How Leaders Scale AI-Enabled Work Without Expertise Constraints
AI reduces the need for expert involvement in routine work while increasing the strategic importance of expert judgment. As employees use AI to perform specialized tasks, throughput rises and routine decisions move faster. Yet the hardest exceptions, approvals, and judgment calls may still concentrate with a small group of experienced professionals. The result is a new constraint: productivity can scale faster than the specialist capacity required to support it.
Succession planning, talent management, retention, cross-training, and knowledge transfer remain essential, but AI changes where those investments are most needed, how much expert capacity the business requires, and whether existing capacity can support future growth.
A strong AI strategy should account for more than productivity and automation. Leaders need to evaluate AI investments against the expert capacity required for oversight, exceptions, and complex decisions. That requires mapping critical expertise dependencies, testing whether alternatives can perform, developing future expertise, factoring scarce expert capacity into investment decisions, and revisiting assumptions as work changes. The goal is to determine where scarce expertise creates value, where it constrains growth or reliability, and which dependencies warrant investment or redesign.
How AI Productivity Changes the Demand for Expertise
AI can broaden access to capabilities that once depended heavily on individual experience. By embedding guidance and established decision patterns into everyday workflows, AI can help less-experienced employees perform more complex work with less direct specialist support.
A 2025 study published in The Quarterly Journal of Economics examined customer-support agents using a generative AI assistant. Access to the system increased issues resolved per hour by 15% on average, with particularly large gains among less-skilled and less-experienced workers. Agents with two months of tenure and AI assistance performed at levels comparable with untreated agents who had more than six months of tenure. The researchers also found evidence consistent with learning rather than simple reliance on generated responses.
AI can reduce specialist involvement per transaction without necessarily reducing total demand for expertise as activity scales.
Klarna illustrates how productivity gains in one part of a workflow can still depend on sufficient human capacity, quality control, and oversight elsewhere. In 2024, the company reported that its AI assistant handled 2.3 million customer-service conversations in its first month, equivalent to the workload of 700 full-time agents, while reducing average resolution time from 11 minutes to less than two. Yet in 2025, Klarna acknowledged that its emphasis on cost reduction had compromised service quality and began recruiting human support agents. The experience highlights why reducing routine workloads does not eliminate the need for human capacity to maintain service quality and handle interactions requiring personal support.
The pattern extends across industries: In healthcare, Sweden’s Mammography Screening with Artificial Intelligence (MASAI) trial found that AI-supported screening reduced radiologists’ image-reading workload by 44% while increasing cancer detection by 29% compared with standard screening. AI directed most examinations to a single radiologist while preserving double reading for higher-risk cases. Rather than eliminating specialist involvement, the technology changed where expert attention was required.
The example illustrates why productivity gains must be evaluated alongside the specialist capacity needed to sustain performance. Leaders should test AI-driven volume growth against the expert capacity required for escalations, approvals, and oversight under both normal and peak demand.
Key Takeaway
Measure AI productivity against the expert capacity required to support it. Lower escalation rates can still increase specialist demand as volume grows.
Explore Our Organization & Human Capital Consulting Services
P&C Global’s organization and human capital practice helps leaders design management capacity for flatter, faster organizations.
Locate the Critical Dependency Behind the Workflow
Traditional key-person risk assessments often begin with people: Which employees would be difficult to replace? Which positions lack successors? Which skills are scarce? A stronger assessment begins with the business outcome that must continue and works backward to the capabilities required to deliver it.
A manufacturing line facing an unusual process failure, for example, may require a maintenance technician who understands the equipment, a process engineer who can interpret operating tolerances, a quality specialist authorized to approve restart, and a software engineer familiar with the control system.
The dependency does not sit with one employee. It can involve four distinct constraints:
- Knowledge and judgment: Specialized understanding and experience required to make sound decisions.
- Throughput capacity: The availability of qualified professionals to handle the required volume of work.
- Decision authority and access: The approvals, permissions, and information needed to act.
- Technology and provider dependencies: The systems, data, platforms, and external capabilities supporting the workflow.
These constraints often overlap, but their remedies differ. Training can address a knowledge gap, for example, but it cannot resolve limited decision authority, insufficient specialist capacity, or dependence on a shared platform.
AI can make these dependencies harder to see. A diagnostic application may allow frontline technicians to resolve more equipment issues without direct engineering support. Yet the workflow may still depend on a small group of engineers to diagnose unfamiliar failure modes, evaluate AI recommendations under unusual operating conditions, and determine an appropriate response. Expertise has not disappeared. It has moved deeper into the workflow.
