Agentic AI support needs resolution metrics beyond contact counts

Resolution should define complex-support AI success

A customer who stops contacting support may still have an unresolved problem. For leaders deploying agentic AI, systems that can take actions to complete a task, that makes reduced contact an incomplete measure of success. The useful test is whether the customer gets the problem resolved, including through an easy transfer to a human who receives the context of the interaction.

Research by Zingtree and CMSWire INSIGHTS shows why this distinction matters. Zingtree’s survey of more than 1,000 U.S. consumers, conducted with CMSWire INSIGHTS for Why Complex CX Is the Real Proving Ground for Agentic AI, examines experiences and perceptions of AI in demanding support situations. Zingtree owns the survey and has a commercial stake in how companies invest in support automation. Readers should weigh its findings and recommendations with that incentive in mind.

The research focuses more narrowly on consumer experience than broader studies of contact-center operations and sentiment. It combines general attitudes toward AI, experiences of unresolved support, and stated willingness to use agentic AI. Each answers a different question: what consumers have experienced, how they respond, and what would make them willing to use AI. Keeping these categories separate helps teams turn the findings into useful deployment criteria.

Those criteria need to connect operating measures to customer outcomes. Ticket containment measures whether a case stays within an automated support process, while deflection measures contacts diverted from another support channel. Both help organizations understand activity and operating costs, but knowing whether a problem is solved requires evidence of resolution. Consumer feedback helps reveal whether lower contact volumes reflect completed work or unresolved demand that could affect loyalty.

Complex support includes repeat contacts, long cases, and high stakes

The gap between reduced contact and resolution matters especially when an issue requires sustained attention or carries serious consequences. Zingtree and CMSWire INSIGHTS define complex support through three qualifying conditions: repeated interactions, resolution taking more than 30 minutes, or high-stakes subject matter. Any one condition is enough for a case to qualify.

Under this definition, complexity extends beyond a question’s technical difficulty. A customer may face a demanding process because they must return repeatedly, spend substantial time getting help, or deal with an error that has serious consequences. Financial and healthcare issues illustrate those stakes. Healthcare inquiries involving sensitive information, insurance claims, and banking transactions also make information handling part of the support problem.

These cases differ from the work that Zingtree and CMSWire INSIGHTS say most customer-experience teams have already automated: password resets, balance checks, and routine tracking inquiries. The organizations describe teams moving toward more valuable and consequential cases, where agentic AI’s effects are less well understood. As the stakes rise, the practical question becomes whether AI can complete the work on the first attempt. Repeated failures can exhaust the customer’s willingness to continue before the problem is solved.

Fewer contacts can mean customers have given up

The survey’s broader support findings show why that willingness matters. Half of consumers surveyed said they had experienced a customer-support issue during the past year that was never fully resolved. Among consumers with unresolved cases, 50% eventually stopped trying to solve the issue. The second figure describes abandonment within the unresolved-case group: people already dealing with unfinished work.

Abandonment can look like a favorable operating result because the company receives no further request. Yet the customer may still need the original problem fixed, leaving a gap between the company’s contact count and the customer’s outcome. Zingtree and CMSWire INSIGHTS identify friction in the support process as a frequent contributor to unresolved problems. When seeking help takes more effort than a customer is willing to spend, unfinished work can disappear from active contact volumes while remaining a business risk.

These support experiences explain why an end to contact is a poor substitute for resolution, while a separate finding shows how consumers respond specifically to automated support. Forty-four percent of respondents in the Zingtree and CMSWire INSIGHTS survey said they try to bypass AI whenever they can. They report already seeking another route to help. A decline in AI interactions can therefore reflect demand moving elsewhere, so teams need to track whether customers complete their task through another channel.

Consumers also report little tolerance for repeated AI failure. More than two-thirds, 68%, of survey respondents said they would abandon AI support after two failed attempts. This is a stated threshold for future behavior, whereas bypassing AI is behavior respondents say they already practice. Zingtree and CMSWire INSIGHTS connect limited tolerance with AI avoidance as related concerns; deployment teams should track failed attempts and channel changes separately to understand how each affects completion.

The survey also identifies skepticism about whose interests support systems serve. Sixty-seven percent of respondents believe companies structure support around their own cost reduction and efficiency instead of solving customers’ problems. Even automation that works as designed faces this challenge, because customers judge the process’s purpose through the help they receive. Efficiency remains a legitimate objective, but teams need to show how it contributes to completed customer tasks.

Easy human access is part of willingness to use AI

Completion is one explicit condition for AI acceptance. Sixty percent of respondents in the Zingtree and CMSWire INSIGHTS survey said they accept interacting with AI if it completely resolves their issue. For deployment teams, this puts the outcome at the center of the interaction: customers need to be able to do what their support request was meant to enable. An interaction that leaves the original problem open falls short.

