What comes after AI takes over the easy work

AI concentrates complexity in human work

Twenty years ago, routine requests filled most of a contact center agent’s day. Password resets and billing questions followed predictable scripts. AI can now absorb much of this work. That changes the economics of automation, but it also changes the job that remains.

Human agents increasingly receive the cases that automation cannot resolve with confidence. These interactions are less predictable and often more emotional. They can involve conflicting customer needs, company policies, operational constraints, and incomplete information. The agent must make a judgment rather than retrieve a standard answer.

Bryan Stoller, Vice President and Global Head of Customer Care, Contact Centers, and Solutions at United Airlines, described a roughly 20-year shift in this workload. Routine interactions have declined as a share of agents’ work, while moderately and highly complex interactions have grown. Greater AI adoption will accelerate that change.

Consider a family stranded at an airport after a weather delay. The standard policy may be to book them onto the next morning’s flight. The actual situation becomes harder when medication is inside checked luggage, a child is distressed, and nearby hotels are full. As Stoller explained, most customers in these situations are seeking a reasonable decision based on their circumstances.

This changes how executives should evaluate contact center productivity. As AI handles a larger share of predictable demand, average human interactions can become harder. Handle times may increase. Agents may need more authority. Training must place greater weight on judgment, communication, and exception management.

The operating constraint therefore moves toward the quality of human decisions in difficult cases. Companies that automate routine volume without redesigning human roles risk creating a service organization optimized for work that is disappearing. Workforce planning, quality management, training, and performance measures should reflect the higher complexity of the remaining workload.

The goal is clear: use AI to remove predictable work while increasing the capability of people who handle unpredictable work. Automation creates more value when the human operation evolves with it.

Build a “Reasonable department” for exceptional cases

Complex cases need a defined operating model. Stoller calls this the “reasonable department.” It is a capability embedded across the organization rather than a separate organizational unit.

His model has four parts: detect, route, resolve, and learn. Together, these steps define how a company identifies unusual situations, assigns the right expertise, gives employees the ability to act, and uses the outcome to improve the wider system.

Detection comes first. Both AI and conventional workflow systems need to recognize when a request has moved outside predictable boundaries. Confidence matters here. An AI system that cannot reliably resolve a case should recognize that condition and transfer the interaction with its context intact. Effective automation therefore includes effective handoffs.

Routing then becomes a capability-matching problem. Sending a difficult case to the next available employee may optimize queue speed while producing a poor outcome. Complex interactions should reach people with the appropriate experience, authority, language skills, product knowledge, or ability to manage emotionally difficult conversations.

Resolution requires more than routing. Agents need customer history, operational information, usable tools, clear decision boundaries, and enough authority to make a practical decision. Stoller summarized the requirement directly: “You cannot constrain them by black and white policy.” For executives, this means defining where discretion is permitted and how the organization controls the resulting financial, regulatory, and customer risks.

The final step is learning. Difficult cases contain information about weaknesses in the business. Repeated exceptions can expose a poor policy, a broken process, missing self-service functionality, inadequate training, or an AI system that has learned an incomplete set of customer needs. Capturing these cases creates a feedback loop into operations, policy, training, and AI development.

This model has an important implication for AI governance. Escalation should be designed as part of the AI system from the beginning. Leaders need explicit thresholds for when AI can act, when a human must review a decision, and which employees have authority to resolve exceptional situations.

Stoller captures the broader design requirement: “This is about not designing our organizations for the work that AI takes away. This is about designing our organizations for the work that AI leaves behind.”

That is the stronger operating model. As routine volume moves toward automation, competitive service performance will increasingly depend on how well the organization detects complexity, directs it to the right people, enables sound decisions, and learns from every exception.

Start AI deployment with a defined business problem

AI investment should begin with a specific customer or operational problem. The first question is simple: why is the customer contacting the company? Contact data can then identify which interactions occur most often, which follow predictable steps, and which can be automated safely.

This approach gives executives a practical way to prioritize investment. High-volume, low-complexity interactions usually offer clearer opportunities because the process can be defined and outcomes can be measured. Successful automation can reduce repetitive work and release experienced employees for cases that require more judgment.

Jessica Gupta, Chief Operating Officer at InfoPay, used customer behavior to identify such an opportunity. InfoPay discovered a “fairly large” group of customers who preferred voice communication. The company focused its automation effort on a high-volume process in which customers primarily needed a simple confirmation.

That narrow scope produced a useful business outcome. InfoPay was able to contain a significant portion of those interactions through automation. Experienced agents then had more capacity to contribute insights elsewhere in the business. The technology addressed a clearly observed demand pattern instead of forcing customers toward a different channel.

Customer contact data is particularly valuable in this process because call volume often reveals problems elsewhere in the organization. A large number of calls about the same issue can signal an inefficient workflow, poor digital self-service, fragmented customer data, or an access problem. Executives should identify the underlying cause before deciding where AI should operate.

Aaron Johnson, Interim Chief Marketing Officer at Penn Medical, described this challenge in healthcare. Growing patient demand and unsustainable call volumes pushed the organization to redesign its access operations. Previous mergers and acquisitions had also created integration problems. Agents working in one part of the system could not always schedule appointments in another part, which prevented some patients from reaching care.

