How CX teams can get the human and AI work split right

Set AI authority based on risk and judgment

74% of consumers prefer human channels for issue resolution and technical support, according to Qualtrics. That finding sets a clear constraint for CX automation. Customers still value human access when the outcome matters or the issue becomes difficult.

The right unit of automation is a bounded task. AI performs well when inputs are clear, rules are defined and the expected outcome is predictable. It can classify customer intent, retrieve approved information, summarize conversations, provide order updates and guide customers through standard processes. These tasks give the system a clear operating boundary.

Authority should decrease as the consequences increase. For a refund, credit or account change, AI can collect information and recommend an action. A human can approve cases above a defined financial or risk threshold. For disputed charges, suspected fraud, unusual accommodation requests and serious service failures, employees should retain decision authority. These cases require policy interpretation, discretion or an assessment of circumstances that may be absent from the customer’s explicit request.

Chris Rozum, founder of Insite Managed Solutions, summarized the boundary clearly: “If a human’s judgment is changing the result, keep a person there even if the task looks routine at first glance.”

This distinction matters because process maps can hide judgment. A workflow may appear repetitive while experienced employees routinely detect conflicting information, infer an unstated concern or apply an exception. Automating the visible steps can remove the judgment that makes the process work.

Executives should therefore define autonomy by use case before deployment. Each AI workflow needs a clear scope, an escalation condition and an approval threshold. The core question is simple: What happens if the system gets this decision wrong? Financial loss, compliance exposure and damage to customer trust justify tighter human control.

High contact volume remains useful for identifying automation opportunities. It should not determine AI authority by itself. The stronger operating model combines automation potential with the cost and reversibility of an error. This allows AI to remove routine work while concentrating human judgment where it has the greatest value.

Design AI-to-human handoffs as part of the resolution process

74% of consumers find it frustrating to repeat their story to different agents, according to Zendesk’s 2026 CX Trends research. AI escalation creates the same problem when customer context disappears during the transfer.

An AI-to-human escalation can be the correct system behavior. AI has reached a predefined boundary, and the workflow has moved the case to someone with greater authority or judgment. Success depends on whether the employee can continue the resolution process immediately.

Three pieces of information should move with the customer: the conversation history, actions already taken and the reason for escalation. Employees also need visibility into relevant AI decisions. They should be able to determine what information the system used, what it recommended or changed, and why it transferred the case.

Sergey Matikaynen, co-founder of GoGloby, describes failures in this process as “context starvation.” He said: “A successful handoff preserves the full conversation history and reasoning trace, giving human agents step-level visibility into what the AI did and why it escalated.”

That visibility has operational value. An employee who receives a conversation transcript without understanding the AI’s previous actions may have to investigate the interaction from the beginning. This increases handling time and customer effort. It can also make automated errors harder to detect and correct.

Context alone is insufficient when the employee lacks authority. A customer can reach a human and still remain stuck if that person cannot reverse an incorrect automated decision, grant an appropriate exception or resolve the underlying account problem. Escalation design therefore needs to connect information with decision rights.

CX leaders should test the entire handoff as one workflow. Measure the time required to reach an employee. Check whether customer identity, history and previous actions transfer correctly. Confirm that the employee can see why the AI escalated. Then verify that the employee has the permissions needed to resolve the case.

This changes how leaders should interpret escalation rates. A transfer can show that risk controls are working as designed. The more important measures are escalation quality, customer effort and final resolution. A well-designed human-AI operation knows when automation should stop and ensures that the next person has the context and authority to finish the work.

Preserve human access and make AI actions accountable

89% of respondents believe companies should always offer the option to speak with a human, according to SurveyMonkey’s 2025 consumer research. For CX leaders, human access should be a defined part of the operating model.

Customers need a clear path to an employee when automation cannot resolve an issue, when they dispute an automated outcome or when the case involves sensitive circumstances. The route should be visible and easy to use. Forcing customers through repeated automated steps increases effort at precisely the point where the system has reached its limits.

Human access also needs decision authority behind it. Employees handling escalations should have permission to review AI actions, correct errors, apply appropriate exceptions and resolve the underlying problem. Without these permissions, the organization simply adds another interaction to the customer journey.

Accountability requires a reliable record of automated decisions. Businesses should be able to determine what information the AI used, what action it took, whether an employee approved or modified that action and what happened afterward. These audit trails become especially important when AI can issue credits, alter account information or take other actions with financial, regulatory or customer consequences.

This record also improves operations. Teams can identify recurring errors, determine whether problems originate in knowledge, instructions, routing or system permissions, and adjust the workflow. An isolated correction fixes one customer case. A reviewable history can reveal a repeated failure affecting many cases.

The level of control should match the potential impact. Low-risk actions with well-defined rules can receive greater autonomy. Financial decisions, account changes and policy exceptions may require thresholds or explicit approval. Sensitive and regulated decisions warrant stronger human authority and review.

For executives, the key design question is accountability: who owns the outcome when AI takes an action? That responsibility must remain clear across CX, operations, technology and risk teams. Automation can execute decisions at scale, while the business remains responsible for the customer outcome.

Measure resolution quality across the human-AI team

AI containment rate answers a narrow question: how often did the system complete an interaction without transferring it to an employee? CX performance requires a broader view of what happened afterward.

A high containment rate can exist alongside repeat contacts, complaints and unresolved problems. An interaction may end inside an automated channel because a customer gives up or chooses another channel later. Executives should therefore connect automation metrics to final customer outcomes.