Leaders should map critical dependencies backward from business outcomes.
For each material workflow, determine:
- Which decisions can be standardized, delegated, or automated?
- Which activities still require specialized knowledge or judgment?
- Is enough qualified capacity available to meet routine and peak demand?
- Who holds the decision authority or access required to act?
- Which technologies, data sources, or external providers support the workflow?
- Which other workflows depend on the same people, capabilities, or systems?
Together, these questions reveal dependencies that a conventional skills inventory may miss, including constraints involving capacity, decision authority, and shared technology.
Find Concentration Across Units, Systems, & Providers
Business units may appear operationally independent while relying on the same technical authority, regulatory specialist, data-science team, or client relationship leader. Growth in one area can consume capacity required elsewhere.
Technology introduces another layer. A capability distributed across employees may still rely on the same model, platform, data source, provider, or small internal team. The workforce may be distributed while the infrastructure supporting its expertise remains concentrated.
In financial services, the Financial Stability Board identifies third-party dependencies, provider concentration, and substitutability as important vulnerabilities to monitor as AI adoption expands. The same logic applies more broadly wherever critical workflows depend on shared technology, external providers, or limited internal expertise.
This is why a skills inventory alone is insufficient for key person risk. Two organizations can employ the same number of specialists yet face very different exposure depending on how many critical workflows rely on the same people, systems, providers, or decision rights.
Key Takeaway
Map expertise backward from critical business outcomes. The most consequential bottleneck may not be a single employee, but a shared capability spanning people, systems, providers, and decision authority.
Explore Our Workforce Development Services
P&C Global’s workforce development services help build and replenish the judgment behind AI-enabled work.
Prove That Expertise Alternatives Can Actually Perform
Organizations can overstate resilience by counting alternatives that exist on paper but have never demonstrated they can perform when needed.
In succession planning, a successor may have completed training but never handled the hardest decisions independently. A cross-trained employee may understand the process but lack approval authority. An external specialist may possess the required expertise but take days to mobilize. An AI system may perform reliably on familiar cases yet struggle when conditions change.
Each provides nominal coverage. None necessarily provides usable capacity. Leaders should apply a stricter standard: a proven alternative.
For every material dependency, define what an acceptable alternative must be able to deliver, under what conditions, within what timeframe, and at what performance threshold. Then test whether the alternative can meet that standard. The test should reflect the dependency being addressed. A successor must demonstrate independent judgment, additional capacity must handle the required workload, delegated authority must function in practice, and fallback technology must remain available when primary systems fail.
Evidence may come from live work, supervised decisions, simulations, rotations, recovery exercises, or other tests appropriate to the function. The objective is not simply to confirm that a backup exists. It is to determine whether that backup can restore the required capability before disruption becomes unacceptable.
Testing alternatives is not limited to technology resilience. NASA, for example, uses simulations of system failures and emergency scenarios to assess flight controllers before they assume mission-control responsibilities. In healthcare, multidisciplinary simulations allow clinical teams to practice emergency responses, assess coordination, and identify capability gaps before real crises occur. Both illustrate why training or designated backup roles alone do not establish readiness. Alternatives must demonstrate that they can perform under realistic operating conditions.
Research into generative AI reinforces the same distinction. A 2025 Harvard Business School working paper found that AI helped marketing professionals produce work comparable to specialist web analysts, but was less effective for software developers and data scientists whose existing skills were less closely related to the task. AI helped participants organize ideas but could not fully compensate for the experience needed to execute unfamiliar work. The findings reinforce a key principle: access to the same tool does not make employees interchangeable.
Test for Capability, Recovery, & Independence
The proof standard should extend to people, external providers, and fallback technologies. An outsourced specialist offers little protection if the provider cannot mobilize before the business exceeds its tolerance for interruption. A backup platform offers limited resilience if it depends on the same infrastructure, data source, internal administrators, or approval authority as the primary system.
Three measures help distinguish apparent coverage from a proven alternative:
- Capability coverage: Which critical decisions or activities have at least one alternative that has demonstrated it can perform to the required standard?
- Recovery time: How long would it take to restore an acceptable level of capability if the primary arrangement became unavailable?
- Dependency independence: Does the alternative rely on the same people, providers, technology, data, infrastructure, or decision authority as the primary arrangement?
These measures should remain separate. An organization may have broad coverage but unacceptable recovery times, or a fast backup that fails because it shares the same underlying dependency.