That example matters because AI performance depends on the systems and processes around it. An automated assistant still needs access to accurate schedules, customer or patient records, business rules, and downstream systems. A weak integration layer limits what the AI can accomplish.

For C-suite leaders, the investment sequence should therefore start with demand analysis. Identify the largest contact drivers. Determine which processes have predictable outcomes. Map the systems required to complete each transaction. Then select AI where automation can produce a measurable customer or business result.

This keeps the business case concrete. AI deployment becomes tied to outcomes such as resolved contacts, improved access, reduced repetitive workload, or better use of skilled employees.

Use small pilots and fast iteration to control AI deployment risk

AI systems encounter real customer behavior only after deployment. Customers use unexpected words, change topics, provide incomplete information, and make requests that designers may have missed. Voice systems face additional variation in phrasing and conversational structure. Controlled pilots expose these conditions while keeping operational risk manageable.

Jessica Gupta, Chief Operating Officer at InfoPay, described an intentionally small starting point. “We only had 20 calls,” she said. That early test produced one success. The team treated the result as information for further development rather than a reason to expose a large share of customer traffic immediately.

The principle is to increase deployment volume as evidence improves. Teams can begin with a narrow use case and limited traffic, measure outcomes, inspect failures, adjust the system, and test again. Each stage creates operational evidence for the next expansion decision.

Neville Letzerich, Chief Marketing Officer at Talkdesk, advocated this experimental approach. “Get started. Try something. Don’t go overboard. Be intelligent about it. Go get a win, leverage that, or lose and say, ‘Wow, that didn’t work,’ try again.”

For executives, failure at pilot scale can be useful. It can reveal incorrect assumptions about customer behavior, routing rules, process design, data quality, or system integration before those issues affect a much larger customer population. The important management discipline is rapid diagnosis and correction.

Penn Medicine experienced the consequences of scale when it deployed a voice assistant. According to Aaron Johnson, Interim Chief Marketing Officer at Penn Medical, the range of language patients used during actual calls was much broader than the team had expected. “In retrospect, we may have wanted to start with a smaller pilot,” Johnson said.

The organization responded with a tighter feedback process. Business leaders worked directly with technical tuning teams. They reviewed what patients were actually saying and adjusted routing logic in real time. This connected operational knowledge with technical changes and allowed the system to improve from observed customer behavior.

A strong pilot therefore needs more than a small user population. It needs explicit success criteria and a short feedback cycle. Teams should know which interactions were resolved, which required human intervention, where routing failed, and what customers actually attempted to accomplish. Those findings should feed directly into the next system version.

Executives should also define expansion gates before deployment. A system can move from pilot to wider use once it demonstrates acceptable performance for its intended task and the organization can manage its failure cases. This creates a disciplined path from experiment to production.

The objective is controlled learning at increasing scale. Start with a bounded problem. Observe real behavior. Correct weaknesses quickly. Expand when the operating evidence supports expansion.

Build employee trust before AI changes the work

A hybrid workforce changes how work is assigned, performed, and measured. AI handles a growing share of customer interactions while employees manage complex cases and increasingly supervise or work with automated systems. That transition requires deliberate workforce management.

Jessica Gupta, Chief Operating Officer at InfoPay, described an early communication strategy. Years before the company’s AI rollout was fully developed, leaders told employees that AI would change how they worked. Management did not yet know every detail of that change. It still communicated the expected direction.

That transparency created time for employees to understand the transition and participate in it. InfoPay brought high-performing agents into AI testing early. These employees used the technology, evaluated its behavior, and provided feedback based on their customer-service experience.

Participation also gave frontline expertise a direct role in system development. Experienced agents know where customers become confused, which exceptions occur repeatedly, and which information is required to resolve difficult cases. Incorporating that knowledge during testing can improve AI workflows, escalation criteria, and human handoffs.

The approach improved employee acceptance at InfoPay. “Now they’re actually looking at AI favorably,” Gupta said after describing the involvement of agents in the implementation process. Early participation gave employees a degree of ownership over the operational change.

Executives should treat this as a capability issue as well as a change-management issue. Hybrid customer operations require people who can handle difficult interactions while using AI effectively. Some employees will adapt faster than others. Companies need to identify those skills, develop them through training, and reflect them in recruitment and career paths.

Communication also needs precision. Leaders can explain which processes are being automated, how responsibilities are expected to change, how employees will participate in testing, and what skills will become more valuable. Where future effects remain uncertain, leaders can state that uncertainty clearly.

The central management task is to build the workforce alongside the technology. Employees closest to customers can improve AI systems when they have structured ways to test them, challenge their behavior, and feed operational knowledge back into development.

Measure AI by business outcomes

AI changes the meaning of traditional contact center metrics. Average handle time and transfer rates remain operationally useful. Their interpretation becomes more difficult when AI resolves routine interactions and sends harder cases to human employees.

The underlying workload has changed. A human agent who spends longer resolving an unusual customer problem may be creating more value than an agent who closes several simple requests quickly. Executives therefore need measures that connect operational performance to customer and business outcomes.

