AI Value Requires More Than Deployment. It Requires Redesigning the Work.

An employee can use AI to complete an analysis in half the time while the organization gains almost nothing from the improvement.

The analysis may still wait for several approvals. Another team may need to reformat or validate it. The next stage of the process may already be operating at capacity. The employee may simply receive more work, without any improvement in service, quality, cost, or cycle time.

This is one of the central challenges facing organizations investing in artificial intelligence. AI can make individual tasks faster, but individual productivity does not automatically become organizational value.

That distinction is becoming increasingly important. In McKinsey’s 2025 State of AI survey, workflow redesign had the strongest relationship with respondents’ reported EBIT impact among the practices examined. Yet only 21 percent of organizations using generative AI reported that they had fundamentally redesigned at least some workflows. The findings are self-reported and do not establish causation, but they point to an important pattern: organizations appear more likely to realize value when they change how work operates, not simply when they give employees access to new tools.

The executive question is therefore no longer only, “Where can we use AI?”

It is, “How should the work now operate?”

Individual productivity is not organizational effectiveness

AI tools can create meaningful improvements in personal productivity.. Employees can draft documents, summarize information, analyze data, retrieve knowledge, and prepare communications more quickly. These gains matter.

But an organization is not simply a collection of individual tasks. It is a system of connected activities, decisions, handoffs, controls, and responsibilities.

A faster task creates broader value only when the surrounding system can use the improvement.

Consider a report that once took four hours to prepare and can now be completed in two. The time savings may appear significant. But the organizational outcome may remain unchanged if the report still sits in an approval queue for three days, if another employee must verify every conclusion, or if the decision-maker receiving it has no additional capacity to act.

The organization must also decide what the saved time is intended to accomplish. Should it increase output, improve quality, reduce cost, strengthen customer service, support innovation, or create capacity for more complex work?

Without a clear answer, productivity gains can disappear into the operating environment. Employees may work faster without the organization becoming more effective.

This is why AI adoption should not be evaluated only through tool usage, employee activity, or time saved. Leaders also need to understand whether AI is improving the performance of the larger system that produces the outcome.

AI changes task chains, not only tasks

Intentional work redesign begins by examining what happens around the task AI performs..

Recent MIT Sloan research on work design argues that AI’s larger value often emerges at the workflow level through changes in how tasks are sequenced, grouped, and handed off between people and machines. This is a more consequential change than simply helping an employee perform an existing task more quickly.

AI may alter which tasks are necessary, when they occur, and who performs them. It may move information earlier in a process, combine activities that were previously separated, or allow employees to make decisions without waiting for another function. It may also introduce new review points, exceptions, and accountability requirements.

These changes can reshape roles.

An employee who previously spent most of the day locating and organizing information may spend more time interpreting it. A manager who once reviewed the completeness of a document may need to evaluate the quality of the judgment behind it. A specialist may move from producing every output directly to overseeing AI-supported work and handling unusual cases.

Morgan Stanley’s use of AI in advisor work offers one illustration. The organization did not simply provide access to a general-purpose tool and expect value to emerge. AI was incorporated into defined knowledge-retrieval and meeting-related workflows, supported by evaluation, quality assurance, and compliance controls.

The example does not prove a universal financial return, and it should not be treated as a blueprint for every organization. It demonstrates a more useful point: serious AI adoption includes the surrounding workflow and control environment.

The technology is only one part of the design.

Work must be designed around an uneven technology

AI does not perform equally well across all types of work..

Research on the “jagged technological frontier” found that AI improved performance on some knowledge-work tasks while reducing performance on other tasks that appeared similarly difficult. This makes broad assumptions about AI suitability especially risky.

A task may look routine but depend on context that is difficult for an AI system to interpret. Another task may appear highly specialized but contain repeatable patterns that make AI assistance valuable. Leaders cannot reliably determine the right design by looking only at job titles or broad occupational categories.

They must consider the characteristics of the work itself.

What happens when the output is wrong? Is the error easy to identify and reverse? Does the work require empathy, tacit knowledge, negotiation, or accountability? Is reliable data available? Can a qualified person review the output? How frequently do unusual cases occur?

Customer service illustrates the difficulty. Klarna’s aggressive use of AI in service operations demonstrated the potential scale of automation, but the company’s later emphasis on restoring more human support also highlighted the importance of complexity, escalation, customer expectations, and service quality.

The lesson is not that AI customer service failed. It is that automation scope is a design decision, not merely a technical possibility.

AI-enabled workflows may also create new work. Outputs may require validation. Exceptions must be handled. Performance must be monitored. Customers and employees need routes to human support. Someone must remain accountable when an AI-supported process produces an adverse result.

Those activities are part of the true cost and design of AI-enabled work.

The surrounding organization must change with the workflow

Workflow redesign cannot be separated from learning, management, and governance..

When a workflow changes, employees need more than general awareness of AI. They need to understand how their own responsibilities are changing, where AI is useful, when it should be questioned, and what standards still apply to the outcome.

The capability requirement is role-specific.

