The AI-era leader’s real job: redesign work, not deploy tools.

Many organisations still approach artificial intelligence as a cost-reduction programme: add a few chatbots, place an AI layer over existing applications, track usage and expect transformation.

That may produce local efficiencies. It rarely changes how the organisation works.

The distinction matters. An employee can draft an email faster while the decision behind it still takes a week. A team can generate a new product offer in five minutes while fragmented data and a long approval chain prevent it from reaching the market. Testing can be automated and completed in days, yet customer acceptance can still take months because the customer does not trust the evidence and repeats the tests manually.

In each case, AI has accelerated a task. It has not accelerated the outcome.

This is why the real job of an AI-era leader is not simply to deploy tools. It is to redesign work end to end—including data, decisions, authority, controls, trust and learning.

The productivity trap

Emerging research gives leaders good reasons for optimism, but also a warning against confusing individual productivity with organisational transformation.

A field experiment involving 7,137 knowledge workers across 66 firms found that frequent users of an integrated generative-AI tool spent around two fewer hours on email each week. Yet the researchers did not detect a broader change in the quantity or composition of work. Their central observation is important: people can independently change how they write an email, but changing meetings, handovers or coordinated workflows requires others to change too.

Other studies demonstrate how powerful AI can be inside the right task. Research involving 5,172 customer-support agents found a 15% average increase in issues resolved per hour, with much larger gains among less-experienced workers. In another field experiment, individuals using AI matched the performance of conventional two-person teams on the innovation tasks being studied. AI also helped technical and commercial specialists produce more balanced solutions.

These are meaningful results. But they do not support a simple instruction to “use AI everywhere.” They show that AI can change the economics of particular tasks, redistribute access to expertise and even alter the boundaries of a team. The leadership challenge begins after that capability becomes available.

When the bottleneck moves

Consider offer creation in a telecom business. AI may help generate the design of an offer in minutes rather than days. But end-to-end lead time may barely change if the underlying customer, product and pricing data remain fragmented across systems. Even with adequate data, the offer may wait a week for approval through the existing chain of command.

The faster design step has not removed the constraint; it has exposed the next one.

The same pattern appears in software testing. AI can generate tests, expand coverage and execute a user-acceptance test cycle in a few days. But if the customer does not trust AI-produced evidence, they may repeat the tests manually before signing off. A technical bottleneck has then become a trust and assurance bottleneck, and a process that could finish in days may still take months.

This is the mistake leaders make when they optimise an activity without examining the system around it. The relevant measure is not how quickly AI performs its step. It is whether time, quality, risk or customer outcomes improve across the full value stream.

The answer may be data readiness, an approval forum, an unresolved exception, a contractual control, a handover between functions or simply a lack of trust. That next constraint belongs inside the AI transformation—not outside its scope.

“Human in the loop” is not a design

The usual response to AI risk is to put a human in the loop. It sounds responsible, but it is incomplete. Which human? Reviewing what? At what point? Against which evidence? With what authority and accountability?

A meta-analysis of 106 experiments found that human–AI combinations performed better than humans alone on average, but worse than whichever of the human or AI performed best independently. The results varied substantially by context, and decision tasks were especially difficult. The conclusion is not that people and AI cannot work well together. It is that combining them does not automatically create a superior system.

Requiring human approval for every AI output can also recreate the old process with an additional step. It limits scale, diffuses accountability and encourages superficial rubber-stamping. Good governance is not the maximum possible human involvement. It is the right human involvement at the right decision point.

Decide autonomy through risk and reversibility

A more useful principle is to calibrate AI autonomy according to two factors: the potential impact of an error and the ease with which the decision can be reversed.

If an AI system wrongly grants a customer a ten-dollar billing adjustment, the loss is small and the action can be corrected. Subject to appropriate limits and monitoring, the system may be allowed to decide and act.

If an AI system could terminate a ten-million-dollar enterprise contract, the potential damage is high and the decision may be difficult—or impossible—to reverse. That decision needs explicit human authority.

The boundaries should not remain static. As evidence accumulates, confidence improves and reversal mechanisms become stronger, leaders may expand autonomy. If failures emerge or operating conditions change, they should narrow it again.

Five questions for redesigning work

Leaders can use a simple sequence to move from tool deployment to operating-model change:

1. Task — What should AI actually do?

Define the unit of work precisely. Is AI generating, predicting, deciding, executing or monitoring? Avoid treating an entire role or process as one uniform task; AI capability is uneven.

2. Decision — What judgment follows the output?

An output is not automatically a decision. Identify who interprets it, what other context is required and how disagreement is handled.

3. Authority — Who may act, and who remains accountable?

Specify whether AI advises, acts within thresholds or executes autonomously. Accountability must remain explicit even when execution is automated.

4. Guardrail — What determines review or escalation?

Use risk, reversibility, confidence, monetary exposure and affected stakeholders to define boundaries. Do not apply one universal approval rule to every decision.

5. Learning — How will the system and the people improve?

Track overrides, exceptions, errors and delayed consequences. Preserve opportunities for people to build expertise rather than allowing AI to remove every developmental task.

This sequence forces leaders to examine the complete work system rather than a single automated step.

The leadership shift

AI changes leadership before it eliminates work. As more first drafts, analyses and routine actions are produced by machines, leaders will spend less time supervising task execution and more time designing the conditions under which decisions are made.

That requires several capabilities at once: enough AI fluency to understand what the technology can and cannot do; enough domain depth to recognise when an answer is plausible but wrong; systems thinking to anticipate where constraints will move; judgment to calibrate autonomy; and human leadership to build trust, accountability and learning.

The organisations that create the most value from AI will not necessarily be those with the most copilots, agents or licences. They will be those whose leaders are willing to question processes built for a different era.

The test of transformation is not whether AI is present in the workflow. It is whether the workflow itself has become faster, better, safer and more capable of learning.

The practical examples and interpretations in this article reflect Saket Kumar’s personal perspective. They do not represent the views of any employer or customer.