AI productivity gains create harder management decisions

Last week, a group of CHROs in Atlanta sat down to discuss the ROI from scaling AI across their businesses. The discussion moved from ROI to harder questions: which problems justify AI investment, what happens to the capacity AI creates, and when accumulated task changes require a job redesign. Successful AI use therefore creates a new set of management decisions about work.

Those decisions matter because a productivity gain has no predetermined organizational outcome. Leaders can turn released capacity into more output, different responsibilities or lower future hiring, and they can turn a successful team’s productivity into a broader performance expectation. Each move requires judgment about jobs, operating conditions and access. The question for the next year is whether management systems can make sound decisions as AI starts to work at scale.

First, learn to choose problems worth solving

Before leaders face those downstream decisions, they have to find applications worth pursuing. Kia Painter, who recently retired after 28 years at Cox Communications and most recently served as chief people officer, proposed a specific sequence. Leaders should identify the business problem, understand its workflow, redesign the process and then determine where AI belongs.

That sequence changes what AI competence means for a manager because operating a tool is only part of the work. The manager has to inspect how work moves, find where time is lost or quality suffers, and determine whether AI can materially improve the outcome. Technical success alone cannot establish business value when the underlying problem has little value.

The need for that judgment also appeared a few weeks before the Atlanta dinner, when the organizer ran an informal poll in the AI Signal newsletter about why managers wanted more AI training for their teams. The leading reason was uncertainty about which use cases mattered. Because the poll was informal, it signals the problem rather than providing general evidence about employers, but it exposes an important distinction. Learning to operate AI and learning where to apply it are different skills.

That distinction matters because conventional AI training often teaches people to write prompts, summarize documents and generate content. Those exercises build familiarity, while business-use-case judgment requires people to connect a technology intervention to a meaningful loss of time or quality in an existing workflow. Companies can develop increasingly comfortable AI users while their managers still struggle to identify work worth changing.

Angela Cheng-Cimini, CHRO at The Chronicle of Philanthropy, described a learning model built around that problem. Her organization planned to crowdsource everyday challenges from employees, choose several with broad applicability, and run short hackathons to address them. Employees would start with work they understood and use AI to solve a concrete problem, making problem selection part of the learning process.

One employee had already followed that pattern independently when repetitive invoice requests began flooding a customer-service inbox. The employee built an agent to handle them. The sequence is clear in the example: an observable workflow problem came first, and the decision to use an agent followed from understanding that work.

Use-case judgment may consequently be a scarcer organizational capability than basic tool-operation skill. It also establishes the first discipline needed for meaningful AI returns: leaders have to choose worthwhile work. Once an application succeeds, the next discipline begins because management has to decide what the resulting change should become.

A productivity gain creates capacity; leadership decides what that capacity becomes

A successful use case creates capacity when AI removes administrative work or reduces cycle times. Management then has to decide where that capacity goes because saved time does not choose its own organizational purpose. That destination determines both the return and what employees will subsequently be expected to do.

Released time can support several operating choices. Leaders can increase output, improve service, avoid some future hiring, broaden employees’ responsibilities or restore work that other demands had crowded out. The same number of hours saved could produce materially different outcomes, which means the initial time reduction cannot settle the ROI question by itself.

That distinction became explicit during the headcount discussion in Atlanta. Painter recalled an unnamed CFO asking whether AI could reduce labor costs. Her response returned the analysis to the work, examining AI through the operating model and work design before making assumptions about the workforce.

Starting with the work changes the sequence of a headcount decision because leaders first identify which activities have disappeared or accelerated. From there, they can determine what the remaining activities demand from employees and which responsibilities could use the released capacity. Labor economics then follows from an operating decision grounded in how the work has changed.

Once management directs saved time toward higher output, wider responsibility or a different service level, the employee’s role in producing value changes with it. AI success therefore creates a second-order management problem: organizations have to turn capacity into an explicit operating choice. Repeated choices of that kind eventually affect the design of the job itself.