Research involving customer support agents found that an AI assistant was associated with an average productivity improvement of 14 percent, with larger gains among novice and lower-skilled workers. The findings came from one organizational context and should not be generalized across all work. They nevertheless show that AI can influence both performance and the distribution of expertise.

This creates opportunities, but it also raises questions about how expertise will develop in the future.

If AI helps less-experienced employees perform more effectively, it may accelerate learning. If it absorbs too many of the tasks through which employees traditionally built judgment, it may weaken future capability. The outcome depends on how the work and learning experience are designed.

Managers are also essential. They are often the people closest to emerging patterns of use. They can observe where AI improves work, where employees are over-relying on it, where expectations are inconsistent, and where the workflow produces new risks or friction.

Governance must also move closer to the work.

Frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 treat AI risk management as an ongoing organizational responsibility. Policies and review committees may be necessary, but they are not sufficient. Governance becomes operational through permissions, decision rights, evaluation standards, review points, escalation paths, and clear accountability.

When workflows, learning, management, and governance are designed separately, employees receive conflicting signals. They may be encouraged to experiment while facing unclear boundaries. They may be trained to use a tool that their manager does not understand. They may be held accountable for an outcome without knowing when human review is required.

Organizations work better when these elements are designed to reinforce one another.

AI creates executive design decisions, not one universal model

There is no single correct operating model for AI-enabled work.. Different strategies, risk environments, customer expectations, and workforce needs will produce different choices.

The role of leadership is not to adopt a universal position. It is to make the choices deliberately and understand their consequences.

What should AI automate, and what should it help people do better?

Some organizations may use AI primarily to expand human capability.. Others may pursue significant automation, cost reduction, or workforce restructuring. Many will use a combination of approaches.

The appropriate direction depends on the repeatability of the work, the consequences of error, the availability of qualified review, customer expectations, and the strategic importance of human expertise.

Leaders should also consider whether automating a task removes an important learning opportunity. Entry-level work often provides the experience through which people build judgment for more advanced responsibilities.

The immediate efficiency gain should therefore be considered alongside the organization’s future capability needs.

What must be consistent, and what can be designed locally?

Enterprise standards can reduce risk, improve interoperability, and create a consistent customer or employee experience.. They may be essential when AI use involves shared data, regulated decisions, or work that crosses business units.

Local experimentation can be equally important. Frontline teams often understand the details of their workflows better than a centralized function, and many valuable use cases emerge through practical experimentation.

The decision need not be complete centralization or complete decentralization. Organizations may establish common boundaries for data, risk, and accountability while allowing teams to redesign work within those boundaries.

The more important question is how effective local practices will be identified, evaluated, and scaled.

What level of oversight matches the consequences of failure?

The common framing of speed versus governance is too simplistic..

The more useful decision is how much oversight is appropriate for the risk and reversibility of the work. A low-risk internal task may support rapid experimentation. A customer-facing, financially consequential, regulated, or safety-sensitive process may require stronger evaluation and human approval.

Governance should make responsible movement possible. When it is disconnected from the work, it can become a layer of delay. When it is built into decisions, reviews, and escalation paths, it can create clarity and confidence.

What organizational outcome should new capacity serve?

Perhaps the most important decision is what the organization intends to do with the capacity AI creates..

The answer may be greater volume, faster service, higher quality, lower cost, increased innovation, expanded employee responsibilities, a different workforce structure, or greater organizational resilience.

Each choice requires a different redesign.

An organization seeking faster service may need to remove downstream approval delays. One seeking higher quality may reinvest time savings in deeper analysis or customer interaction. One seeking lower cost may need to redesign roles and staffing. One seeking innovation may need to protect time for experimentation.

Unless leaders define the intended outcome, task-level gains may remain scattered across the organization without producing a measurable change in performance.

Begin with one consequential workflow

Organizations do not need to redesign every process at once..

A practical starting point is one consequential workflow where AI use is already emerging, the outcome matters, and existing friction is visible.

Leaders can examine what happens before and after the AI-supported task. Where does information come from? Who reviews the output? What decisions depend on it? Where does the work wait? What happens when the AI is wrong? What must employees learn? What should managers observe? Who remains accountable?

This shifts the conversation from a general discussion of AI capability to a more useful examination of organizational performance.

It also makes the tradeoffs visible. The organization can see where AI removes work, where it creates new work, where decision rights need to change, and whether the surrounding process can convert the improvement into a better outcome.

AI deployment may create many isolated moments of productivity. Intentional work redesign determines whether those moments contribute to stronger service, greater capacity, better quality, lower cost, or new forms of value.

The consequential question for leaders is not simply where AI can perform a task.

It is how the surrounding work must change so that the task produces a better organizational outcome without weakening judgment, accountability, or capability.

Executive Questions

As your organization explores AI-enabled work, consider discussing:

• Which workflow would create the greatest organizational value if it were intentionally redesigned?
• Where are employees already using AI without corresponding changes to roles, workflows, or accountability?
• What organizational outcome should the capacity created by AI actually produce?
• Which management practices, learning investments, or governance mechanisms would need to evolve for that redesign to succeed?