AI deployment falls short without changes to work

AI activity can leave a company largely unchanged

A company can buy AI licenses, run pilots, and give employees chatbot access while leaving its business largely unchanged. For CEOs and boards reviewing disappointing returns, the immediate question is what has changed: who makes decisions, how work moves, and where people spend their time. Adopting technology creates an opportunity to change those arrangements. Leadership has to decide whether to use it.

Disappointing business returns belong alongside AI’s wider risks. Researchers and technology leaders are debating whether increasingly capable systems could escape human control, while job displacement, misinformation, and technological power concentrated in a handful of companies also demand attention. Executives can address business value directly through investment choices and changes to the organization. It deserves a place in leadership discussions about AI.

This practical concern surfaced during a Tokyo panel with Japanese CEOs, convened by Nikkei, the Japanese business and financial newspaper. An adviser in the discussion heard frustration that substantial investment and experimentation were producing less value than leaders expected. The panel covered talent, organizational design, and the responsibilities of boards and CEOs, all choices that shape what companies gain from technology. About a week later, a conversation in Singapore helped the adviser explain the gap.

Disappointing returns raise a question about business change

The frustration in Tokyo also appears in a recent Bain & Company survey of CEOs worldwide. Bain is a consulting firm with a commercial interest in advising companies on AI transformation, so the finding comes from a business that benefits from demand for that work. The survey compared reported results with the CEOs’ own expectations.

To explain why extensive activity can produce disappointing returns, the adviser drew on the Singapore conversation. A senior government digital leader presenting there said, “Deployment is not transformation.” Deployment means making technology available and putting it to use. Transformation means changing the work around it: the order of tasks, who has authority to decide, and how information reaches the people or systems that need it.

This distinction changes how leaders should interpret the activity described earlier. In the adviser’s account, licenses provide access, pilots let teams experiment, and extending chatbot access to thousands of employees broadens participation. These steps can be useful, but their scale shows how widely the technology has been introduced. To understand the business result, leaders must also examine what employees now do differently.

A pilot can show that AI performs a task while employees still pass work through the same stages, seek the same approvals, and handle information in the same ways. The Singapore presenter also observed that adding a chatbot to a defective process leaves the process defective. A new interface may make the sequence easier to use, but it cannot by itself remove an unnecessary step or settle an unclear responsibility. Fixing those problems requires decisions about how work is organized.

Changing how work is organized can mean removing or reordering workflow steps (the tasks that make up a process), redistributing decision rights (the authority to make particular decisions), or changing how information moves. These choices also affect where employees spend their time. Boards can therefore review deployment and business change separately, asking how widely AI is used and how that use changes the way the company produces a result. The adviser’s view is that much corporate AI activity still helps employees perform familiar jobs in familiar ways, with some added speed.

Efficiency and differentiation require different tests

If much AI activity makes existing work faster, leaders need to distinguish that goal from more ambitious changes. At the Nikkei panel, the adviser proposed three stages: Improve, Innovate, and Build. The framework gives leaders a vocabulary for choosing the business change they want before deciding how to make it happen.

Improve matters because completing useful tasks more efficiently can improve the economics of work. The adviser argues that much company effort remains at this stage, making unchanged jobs somewhat faster. A productivity gain can justify investment even when competitors can gain a similar benefit from the same technology. So leaders need to assess efficiency and lasting competitive advantage separately.

For competitive advantage, the adviser expects companies to pull further apart through Innovate and Build, where AI changes the business or makes a new offering possible. Companies can deliberately pursue efficiency or choose parts of the business that need deeper reconsideration. The stages make that choice explicit without imposing a mandatory sequence. Accumulating Improve projects does not automatically produce the outcomes associated with Innovate or Build.

Choose a painful workflow before deciding AI’s role

Once leaders have chosen their ambition, they need to select work where it can matter. The adviser recommends focusing on a handful of areas that could materially change business economics or customer experience. In the adviser’s view, hundreds of disconnected use cases can spread attention across many small tasks, while concentrating on consequential work gives leaders room to address the organizational decisions involved. The business problem should set the scope.

Finding that problem begins with friction: delay, excessive cost, error, or a poor customer experience. Each symptom points to work worth examining, but leaders still need to understand its cause. Introducing AI where the symptom appears may leave the underlying difficulty untouched. Investigating the cause helps the company decide what success should look like.

Success then gives the redesign a clear purpose. If delay is the concern, the outcome should describe how work reaches completion sooner; if customer experience is the concern, it should explain what improves for the customer. Only then should the company decide AI’s role. An unattributed phrase heard in Singapore captured this sequence: “Start with a painful workflow, not a model.”