When enough tasks change, the job itself has changed

Those capacity choices accumulate at the job level. Kesi Dorner, AGM of Human Resources at Metro Atlanta Regional Transit Authority (MARTA), described the possibility of consolidating work currently distributed among specialized HR positions into broader business-partner roles. AI-driven work redesign can therefore alter the boundaries between roles as well as the speed of individual tasks.

Those boundaries can begin to change almost invisibly because each adjustment may be small. One task disappears, another becomes faster, routine activity contracts, and the work left behind requires more judgment. New responsibilities can then fill the capacity AI released, gradually changing both what an employee does and what capabilities the organization expects.

Because the changes accumulate, no single step has to look transformational. The key question is cumulative: at what point do enough changes to tasks, judgment and responsibility mean that the organization is dealing with a different job? When that question was put to the CHRO group in Atlanta, it went unanswered.

That unanswered question matters for workforce design because there may be no defensible universal percentage of automated or altered tasks that determines when a role needs reconstruction. A job combines responsibilities, required skills and expected judgment, so leaders have to monitor how those elements change together. The Atlanta discussion exposed a decision organizations need to make without imposing false precision on where its boundary sits.

The boundary has direct consequences for HR because job architecture drives other management systems. Job descriptions have to reflect the work employees actually perform, skill requirements have to match what the work now demands, and compensation and career paths have to correspond to the resulting role. Performance expectations likewise become unreliable when they continue to describe a job whose content has substantially changed.

Incremental adoption makes that mismatch easy to create. Leaders can introduce one AI change at a time while leaving the existing role formally intact, even as routine work contracts and more consequential judgment takes its place. Employees can eventually be doing different work under an old job design, with compensation, progression and evaluation still tied to assumptions that preceded the change.

Broader roles such as the business-partner model Dorner discussed make the consequence concrete. Combining previously specialized work expands the range of responsibilities an employee carries and can change the skills required to succeed. At that stage, the efficiency gain has become an organizational change because management has redesigned who is responsible for what.

The unanswered threshold also limits what deployment metrics can tell leaders. Organizations may be able to measure that a task became faster while still lacking a process for deciding when those measured changes invalidate the design of the job around it. That decision becomes increasingly important as companies move from isolated uses of AI toward expectations applied across teams.

A strong pilot establishes evidence under specific conditions

The same concern with conditions applies when leaders move from one job or team to enterprise-scale expectations. If one team becomes more productive with AI, its new output level can look like an obvious candidate for a broader standard. Before using it that way, leaders need confidence that the gain will persist, that work elsewhere is sufficiently similar, and that employees have comparable access to the technology.

Those checks matter because pilot performance can depend heavily on local conditions. A team may have a structured workflow, clean data, a manager who understands the technology and enough time for employees to experiment. Another function may lack those conditions, so copying the observed productivity target can separate the expectation from the environment that produced it.

Scaling therefore begins with understanding why the pilot succeeded. Leaders have to distinguish elements that can be standardized from those that depend on a function’s work, workforce or risk profile, then coordinate the investment and governance needed to reproduce what transfers. A successful implementation becomes enterprise evidence to the extent that its relevant conditions can travel with it.

Painter described how Cox Communications approached that coordination with its head of AI. They developed a target-state roadmap connecting technology decisions with the company’s future operating model. The roadmap placed individual AI choices in the context of how the organization intended work to operate later, giving deployments a common organizational destination.

That roadmap was paired with a cross-functional AI council involving HR, technology, finance and legal. Those functions coordinated investment and governance around business priorities instead of independently accumulating their own tools and projects. Scaling spans several forms of accountability at once: technology determines feasibility, finance has to understand investment and returns, legal participates in governance, and HR owns much of the impact on jobs and people.

Cross-functional coordination also places a useful boundary around productivity claims. Evidence from one strong pilot supports a conclusion about that use case under its particular conditions. Workforce-wide productivity or headcount assumptions require further evidence that the relevant work and enabling conditions are sufficiently comparable.