Testing should reflect the conditions in which the alternative would be needed. A successor who performs well with supervision may respond differently under pressure; an external provider may lack capacity during a broader disruption; a fallback technology may fail when dependent systems are unavailable.
Availability is not the same as readiness. What matters is whether the alternative can carry the required work, at the required level, when the primary capability is unavailable.
Key Takeaway
Treat an alternative as proven only when it has demonstrated the required capability, recovery speed, and independence under realistic operating conditions.
Protect the Ability to Develop Future Expertise
AI can improve employees’ immediate performance while changing how they develop the judgment needed for more complex work. The risk is that automation may reduce opportunities to build experience, even as the business continues to depend on independent expert decisions.
AI does not necessarily weaken learning. A 2025 study published in The Quarterly Journal of Economics found that generative AI improved performance among less-experienced customer-support employees, with evidence of learning that persisted even without AI assistance. However, those gains do not establish that AI develops the independent judgment required across all professions.
A September 2026 NBER randomized trial involving over 130 patent lawyers across U.S. intellectual-property firms found that a custom AI drafting assistant improved assisted quality by 0.34 standard deviations after 10 days and 0.38 standard deviations after 90 days. After three months, however, improvements on an unaided judgment task were concentrated among senior lawyers. Junior lawyers showed no average improvement and greater variation. While this does not establish skill deterioration, it raises concerns about whether AI-assisted work consistently develops independent judgment.
A separate 2026 randomized study of software developers found that participants learning an unfamiliar programming library with AI scored 50% on a subsequent skills assessment, compared with 67% for those working without AI. Participants who used AI to deepen their understanding, rather than simply delegate tasks, performed better on the assessment.
Together, the findings suggest that AI can improve immediate performance without necessarily developing the expertise required for independent work.
Distinguish Work to Automate From Experience to Preserve
Automation decisions should distinguish between work humans still need to perform and experience humans still need to acquire.
Routine assignments expose employees to recurring patterns, edge cases, customer behavior, equipment conditions, and decision consequences that later inform more complex judgment. If automation removes those assignments, organizations need other ways to build that experience.
Those knowledge transfer mechanisms may include supervised AI-assisted work, simulations, rotations, structured exception reviews, and progressive decision authority.
Development should reflect the capabilities the business will need next. If technology, regulation, customer behavior, or product design will materially change the expertise required, reproducing today’s specialist role may have limited value.
Preserve the Capacity to Develop Others
Developing future expertise consumes senior capacity.
Experienced professionals cannot simultaneously maximize client delivery, resolve escalations, improve AI-supported workflows, and develop successors without tradeoffs. If leadership expects experts to develop others, that demand should be reflected in workload and workforce planning. Mentoring, validation, and guided review replenish organizational expertise; they are not peripheral work.
At software company Bonterra, for example, junior engineers work alongside experienced colleagues as AI agents take over routine coding tasks. Rather than simply reviewing AI-generated output, junior engineers remain responsible for project outcomes, learning to direct AI agents, challenge their work, and develop the judgment needed for more complex assignments.
The challenge is to give employees opportunities to develop independent judgment, supported by experienced professionals, without recreating the routine work that AI was intended to reduce.
Key Takeaway
Build expertise development into AI-enabled workflows, reserving sufficient senior capacity for mentoring, review, and the development of independent judgment.
Price the Full Economics of Expertise
The economics of AI depend not only on the work technology removes, but also on the scarce human capacity the resulting workflow still requires.
A 2026 NBER study of more than 500,000 GitHub developers shows how those economics can diverge. Autonomous coding agents increased coding activity, measured by code commits, by 240%. Yet the increase fell to 80% for the number of projects and just 30% for software releases. The findings are consistent with human bottlenecks in reviewing, integrating, testing, and releasing code, which can limit how much additional coding activity translates into finished software. These activities can constrain how much AI-generated work becomes finished software, limiting the productivity gains realized across the full workflow.
Model the Expert-Capacity Constraint
Consider an illustrative AI-enabled technical-support operation. Management plans to increase weekly case capacity from 10,000 to 16,000 while AI reduces the specialist escalation rate from 5% to 3.5%.
The headline case looks attractive:
- Processing capacity increases 60%.
- The escalation rate falls 30%.
- More cases are resolved without specialist involvement.
Yet specialist escalations still rise from 500 to 560.
If existing specialists can sustainably handle 520 cases a week, the workflow exceeds available expert capacity despite the lower escalation rate. Scaling now depends on reducing exception demand, adding usable specialist capacity, or moderating volume growth.
