Stop Buying Copilots. Start Redesigning Workflows

Stop Buying Copilots. Start Redesigning Workflows

The most common AI investment gives an employee a faster way to perform an existing task.

The employee drafts the email sooner. Summarises the document faster. Creates the presentation in less time. The demonstration looks impressive, usage rises and the organisation reports hours saved.

Yet the customer still waits. The decision still needs four approvals. The information still moves between disconnected systems. The cost base barely changes.

That explains one of the central contradictions of enterprise AI: individuals are becoming more productive, but relatively few organisations are capturing meaningful value.

The problem is not necessarily the copilot. It is the workflow surrounding it.

Faster tasks, unchanged outcomes

McKinsey’s 2026 global AI survey identifies just 6% of respondents as AI high performers—organisations reporting both significant value and at least a 5% contribution to EBIT from AI. Nearly three-quarters of those high performers say they are fundamentally redesigning workflows, compared with approximately one-quarter of other respondents. McKinsey’s 2026 State of AI survey

This does not prove that workflow redesign alone caused the superior results. High performers also invest more and report stronger leadership commitment. But it reveals an important association: the organisations reporting substantial value are not simply asking employees to perform the same work more quickly.

They are changing how work moves from beginning to end.

A copilot improves an activity. A redesigned workflow improves an outcome.

That distinction determines whether AI produces convenience for an individual or an advantage for the business.

The bottleneck merely moves

Consider a customer request that passes through six stages:

Suppose AI reduces the information-collection stage from four hours to ten minutes. That sounds like a major productivity gain.

But if the request then waits two days for managerial approval, the customer experiences almost no improvement. The employee may be faster, while the workflow remains slow.

BCG describes a comparable case in which AI reduced a task from ten days to one, yet customers continued waiting ten days because the surrounding process had not changed. The organisation improved the outcome only after removing intermediate steps and unnecessary committees. BCG’s analysis of why AI pilots fail to create business value

This is the danger of task-level automation: it can optimise one section of a system while leaving the constraint somewhere else.

In some cases, it produces more work. Faster content generation creates more material to review. Faster coding produces more software to test and maintain. Faster lead generation creates a larger qualification backlog.

The organisation has not eliminated work. It has moved—and sometimes multiplied—it.

The strategic choice

Leaders now face three different investment paths.

These are not mutually exclusive. General-purpose copilots can build familiarity and help employees discover use cases. Bottom-up experimentation remains useful because employees closest to the work often see opportunities that executives miss.

But discovery is not transformation.

Andrew Ng argues that bottom-up AI experiments frequently fail to produce substantial gains until leaders redesign the broader workflow. His example of loan processing shows how accelerating one approval step becomes strategically meaningful only when marketing, routing, final review and execution are redesigned around a faster customer proposition. Andrew Ng on moving from AI experiments to AI products

The question is therefore not whether employees should have copilots. It is whether the organisation knows how a promising individual use case graduates into a redesigned business capability.

What successful redesign requires

The Stanford Digital Economy Lab studied 51 successful AI deployments across 41 organisations. The researchers found that similar technology and use cases could produce markedly different results. Organisational readiness, process design, leadership and willingness to change were more consequential than the choice of model alone. Stanford Enterprise AI Playbook

This suggests a practical sequence.

Start with an outcome

Choose a business measure, not a tool or task.

Examples include reducing customer-resolution time, increasing the proportion of claims processed correctly, improving conversion, shortening the cash cycle or lowering the cost of producing a compliant proposal.

Map the complete workflow

Identify every activity, decision, handoff, queue, exception and approval between demand and outcome.

The most important discovery may have nothing to do with AI. A report may exist only because an old manager once requested it. An approval may persist despite the underlying risk disappearing years ago.

Simplify before automating

Remove unnecessary work first. Consolidate duplicate checks. Clarify decision rights. Standardise the inputs required for repeatable decisions.

Automating a process before simplifying it makes obsolete work faster and more difficult to remove later.

Redesign human and AI responsibilities

Decide which work AI should prepare, recommend, execute or monitor. Specify what humans must judge, approve or own.

The boundary should depend on consequences and reversibility. A low-value formatting error and an incorrect financial approval should not receive the same level of autonomy.

Measure the outcome and the checking burden

Track cycle time, quality, cost, revenue, exceptions and customer impact. Include the time employees spend reviewing and correcting AI output.

Minutes saved in one activity are not valuable if they create more minutes of verification elsewhere.

The recommendation

Do not eliminate general-purpose copilots. Stop treating their distribution as the centre of your AI strategy.

Maintain a controlled, proportionate layer of general AI access for learning and individual productivity. Then redirect the next portion of incremental AI investment toward two or three end-to-end workflows tied to material business outcomes.

Each workflow should have:

Avoid beginning with the most complex or politically important process. Start with a workflow that repeats frequently, has measurable outcomes and contains errors that can be detected and recovered economically.

The recommendation changes when personal productivity is itself the product—for example, in highly autonomous creative or advisory roles. In those cases, broad copilot access may directly increase valuable capacity. Even then, leaders must decide how that capacity will be converted into revenue, quality, learning or customer value.

Otherwise, “hours saved” remains an unclaimed benefit.

Your Monday-morning move

Choose one AI use case currently described as successful.

Draw the complete workflow from initial demand to customer or financial outcome. Mark every queue, approval, handoff and exception. Then ask:

That remaining constraint—not the fastest available model—should determine your next investment.

A copilot helps an employee do the work. Transformation begins when the company changes the work itself.

Sources and why they matter