The Agent Took the Easy Calls

An automated support system can lower the total cost of customer service while making each remaining human contact slower and more expensive.

That is not necessarily a contradiction. Remove routine billing questions and documentation lookups from the human queue, and what remains will be harder: locked accounts, disputed charges, missing information and requests that require privileged tools or someone with authority to act. The average human ticket rises because the easy tickets are gone.

The harder question is what happened before the person entered the conversation. A customer may arrive after the automated system failed to solve the problem, delayed the handoff or made them repeat themselves. The worker inherits both the unresolved task and the customer’s experience of not being resolved.

The queue changes before a person sees it

On October 21, Anthropic introduced Claude Duet, a research-preview product that checks connected work and decides what needs the user, what can wait and what should be checked first. Duet is not a customer-support study, and Anthropic has not published field outcomes showing what happens to the work it leaves behind. It is a clear product signal: agents are being sold not only as task performers but as systems that decide which tasks deserve a person.

Customer support has been testing the consequences of that decision for years.

Bloomberg reported in June 2025 on Jessica Lindsey, who handled American Express contacts for Concentrix. Some callers reached her after automated triage had already frustrated them. Her required opening script could sound like another machine, so she changed her delivery to prove she was human. Some callers yelled or hung up, and the experience sometimes left her in tears in her Oklahoma home office. The publisher-distributed audio is available through iHeart. Bloomberg said Concentrix did not respond and American Express declined to comment.

Lindsey’s experience does not describe every contact center. It does show the handoff from the worker’s side. The person receives the unresolved problem and the emotional condition created by the path that came first.

Easier contacts disappear from the average

Scott Falconer, who works in applied AI and agent success engineering, described a related pattern in email correspondence with the show. He spoke from personal observation, not for his employer, and kept the deployment unnamed.

Falconer said routine questions that could already be answered from documentation moved to an agent, including general billing and invoicing questions. Other contacts exposed missing or hard-to-find documentation. Those became repair work: improve the documentation so the agent could answer the next similar question.

The requests increasingly reaching engineers required privileged access, internal tools or the authority to take an action the customer could not perform. Falconer observed that average human time and cost per ticket rose while total time and cost, including the vendor cost, fell.

The operation tracked customer satisfaction, dissatisfaction, escalations and deflection. Falconer used deflection to mean a question handled through self-service or the agent without a human interaction. His warning was direct: “A ticket that never reaches a human isn’t automatically a successful resolution. You still need evidence that the customer got what they needed.”

His account is one person’s observation of one unnamed deployment. A randomized field experiment on Alibaba’s Taobao marketplace provides a measured comparison. Alibaba supplied the workplace and data, and one author was Alibaba-affiliated.

The randomized results found that AI-eligible chats were 16.8 percent shorter, while customer ratings fell by 0.412 points on a five-point scale.

A separate, non-randomized analysis in the same paper cannot establish causation. It found that technical-limit handoffs preserved ratings and repeat-contact rates relative to matched human-only chats. Emotional escalations tended to happen later and, in those matched comparisons, were associated with six percentage points more repeat contacts within seven days and ratings 0.928 points lower. Grit Daily reported the results in October.

Timing changes the job. A person called because the system lacks technical ability can still solve the problem. A person called after the customer has become angry must first repair the interaction, then solve the original problem. Both may appear on a dashboard as escalations. They are not the same work.

Customer Experience Dive reported in May that routine issues are disappearing, representatives are receiving repeated edge cases, average handling time is rising and workers are losing easier calls that once provided a mental reset. The sources quoted advise companies on customer experience, so the account is not a neutral workplace audit. It does match the mechanism visible in Falconer’s observation, the Taobao results and Lindsey’s experience.

Deflection is not resolution

A useful dashboard should measure whether the customer was resolved, whether they returned with the same problem, when a handoff occurred and what kind of work reached the person. It should also track what authority the worker needed and whether the automated path made the contact harder before surrendering it.

Otherwise the dashboard rewards an agent for keeping a customer away from a person even when all it accomplished was making the eventual conversation worse.

Finding where the human sits in an agent workflow is not enough. By the time the contact arrives, the system may already have selected the hardest work, withheld the authority needed to finish it and changed the emotional condition in which the person has to begin.

Listen to Episode 71

Episode 71, “The Agent Took the Easy Calls”, is available now.

Download the episode or subscribe to the feed.

Sources

If you work in customer support after automation removed the routine contacts, tell the show what changed about the cases, authority and customers that reached you. Email [email protected]. Use the subject line “The calls left for people.” Anonymous notes and source-protection requests are welcome.