AI time savings depend on redesigning how teams work

AI adoption is different from work redesign

An AI rollout can save time while leaving an organization with harder decisions: which work machines should perform, where people should keep judgment, who gets authority to act on new information, and where saved capacity should go. Rebecca Hinds, Head of the Work AI Institute at Glean, frames AI adoption around those decisions. Her institute studies how AI changes people and organizations, and her central question for leaders is simple: “How do we use AI to make work better?” Glean sells workplace AI technology, so it benefits commercially when organizations expand productive uses of AI.

Hinds approaches that question through organizational behavior. She earned a B.S., M.S., and Ph.D. from Stanford and says her career has focused on how collaborative and emerging technologies change organizations. For the past fifteen years, she says, she has worked with organizations around the world on dysfunctional collaboration, including meetings that consume excessive time and energy. Her recent bestselling book, Your Best Meeting Ever, covers much of that work on meetings and collaboration.

Over the past seven years, Hinds says her focus has shifted increasingly toward AI, which she treats as an organizational change as well as a technology deployment. That makes work design the management problem: “What should AI automate?”, “What should AI augment?”, “What should remain deeply human?”, and “And what will we do with the time we gain?” Each question assigns a role to technology and people. Together, the answers determine whether an efficiency gain changes how the organization works.

The first payoff is deciding where saved time goes

The fourth question becomes concrete at the management layer because much managerial work consists of coordination. Hinds says managers she describes as high AI achievers offload 32% more coordination work to AI. That transfer creates capacity managers can assign elsewhere. Hinds focuses on what they do with it.

The intended sequence moves from coordination work shifting to AI, through managers recovering time, to managers deliberately spending more of that time on people. Hinds identifies coaching and mentoring as examples of work where human involvement remains especially valuable. She also points to helping employees develop AI skills, so some capacity created by AI can be reinvested in improving how people use the technology itself. Saved hours become valuable when management gives them a destination.

That destination requires a management decision because automation cannot decide which neglected responsibility deserves more attention. A manager who spends less time coordinating work has several ways to use the difference, and the choice depends on the team’s needs. Hinds’s framework makes the allocation of time part of deployment itself. The operating question becomes what the organization wants managers to do more of once routine coordination demands less of them.

Coaching and mentoring then clarify what Hinds means by work that should remain deeply human. Her framework separates transferable tasks from activities where human judgment and development are central, then uses automation to create capacity for the latter. That distinction changes how leaders should evaluate efficiency because recovering time is an intermediate result. The organizational result depends on how that capacity is reassigned.

Better AI information can fail when nobody has authority to act on it

Reassigning capacity addresses existing work, but AI can also produce information that existing roles were never organized to use. Hinds says research she conducted with colleagues found that AI can conflict with static organizational charts. Pattern-finding algorithms can identify relationships that cross departments and other organizational categories. A fixed reporting structure can then leave an important finding between people who each have formal responsibility for only one part of it.

That gap turns information discovery into a question of authority because someone needs a mandate to respond across the same boundaries when AI surfaces a useful relationship across organizational silos. Existing roles may give several people responsibility for pieces of the issue while leaving authority over the whole unclear. Hinds argues that human roles may consequently have to be rebundled so a person can act on the connections the system identifies.

Rebundling matters because better analysis has limited value when the decision structure prevents action. An organization can deploy AI broadly and generate cross-functional insights while preserving job boundaries designed around older information flows. Adoption can then move faster than organizational design. Leaders need to examine decision rights and role boundaries alongside workflows when AI starts finding relationships across functions.

Those cross-functional findings also affect technology selection because the tools themselves may need to span functions. Hinds argues that organizations should consider tools that work across functions because AI adoption crosses functional lines. She also characterizes higher-performing organizations as strengthening the relationship between HR and IT. That pairing follows from the work involved: IT decisions about AI affect roles, trust, culture, and everyday practices that HR and business leadership also govern.

Once functions become more permeable, the same logic can change how teams are assembled. Hinds expects AI to surface employees according to their skills, available bandwidth, and career ambitions when a project needs staffing. Those signals can point beyond the relationships represented by a conventional reporting chart. An employee with the right expertise and capacity could become visible for a project even when that person sits elsewhere in the formal hierarchy.

That broader visibility addresses a limitation of staffing that starts from reporting relationships. Relevant skills, spare capacity, or an employee’s ambitions can remain hidden when the formal structure primarily records organizational position. Employees whose useful attributes are absent from that structure can consequently miss work for which they might be suitable. AI-based staffing, as Hinds describes it, would make those additional signals part of deciding who performs the work.

As staffing becomes more responsive to those signals, Hinds still expects the hierarchy to have a purpose. She expects organizational charts to become more flexible and increasingly represent where people and responsibilities formally sit, while project staffing responds more directly to the work. Hinds forecasts substantial changes along these lines over the next five years. The direction follows the authority problem in her cross-silo research: richer signals make reporting lines a less complete description of who should do particular work.

For leaders, the result reaches beyond selecting another AI use case because information and authority have to meet. AI may reveal relationships that the organization previously missed, but somebody still needs the authority, capacity, and role definition to act. Work redesign consequently extends into organizational structure itself. Better information has limited value when formal responsibility ends at the boundary the information crosses.