First-contact resolution shows whether the underlying issue was solved without further customer effort. Repeat-contact rate identifies customers returning with the same problem. Customer-effort measures show how much work customers had to do to obtain a result. Escalation quality assesses whether the case reached an appropriate employee with sufficient context and authority. Error-recovery measures track how quickly incorrect AI actions are detected, corrected and explained.

Employee workload belongs in the same measurement framework. AI can absorb routine questions and leave employees with a greater concentration of exceptions, sensitive interactions and difficult cases. Those cases can require more investigation, discretion and emotional effort. Staffing models and performance targets should reflect this change in case mix.

The measurement challenge also includes aspects of CX that conventional dashboards can miss. Sarup Paul, vice president of product management for CRM and industry workflows at ServiceNow, said: “Executives are making decisions based on available metrics: NPS, CSAT, resolution rates, call deflection. What those metrics don’t capture is whether the customer felt heard or whether the interaction felt empathetic.”

ServiceNow’s global study of 34,000 customers, service representatives and executives shows the size of that gap. Half of customers cited a lack of empathy and understanding as their top frustration. Only 23% of executives identified it as a top challenge.

That difference becomes more important as AI changes the distribution of work. Automated systems increasingly handle transactional requests, while employees receive situations involving exceptions, previous failures and customers who need reassurance or flexibility. The human interaction can therefore represent a disproportionately important part of the overall customer relationship.

Executives should manage human and AI service as one operating system. Containment remains useful for understanding automation activity and capacity. Resolution, repeat contacts, customer effort, escalation quality, error recovery and employee workload reveal whether that automation creates better outcomes.

The target should be durable resolution. AI creates value when it reduces routine effort while preserving effective recovery for difficult cases. That standard gives leadership a clearer basis for judging whether automation is improving CX, reducing operating cost and protecting customer trust.

Redesign frontline CX roles for judgment and AI oversight

AI changes the work that remains for customer service employees. Routine requests increasingly move to automated systems. Human agents then spend a greater share of their time on exceptions, ambiguous cases, escalations and relationship recovery.

That shift raises the value of judgment. Employees need to recognize incomplete or incorrect AI responses, investigate conflicting information and decide when policy requires an exception. They also need to correct bad information before it affects more customers. Strong communication remains important because many escalated customers arrive after an automated process has already failed to resolve their issue.

Jourdan Hathaway, chief business officer at General Assembly, described this problem based on the company’s contact center experience: “When we implemented AI in our contact center, we quickly realized the system couldn’t adapt to context it wasn’t trained on. Any deviation from the typical script broke the experience.”

Hathaway said ambiguous situations were routed incorrectly, reducing resolution rates and customer satisfaction. The experience identifies a central constraint in CX automation: an AI system operates within the information, instructions and workflows available to it. Unexpected circumstances therefore require an escalation process designed around human judgment.

Training must evolve with that responsibility. Agents need practical experience reviewing AI-generated summaries and recommendations, identifying missing or contradictory context, applying policy and documenting automation errors. They should understand when they can override an AI recommendation and how to escalate cases that exceed their own authority.

Supervisors face a related change. They increasingly need to manage the performance of employees and AI-enabled workflows together. This requires skills in quality assurance, AI oversight, coaching and root-cause analysis. When an interaction fails, managers need to determine whether the cause was an employee decision, incorrect knowledge, poor routing, inadequate system permissions or another workflow issue.

Karl Phenix, VP of Global Microsoft Strategy and Go to Market at TTEC Digital, said customer service representatives are increasingly handling “complex cases, exceptions, escalations, and relationship-building activities.” He added that critical thinking, decision-making, quality assurance, AI oversight and coaching become more important as AI becomes embedded in operations. In his description, the future agent becomes “more of a problem solver and customer advocate.”

This change also affects workforce planning. Lower routine contact volume can leave employees with a higher concentration of difficult, sensitive and time-consuming interactions. Average handle time may rise because the remaining cases require more investigation and discretion. Case volume can also become a weaker indicator of individual contribution when complexity varies substantially between interactions.

Executives should align performance management with the new work mix. Measures such as resolution quality, successful exception handling, escalation outcomes and error identification can provide a stronger view of employee contribution. Coaching should develop decision quality and AI supervision alongside customer communication.

The broader workforce decision concerns capability allocation. AI can perform standardized work at scale. Skilled employees can focus their time on cases where context, discretion and accountability affect the outcome. Organizations that redesign training, permissions, staffing and performance measures around that division of work can capture automation efficiencies while maintaining effective human support.

Key highlights

• Set AI authority by risk: Automate bounded, predictable tasks and keep human decision authority where financial, compliance or trust risks rise. Define approval thresholds and escalation conditions for every AI use case.
• Make handoffs preserve context: Transfer conversation history, prior AI actions and the reason for escalation with every case. Give employees enough authority to correct decisions and complete the resolution.
• Keep human access and accountability clear: Give customers a visible route to a person when automation reaches its limits. Maintain audit trails that show what AI did, what information it used and who approved or changed consequential actions.
• Measure resolution over containment: Track first-contact resolution, repeat contacts, customer effort, escalation quality and error recovery alongside AI containment. Evaluate the human-AI operation as one system tied to customer outcomes.
• Redesign frontline roles for complex work: As AI absorbs routine requests, develop employees for judgment, exception handling, AI oversight and customer recovery. Update training, staffing and performance metrics to reflect the higher complexity of human-managed cases.