AI returns depend on who can change workflows

The consequential AI decision is who can change the work

A team can have AI access and still lack permission to improve delivery. If every workflow change requires committee approval, the team remains accountable for results while leaders control how it can achieve them. For an executive pursuing enterprise-wide gains, the key decision is how much authority teams will have to change the work and what evidence they must produce afterward.

The conventional adoption model leaves that decision unresolved. Organizations buy tools, preserve existing procedures and report seats or usage, even as efficiency remains weak, competitiveness improves slowly and quality stays expensive. Individual AI use can make particular tasks easier or faster. To turn those gains into enterprise performance, leaders must examine how faster tasks become completed work, lower costs and better business results.

Enterprise returns make workflow redesign worth examining

Financial results give leaders a reason to examine that connection. McKinsey reports that 39% of organizations see any AI-related impact on EBIT, or earnings before interest and taxes, with most reporting an impact below 5%. Buying and deploying AI therefore leaves a substantial business question unanswered: whether the investment changes the economics of the work it supports.

MIT’s findings bring that question down to individual pilots. It found that most enterprise generative-AI pilots produced no measurable return on the P&L, or profit and loss statement, and attributed weak results to integration difficulties and a learning gap. A pilot can show that a tool performs a task while the organization still struggles to make it part of everyday operations. Financial value depends on what happens next.

Integration difficulties can begin with the structure of the existing workflow. An organization divides work into stages on the assumption that people will carry out each stage manually, then adds an AI model while preserving the sequence. The model may speed up an activity, but the surrounding handoffs, reviews and decisions continue as before. People learn to use the tool within established responsibilities and procedures, while the process retains much of its original cost and delay.

McKinsey’s findings make workflow redesign relevant here: high performers capturing AI impact were nearly three times more likely to have fundamentally changed their workflows. The reported relationship is an association between stronger performance and redesign. It gives leaders a specific practice to examine when investment disappoints.

MIT and McKinsey address different parts of the management problem. MIT links weak pilot returns to integration and learning difficulties, while McKinsey links stronger impact to substantial workflow changes. Their findings concern different measures and should remain distinct. To understand how much an individual developer gains from AI access, leaders need task-level evidence.

Metr’s experiments show why productivity requires careful measurement

The task-level question produced a striking result in Metr’s early 2025 randomized trial. The AI research organization studied experienced open-source developers and varied whether they could use AI. When AI was permitted, task completion took about 19% longer, yet the developers believed they had completed the work roughly 20% faster. Their experience of using the tool pointed in the opposite direction from their measured performance.

The gap matters because confidence in a tool can become an informal substitute for evaluation. A developer’s judgment that AI helped may accurately describe how the work felt while misrepresenting how long it took. So leaders relying on enthusiasm alone would have drawn the wrong conclusion about task speed in this trial.

When Metr repeated the experiment later in 2025, the estimates changed direction. Among the 10 developers who returned from the original study, the estimate suggested work was completed about 18% faster. The estimated improvement among 47 newly recruited developers was about 4%. These were separate groups, so combining their estimates into a single result would obscure differences between the participants.

Both later estimates also carried substantial statistical uncertainty. Their confidence intervals, the ranges expressing uncertainty around the estimated effects, straddled zero, including both a possible speedup and a possible slowdown. So the direction of the effect remained uncertain for each group. The percentages alone would give an incomplete account of the results.

Participation introduced a separate measurement problem. In Metr’s later experiment, a growing share of developers declined to participate because the experiment required them to spend half their time working without AI. Those willing to accept that restriction could differ from developers who considered access too important to give up. The restriction therefore affected which developers the experiment could measure.

Task selection narrowed the evaluated work further. In Metr’s later experiment, between 30% and 50% of participants admitted withholding the tasks for which they particularly wanted AI, so the experiment missed work they expected to benefit most from assistance. Metr treated its estimate as a lower bound: the actual gain was probably larger, but it could not quantify how much larger. That interpretation accounts for selection’s likely downward effect while leaving the statistical uncertainty intact.

The two experiments show how evaluation can mislead in different directions. In the first trial, developers’ impressions overstated their performance; in the later experiment, participation decisions and withheld tasks likely understated the measured benefit. Even an organization specializing in AI measurement struggled to determine the effect’s size clearly. These behaviors describe developers’ responses to experimental restrictions on access and task submission, so they provide a narrow basis for interpreting the experiment.

For a business, the next question is how task-level gains affect delivery. The later findings leave room for useful individual assistance without establishing that participants fundamentally redesigned their workflows. Organization-wide transformation is therefore unnecessary for every beneficial AI task. Enterprise returns require a further connection: improvements in particular activities must translate into better delivery and financial outcomes.