Human judgment matters where AI looks more capable than it is

Structural authority solves the problem of who can act, but it cannot establish whether an AI system understands enough to support the decision. Hinds identifies apparently polished or finished work that is actually inadequate as one failure mode teams discover as they use AI more extensively. She also identifies tasks where human judgment remains necessary. Both cases make presentation a weak basis for deciding how much responsibility a system should receive.

Hinds instead proposes tying trust to context: how much the AI understands about the person involved, the situation, and the history behind the decision. A fluent or confident answer can coexist with missing context, so surface quality cannot establish whether the system understands enough for a consequential judgment. The contextual requirement rises with the stakes of the decision. Trust therefore becomes a workflow-design choice tied to what the system needs to know.

Those stakes already extend to major workplace decisions. Hinds says AI has moved into hiring, firing, and performance decisions, and she reports that 29% of workers are comfortable with AI firing human colleagues. The figure shows the level of responsibility involved when organizations decide how much judgment to delegate. For decisions that materially affect a person’s employment, Hinds’s context test makes the system’s understanding of the person, situation, and history central to the design.

Delegation also reaches collaborative settings where AI can stand in for an employee. Hinds says one out of six meetings now receives a “digital twin” instead of a person. Representation creates its own contextual demand because a system acting in place of an employee needs enough information to represent what matters in that setting. The risk grows when the apparent completeness of the system’s participation exceeds what it knows about the person and situation.

That gap between apparent capability and adequate understanding also explains why “Botsitting” is a useful concern to distinguish from productive human oversight. The term suggests a productivity problem in which people spend newly saved time supervising AI, while Hinds’s examples establish a concrete need to inspect polished-looking work and preserve human judgment for some tasks. Leaders can address that need by defining the context and review a task requires before assigning responsibility to AI. The design problem is how much judgment the workflow can safely delegate.

Use AI intensively to discover its value and its boundaries

Deciding that boundary in advance can be difficult, so Hinds proposes a period of deliberately aggressive experimentation. Her “AI Immersion Week” asks a team to go “all in” on AI for an entire week and encourages people to use it on every task, whether large or small. The purpose is diagnostic. Broad use gives a team evidence about where automation and augmentation help across work that employees might otherwise never test.

The first stage of the exercise is discovery because team members extend AI across their full range of work. That wider use can reveal useful applications beyond the obvious use cases. Hinds argues that people’s assumptions about AI can constrain experimentation before the technology itself does. People interact with the same tool differently according to the mental model they bring to it, so leaders have to shape those mental models deliberately.

Once broader use reveals unexpected successes, the same exercise can expose meaningful boundaries. Intensive use brings out tasks where AI performs badly, including outputs that seem polished but fail closer inspection, as well as tasks where human judgment remains necessary. Hinds calls the value of the exercise “twofold” because successes and failures both provide information. A failure identifies where critical thinking and human involvement need to remain in the workflow.

An immersion week therefore uses extensive adoption as an experiment rather than as the final measure of success. During that week, broad use helps teams classify work according to what can be automated, what benefits from augmentation, and what still requires human control. Those observations give leaders a basis for drawing operational boundaries. Experimentation becomes an input to work design.

That diagnostic purpose also gives implementation leads a concrete role after experimentation. A team should examine what changed in each task: where AI created useful capacity, where it improved a person’s work, and where insufficient context or judgment made the result inadequate. Leadership then has to convert those observations into expectations for future workflows. The mental model leaders encourage should make useful deployment and critical inspection normal parts of working with AI.

AI can also redesign what happens before people enter the room

The augmentation category becomes more specific in Hinds’s own meeting practice because AI can change the preparation before group discussion. She points to prior research finding that teams generate better ideas when people first formulate their thinking independently before a group discussion begins. The mechanism concerns the order in which people think and speak. An early or loud voice can steer immediate discussion toward “groupthink” before other participants fully develop their positions.

AI can create a private stage before that group interaction. A person can develop an idea, challenge it, and prepare a position before committing to it publicly in a meeting. That sequence can matter particularly for quieter participants whose ideas might otherwise be overridden by the first voices in open discussion. The potential benefit comes from changing when thinking happens and how prepared participants are when group interaction starts.

Hinds uses AI this way before meetings by asking it to be an adversarial or contrarian reviewer. She gives it a proposal and asks it to argue against the proposal, identify its weakest assumption, or adopt a position that nobody in the meeting is likely to raise. The practice keeps the eventual judgment with the person while using AI to widen the thinking that precedes it. In this workflow, augmentation changes the preparation stage so the human enters the meeting with a more tested position.

Key takeaways for leaders

• Reinvest the time AI saves: Managers who offload coordination work to AI need a deliberate plan for the recovered capacity. Direct it toward high-value human work such as coaching, mentoring and developing employees’ AI skills.
• Align authority with AI insights: AI can surface relationships that cross existing reporting lines and functional boundaries. Organizations need to adjust decision rights, roles and project staffing so someone has the authority to act on those findings.
• Match AI autonomy to context: Fluent output does not prove that AI understands the person, situation or history behind a decision. Workflow owners need stronger human review as context requirements and consequences rise, especially in employment decisions.
• Use intensive experimentation to set boundaries: An AI immersion period can expose useful applications, weak outputs and tasks that still require human judgment. Implementation teams can use those findings to decide which work to automate, augment or keep under human control.
• Use AI to improve thinking before meetings: Employees can use AI as a contrarian reviewer to test assumptions and develop ideas independently before group discussion. This preparation can broaden perspectives and reduce the influence of early or dominant voices.