That requirement becomes especially important when performance standards follow the technology. A local gain may be sustainable and transferable, but managers need to establish those properties before treating the resulting output as a baseline for other teams. Otherwise, the organization can encode the advantages of one environment into expectations for employees working in another.

Performance standards are defensible when leaders account for unequal AI conditions

The differences between those environments include access to AI itself. Participants in the Atlanta discussion described employees receiving different tools because of their function, budget or company policy. Some workers supplement employer-provided access by paying for AI tools themselves and using them at home or on personal devices, while other employees cannot do so.

Those access differences become relevant to performance management once AI materially affects output or quality. Kara Miller, former vice president of talent, learning and workforce transformation at Visa, raised the case of two employees facing the same output standard while one has access to a more capable tool. Any observed performance difference then becomes ambiguous because it can reflect employee capability, the environment the employer created, or some combination of the two.

That ambiguity extends beyond an immediate rating because access also affects learning. Roles with greater AI access can give employees more opportunities to develop AI skills through daily work, and those accumulated opportunities can influence development and career mobility. An organization that uses AI-influenced performance as evidence of individual capability therefore has to understand how access shaped the opportunity to build and demonstrate that capability.

Accounting for those conditions does not require every employee to receive identical tools. The CHRO group agreed that universal uniformity was unnecessary. The management responsibility begins when access differences materially affect the standards used to judge people. Leaders need to know when function, budget or policy has changed the conditions enough that an output comparison no longer cleanly supports the personnel decision being made.

Productivity and fairness consequently become part of the same management question. Once an AI-enabled level of output enters performance expectations, leaders have to establish whether employees had conditions capable of supporting that standard. Evaluation, development and mobility decisions depend on that distinction.

AI maturity depends on distributed managerial capability

Near the end of the Atlanta evening, the group considered what would have to change over the next year for leaders to feel they had figured out how to scale AI. Their answers centered on judgment distributed through the organization: executives able to discuss the technology with some fluency, employees who understand responsible use, managers capable of identifying worthwhile applications, and clearer decisions about access. In this view, maturity depends on whether different levels of the organization can make the decisions that AI use creates.

That definition sets a higher bar than adoption because employees using AI is only the beginning of the management problem. The organization also needs people with enough judgment to connect AI to worthwhile work and make decisions about its consequences under different operating conditions. Fluency, responsible use and access decisions matter because scaling turns local technology choices into recurring management decisions.

Some of those decisions will remain matters of judgment rather than fixed thresholds. Companies do not yet have a universal threshold for when accumulated task changes constitute a new job, while defensible access decisions do not require giving everyone identical technology. The next stage of AI maturity therefore depends on leaders being able to make those operating choices deliberately as the work changes beneath them.

Key executive takeaways

• Choose problems worth solving: Managers need to connect AI to specific workflow losses in time, quality or service before selecting a tool. Training should build use-case judgment alongside technical fluency.
• Decide where AI-created capacity goes: Productivity gains create choices about output, service, responsibilities and future hiring. Management teams need to define the intended use of released capacity before translating efficiency gains into workforce assumptions.
• Redesign jobs as task changes accumulate: Repeated automation and workflow changes can alter responsibilities, skills and judgment enough to create a different job. HR and operating leaders need to review job architecture, compensation, career paths and performance expectations as those changes accumulate.
• Scale the conditions behind successful pilots: A strong pilot establishes what worked in a particular environment. Cross-functional teams need to identify which workflows, tools, data and management conditions produced the gain before applying its productivity expectations elsewhere.
• Align performance standards with AI access: Differences in AI tools and opportunities to use them can affect output, skill development and career mobility. HR and business leaders need to account for those conditions when AI-enabled productivity influences evaluation and advancement.
• Build managerial capability across the organization: AI maturity depends on distributed judgment about use cases, responsible use, access and changing work. Organizations need managers who can make those decisions consistently as AI moves from isolated deployments to routine operations.