Human access is another condition, particularly for demanding cases. Sixty-six percent of survey respondents said they would use agentic AI for complex issues if reaching a human were easy. These respondents see AI use and access to a person as compatible parts of the service. Making human help difficult to obtain can therefore weaken a condition under which consumers say they are willing to begin or continue with AI.

Easy access also needs a useful transfer process. Sixty-four percent of survey respondents agreed that a smooth AI-to-human handoff that preserves context would make them more willing to use AI again. Preserving context means giving the human relevant information from the preceding interaction so support can continue from the work already done. Because this finding concerns stated willingness to return to AI, the transfer itself matters when evaluating adoption.

Context determines whether the customer’s earlier effort remains useful. When a human receives the issue history and information already collected, the customer can continue seeking resolution without rebuilding the case. This continuity can reduce the work of pursuing help through repeated contacts, a lengthy case, or a sensitive problem. Teams should therefore examine what information reaches the person taking over, alongside how easily the customer reaches them.

Human involvement can also take a different form. A separate 66% finding concerns consumers who prefer dealing with human agents over AI; Zingtree and CMSWire INSIGHTS say this group remains comfortable with those agents using AI tools to help them. This preference differs from willingness to use agentic AI when human access is easy. It shows how a preference for human contact can coexist with acceptance of AI working behind the scenes.

These preferences call for separate evaluation of three service arrangements: AI handling the customer interaction, AI transferring the customer to a person, and a person using AI tools while providing service. Each assigns customer contact and responsibility differently. Teams should measure acceptance and resolution for each arrangement, because willingness to use one gives them a basis for evaluating that particular form of service.

Evaluate resolution across AI and human support

The practical response is to evaluate resolution across the full sequence of assistance. Record successful AI-only resolution separately from successful resolution involving a person, while recognizing that both can meet the customer-outcome objective. Keep containment and deflection as operating measures, and assess them alongside whether the customer’s work is complete. Teams can then distinguish efficient completion from cases that leave the support process unfinished.

Evaluation should also examine the conditions consumers attach to AI acceptance: easy human access and a transfer that retains relevant context. These deployment recommendations follow from the survey’s willingness findings. Teams can assess whether customers can reach a person, what information accompanies the transfer, and whether the issue is ultimately resolved. For customers who have already invested time in a case, a poor handoff adds work to an unfinished task.

The consequences of failure grow as AI takes responsibility for sensitive or high-stakes interactions. Zingtree and CMSWire INSIGHTS describe how greater capability expands exposure to hallucinations, incorrect information generated by AI, and mishandling of customer data. In their account, failures consume consumers’ limited willingness to retry, while weak controls and difficult human access make the experience worse. Teams assessing sensitive work should therefore examine controls for incorrect responses and improper information handling together with the route to human assistance.

These controls give executives a practical way to connect operating benefits with consumers’ demand for complete resolution. Before expanding automation, teams should determine which tasks the system can safely complete and how support continues when it reaches a limit. They should also evaluate consequential failures during deployment, including their effects on customers’ willingness to keep seeking help. Efficiency then becomes a decision about the work the system can complete safely.

The risks of expansion also extend to future AI use. Zingtree and CMSWire INSIGHTS characterize current consumer openness as a “grace period.” They argue that expanding AI into high-value cases while prioritizing deflection and restricting human assistance can lead to customer loss today and broader AI avoidance in the future. This business risk gives teams a reason to examine how failed interactions affect continued AI use and customer relationships before extending automation into more consequential work.

Information handling requires its own examination as automation expands. Zingtree and CMSWire INSIGHTS also describe AI learning from consumer information and inputs associated with successful and unsuccessful interactions. Deployment teams should examine how their systems use those inputs, including whether they enter a learning process and how sensitive information is handled. This review belongs alongside assessment of the customer-facing interaction because information collected during support can have uses beyond completing the immediate request.

Key highlights

• Define complex support broadly: Cases qualify through repeated contacts, resolution times exceeding 30 minutes, or high-stakes subject matter. CX teams evaluating automation need to assess customer effort and consequences alongside technical difficulty.
• Track unresolved demand: Half of surveyed consumers with unresolved cases eventually stopped seeking help. Support operations owners need to pair contact-volume metrics with resolution evidence and track failed AI attempts and channel changes.
• Make human access part of AI service: Sixty-six percent of respondents said they would use agentic AI for complex issues if reaching a human were easy. CX teams need to test access to human help and preserve relevant case history during transfers.
• Set expansion criteria around safe completion: Deployment owners need to record AI-only and human-assisted resolution separately, test controls for incorrect responses and sensitive-data handling, and define how support continues when AI reaches its limits. These checks connect operating gains to completed customer tasks.