Starting with a model encourages a search for tasks it can perform, while starting with painful work opens the whole process to examination. The adviser asks leaders to go further: “If we were designing this process today, knowing what AI can do, would it look anything like this?” This question invites leaders to reconsider existing steps, approval responsibilities, information movement, and employee time. It turns a technology decision into a decision about how the business should operate.

Examining the work this way should guide a fundamental redesign of the selected areas. Teams can experiment with AI’s contribution while leadership remains responsible for the intended economic or customer outcome. Experimentation then has a purpose, with room to learn as the design develops. Repeated experience with consequential work can become a capability the organization carries into its next redesign.

Competitive advantage develops through organizational learning

Experience matters because, in the adviser’s view, higher spending alone may leave disappointing results unresolved. Companies leading with AI are developing proprietary data, redesigned workflows, organizational learning, and experience deciding where people and machines work best together. Proprietary data is information held by the company that competitors cannot readily obtain; organizational learning is the knowledge built through decisions and their consequences. The adviser sees these capabilities as sources of advantage that competitors cannot easily purchase.

These capabilities can reinforce one another as the company changes its work. A redesigned workflow gives teams experience with new responsibilities and ways of dividing tasks, which informs later decisions. Company-held data can support further learning, while that learning helps the business improve its data use and workflows. Their value grows through use in the company’s own operations.

Model access offers a different kind of resource because competitors can obtain the same technology. The adviser states, “You can buy the same model your competitor uses.” The adviser then adds, “You cannot buy several years of learning how to redesign your business around it.” Shared access can therefore coexist with very different levels of experience in putting the technology to work.

For leaders, the practical question is whether each AI effort creates knowledge that improves later choices. An experiment contributes more when teams can use what it teaches in the next redesign. Doing so requires attention to the consequences of changes in information, responsibilities, and how tasks are divided. Some of the most important lessons concern the work people should retain.

Deliberately choose the work humans should retain

Choosing work for people requires a deliberate decision about their contribution. The adviser asks CEOs, “What do we deliberately want humans to keep doing?” This challenges the assumption that every task technology can automate should be automated. It puts the desired role of people into the design from the beginning.

As machines become able to search, summarize, analyze, draft, and coordinate, the adviser emphasizes human judgment, creativity, empathy, supervision, and relationships. These contributions give leaders criteria for allocating work, though their importance will vary by task and organization. Leaders can remove mechanical tasks that obstruct valuable human work while preserving the responsibilities that work requires. The allocation should follow what the business needs people to do well.

These choices affect employees directly, so the adviser also calls for support in dealing with fears about AI. Leaders should help people work through those fears and consider what becomes possible as tasks change. This support belongs within the problem-first approach because employees take part in the redesigned work and experience its consequences. Decisions about their roles need leadership attention throughout implementation.

Governance, the rules for AI’s authority and oversight, makes those decisions explicit. Leaders must identify where AI may act autonomously, where human participation remains necessary, and where AI should be kept out. As capabilities change, they must revisit those boundaries while continuing to consider the effects on people and society. Each boundary should make clear who can act, who must participate, and who remains responsible for the outcome.

Key executive takeaways

• Measure business change: CEOs and boards need to assess how AI changes decisions, workflows, and employee time. License purchases and pilot activity measure deployment; business outcomes reveal whether that deployment creates value.
• Review workflows alongside returns: More than 80% of CEOs in Bain’s 100-person survey reported AI transformation results below 70% of expectations. Boards reviewing disappointing returns can examine approval steps, decision authority, and information flows for barriers that technology leaves unresolved.
• Set the investment ambition: CEOs can use Improve, Innovate, and Build to distinguish efficiency gains from changes to the core business or new offerings. Each investment needs an explicit goal because faster existing work does not automatically create lasting competitive advantage.
• Start with consequential friction: Business owners can select a handful of workflows where delay, cost, errors, or customer frustration materially affect results. Define the desired outcome, redesign the process, and then determine AI’s role.
• Build learning into implementation: Companies develop AI advantage through proprietary data, redesigned workflows, and experience allocating work between people and machines. Project owners can carry lessons from each redesign into subsequent investment and operating decisions.
• Define human roles and AI boundaries: CEOs need to identify where judgment, creativity, empathy, supervision, and relationships matter most. Governance owners can then specify where AI acts autonomously, where people participate, and who remains accountable, revisiting those boundaries as capabilities change.