about two-thirds of its customer service volume. Resolution time fell from 11 minutes to under 2, and repeat inquiries dropped 25 percent, according to its AI provider.
By May 2025 the story had turned. The company’s chief executive said it would hire people again so customers could always reach a human, Bloomberg reported. He said cost had weighed too heavily in the push and quality had suffered as a result.
The two-minute number was real. It just measured the conversation, not the problem. Once quality slipped, the clock that customers felt kept running after the chat closed.
What I measure now
My own rule is simple. If the clock stops when the agent finishes, I do not trust the number. I want it to stop when the customer has the outcome. Review time and rework get charged back to the first request.
This is a governance call first and a design call second. Someone has to own the start and stop events, and someone has to sign off on the baseline. Then the design does the rest. That means a timestamp at every handoff, a review step inside the clock and an eval suite that checks first-pass quality every night.
I’ll admit I am less sure about fraud, where a slow answer is sometimes the right one. There, a fast clock can reward the wrong behavior. For most regulated work, though, agentic AI cycle time from end to end is the first number I would show a board.
It also tells you how far a team has moved on the Five-Stage Roadmap. A team that can only report the agent’s slice is still piloting. A team that can show agentic AI cycle time from request to outcome has earned the right to widen the agent’s autonomy tier.
Where does your clock stop today: when the agent finishes, or when the customer has the outcome?
Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
My Books: Three books on agentic AI for regulated enterprises: Intent In, Outcomes Out, Earned Autonomy and The Governed Loop. Reading order, formats and buy links are on one page.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.