The investment case should account for escalation, validation, and oversight time; external capacity; rework and delayed decisions; supporting technology and knowledge infrastructure; and the opportunity cost of deploying scarce experts in one activity rather than another. Management can then compare those costs with the incremental revenue, margin, service improvement, labor savings, and risk reduction the investment is expected to generate.
Compare the Economics of the Response
The appropriate response depends on the constraint. Knowledge gaps may justify training or knowledge transfer. Capacity limits may require additional specialists, workload redistribution, or process redesign. Authority constraints may call for delegated decision rights, while technology dependencies may require independent alternatives or provider diversification. Each response carries different costs, activation times, flexibility, and residual risk.
Additional redundancy does not always justify its cost. Duplicating highly specialized knowledge may destroy value when demand is infrequent and external access provides sufficient coverage. Conversely, training or broader decision rights may generate substantial returns when expert availability constrains revenue-generating capacity.
In some cases, the highest-return option is to remove the dependency. Years of customization, exceptions, or unclear decision rights can make routine work unnecessarily dependent on scarce expertise. Standardization may create more value than adding specialists to sustain that complexity.
Organizational design also affects AI returns. A 2025 Management Science study on domain expertise and algorithmic literacy found greater market value from AI and algorithmic investments when related capabilities were more broadly distributed among domain experts rather than concentrated in narrow technical groups.
Key Takeaway
Price scarce expertise into the investment case. Scale AI when the value created exceeds the full cost of the expert capacity, oversight, and operating constraints required to support it, and choose the configuration that meets performance and risk requirements at the best risk-adjusted economics.
Govern Expertise as the Work Changes
Key person risk and expertise decisions have a shelf life. A concentration that is manageable today may become a constraint as demand grows, technology changes, providers evolve, or responsibilities shift. Governance should identify when the assumptions behind an existing decision no longer hold.
For each material dependency, management should document four elements:
- Operating assumption: What must remain true for the current arrangement to remain acceptable?
- Trigger: What observable change would call that assumption into question?
- Owner: Who is responsible for recognizing the change and reopening the decision?
- Decision window: How long can management wait before the change begins to affect performance, continuity, or growth?
Consider an organization that depends on a single technology provider for a critical workflow. Management assumes the provider can maintain agreed service levels. A deterioration in performance triggers a review led by the technology executive responsible for continuity. If qualifying an alternative provider takes several months, that lead time must shape when management intervenes. Together, the four elements ensure that warning signals translate into accountable decisions before alternatives become impractical.
Turn Expertise Assumptions Into Management Decisions
The earlier technical-support example also illustrates the importance of timing. Management assumes specialists can sustainably handle 520 escalations per week. A forecast showing demand will exceed that limit triggers a review, with demand projected to reach 560 in eight weeks. The head of support operations must decide whether to expand capacity, reduce escalation demand, or adjust the rollout. If training and authorizing additional specialists takes six weeks, management has no more than two weeks to act, and less if additional time is needed to manage uncertainty.
The same approach applies to other dependencies. Changes in provider availability may require a continuity decision before recovery options become inadequate, while shifts in decision authority may warrant intervention before approvals delay critical work. In each case, management must establish what change warrants action, who owns the decision, and how long the organization can afford to wait.
Defined triggers are not a substitute for periodic review. Underlying assumptions can become outdated as workflows, technology, business priorities, and external conditions evolve, even when no threshold is breached.
Executive reporting should focus on assumptions at risk, approaching decision deadlines, accountable owners, and actions requiring leadership attention.
Key Takeaway
Govern expertise through explicit operating assumptions, observable triggers, accountable owners, and decision windows. Revisit dependencies while management still has viable options.
Scale Expert Capability Along With Output
AI can broaden access to expertise and increase productivity. Higher output, however, does not necessarily mean the business has become less dependent on scarce judgment.
In some workflows, technology will reduce expert demand. In others, it will concentrate that demand around the decisions, exceptions, and improvements that remain difficult to automate. The advantage belongs to organizations that understand where those constraints will emerge before they begin to limit growth, reliability, or investment returns.
Human capital management that treats expertise as an economic resource changes the objective. The aim is not to eliminate specialization or maximize redundancy, but to ensure that critical capability can scale, adapt, and be replenished as the work changes.
The measure of success is therefore not how much output AI enables today. It is whether the organization can sustain the expertise required to support tomorrow’s performance.