Jonathan Rosenberg, Chief Technology Officer at Five9, recommended maintaining visibility into customer return rates, contract renewal rates, and customer satisfaction scores. “Make sure these things remain good, so you’re keeping the top-level picture,” he said.

This distinction matters when evaluating the economics of automation. Faster calls, fewer transfers, and lower contact costs can help an organization understand process efficiency. Customer satisfaction, retention, renewals, and successful service outcomes provide evidence about whether the process is achieving its business purpose.

Penn Medicine encountered this issue after improving AI-driven call routing. Aaron Johnson, Interim Chief Marketing Officer at Penn Medical, explained that costs increased as the system became more effective at connecting patients to care. More successful access generated more downstream activity and therefore more cost.

A narrow efficiency measure could treat that cost increase as deterioration. From a healthcare-access perspective, more patients reaching appropriate care can represent a better operational outcome. The example shows why executives must define what the system is intended to achieve before choosing its performance measures.

The same principle applies beyond healthcare. A customer-service AI could reduce handling costs while affecting retention, satisfaction, or contract renewal behavior. Those outcomes have different economic consequences. Senior leaders need a measurement framework that connects AI activity to the outcomes each business values.

Metric ownership also matters. Operations teams may focus on queue performance. Finance may focus on cost. Commercial leaders may care about retention and revenue. Customer-experience teams may prioritize satisfaction. AI investment decisions become stronger when these measures are evaluated together under a shared definition of success.

Executives should establish that definition before scaling an AI deployment. Set operational measures for system performance and business measures for the result it is expected to produce. Track how the mix of AI and human work changes over time. Revisit targets as routine contacts move toward automation and the average complexity of human cases rises.

The core principle is straightforward: AI performance should be judged by the outcome the organization needs. Efficiency remains valuable, while effectiveness determines whether that efficiency creates business value.

Leadership determines whether AI transformation works

AI transformation creates a coordination problem across the business. Technology teams build and tune systems. Operations teams see how those systems perform with customers. HR manages changing skills and roles. Finance tracks economics. Senior executives decide where investment should go. These groups need a shared operating process.

Middle managers and frontline supervisors are central to that process. They see how AI affects daily work and hear directly from agents when routing fails, customer needs change, or policies prevent effective resolution. They can convert those observations into specific requirements for technology teams and senior management.

The communication must work in both directions. Executives set priorities, risk limits, and performance targets. Managers translate those decisions into operating practices. Frontline teams then generate new information through real customer interactions. Strong organizations capture that information and use it to adjust processes, policies, training, and AI behavior.

This becomes especially important when executive functions have different objectives. A CFO may focus on cost and return on investment. A CTO may prioritize reliability, integration, security, and technical scalability. HR leaders may focus on workforce capabilities and organizational change. Customer-experience leaders are responsible for service outcomes. Operational leadership has to connect these requirements into decisions employees can execute.

Executives should formalize this feedback process. AI programs need clear ownership, defined decision rights, regular operational reviews, and mechanisms for escalating problems that cross functional boundaries. Frontline observations should reach the teams capable of changing the system. Changes should then flow back into employee guidance and operating procedures.

This structure also affects the speed of improvement. Customer interactions can expose unexpected requests, weak routing logic, policy conflicts, and gaps in automation. When operational teams can quickly reach technical and executive decision-makers, the company can diagnose these issues and adjust the system faster.

Neville Letzerich, Chief Marketing Officer at Talkdesk, sees significant economic potential in getting this model right. “This is a huge opportunity, especially for all of us in customer experience,” he said. “We can deliver really great customer experiences at a fraction of the cost. We can serve customers faster. We can give them new experiences they didn’t even know about.”

Those gains require investment in organizational capability alongside AI systems. Companies need managers who understand customer operations, can work effectively with technical teams, and can explain business priorities clearly in both directions.

For the C-suite, the key issue is governance. Assign accountability for customer and business outcomes. Give operational leaders access to frontline evidence. Establish clear channels between operations, technology, finance, and HR. Give managers enough authority to turn that information into action.

AI can change the cost and speed of customer service. Leadership determines whether those technical gains become sustainable business results.

Concluding thoughts

AI can remove routine customer service work quickly. The harder task is redesigning the organization around what remains. Human agents will handle a higher concentration of unusual, sensitive, and consequential interactions. Their tools, authority, training, and performance measures must change with that workload.

For executives, this shifts the priority from deployment to operating design. Start with defined business problems. Test AI with controlled pilots. Build reliable handoffs to skilled employees. Measure customer and business outcomes alongside efficiency. Give frontline teams a direct path to the people who can improve systems and policies.

Workforce strategy deserves equal attention. Employees need early visibility into how their roles are changing and practical opportunities to shape the systems they will use. Middle managers need enough authority to connect frontline evidence with decisions across technology, finance, HR, and operations.

The next phase of AI will be determined by execution. Companies that combine effective automation with stronger human decision-making can serve customers faster, use skilled employees more effectively, and learn from difficult interactions. The technology creates the capacity. Leadership determines what the business does with it.