Redesign requires end-to-end ownership and permission to remove steps

To make that connection, a team must first establish what its work is meant to achieve. It can then reconsider the sequence in light of what AI makes possible, asking which stages still contribute to the result. Purpose opens the process to change. Starting with the inherited procedure encourages the team to preserve it while speeding up isolated tasks.

Reconsidering the sequence creates an opportunity to remove coordination and review that the revised process no longer needs. Candidates include unnecessary handoffs, review layers that have lost their function and steps justified by the assumption that people would complete work at human speed. Each removal should connect to a measurable business effect. Reviews and approvals that still serve a useful purpose should remain.

End-to-end ownership gives someone responsibility for making those judgments across a complete segment of work. When AI reduces coordination previously spread across a whole department, one person or a small team may be able to take responsibility for that segment. The proposed gain comes from changing how responsibilities connect as well as completing individual activities faster. Leaders can test that design against outcomes; team size alone guarantees no improvement.

Such ownership requires explicit decision rights: the operating decisions a team is authorized to make. A team cannot meaningfully redesign its work if each adjustment returns to a steering committee that preserves the original arrangement. Leaders must specify which decisions the team controls and hold it responsible for the consequences. Otherwise, they assign a performance target while retaining the procedural controls that constrain it.

The mismatch becomes concrete when a team cannot remove approval layers, take responsibility for a complete work segment or act without decisions made elsewhere. In each case, the team remains accountable for a result that depends on choices outside its control. Defining authority alongside the target gives the team a way to meet its responsibilities. Leaders can then assess the changes through the work delivered.

Measure delivered work and reward teams for changing how it gets done

Once teams have authority to act, measurement must show what their decisions produce. Seats, logins and the percentage of staff using AI help leaders understand access and participation. Evidence of business value requires a further step: showing whether work was delivered, demand was met, revenue improved or quality increased. A board reviewing AI investment needs that connection between adoption and performance.

Work-based measures make the connection more concrete. Cycle time shows how long a unit of work takes from start to completion, while cost per unit shows what the organization spends to produce it. Output per person indicates how much each person delivers. Together, these measures let leaders test a team’s claimed improvement against changes in the work itself.

Quality and revenue extend that assessment because faster or cheaper production can still miss the business objective. Quality establishes whether increased output remains useful, while revenue connects operational change to commercial performance. The right combination depends on the work being changed. Leaders should require evidence across those measures before treating a pilot or deployment as a demonstrated business result.

The evaluation method matters alongside the measures, as Metr’s experiments illustrate. Users’ impressions can diverge from measured performance, and the people and tasks included in an evaluation can affect its result. Teams should therefore explain both how their results changed and how they assessed them. A favorable number is more useful when leaders understand which work and participants it represents.

An investment review can put that discipline into practice with a few operational questions. What work shipped, and which outcomes measurably improved? Which operating practices changed, and did the team have authority to change them? Useful individual assistance can justify a task-level investment, while a claim of enterprise transformation requires evidence that changed work is producing better business results.

Even clear evidence will have limited influence if rewards favor visible busyness, preserving headcount or obedience to inherited procedures. Those incentives make unnecessary coordination worth protecting. Leaders should reward successful workflow redesign, greater delivery with fewer resources and the elimination of a team’s own unnecessary work. Reducing avoidable work must count as an achievement if people are expected to pursue it.

Recognizing that achievement requires attention to results that may be less visible than usage counts or procedural compliance. A team that delivers more with fewer steps may generate less visible activity even as its performance improves. Leaders need to examine completed work and the resources it consumed so that removing effort earns credit. Otherwise, teams that preserve unnecessary activity can appear more productive than teams that eliminate it.

Key takeaways for leaders

• Connect AI investment to workflow changes: McKinsey found that high performers capturing AI impact were nearly three times more likely to have fundamentally changed their workflows. Investment reviews need evidence of how AI changes the cost, speed and quality of completed work.
• Test productivity claims carefully: In Metr’s early trial, developers believed AI made them faster while measured completion times increased. Evaluation owners need to examine measured performance, statistical uncertainty and which people and tasks the evaluation includes.
• Give workflow owners decision rights: Executives need to define which approvals, handoffs and operating decisions teams control. End-to-end accountability becomes actionable when teams have authority to change how they deliver results.
• Measure outcomes and reward work removed: Business owners need to track cycle time, cost per unit, output, quality and revenue. Performance incentives should recognize teams that improve delivery by eliminating unnecessary work.