McKinsey: AI speeds up work, but EBIT stays flat
4 min read
Eight in ten respondents in the new McKinsey survey say AI has made them faster at work. Little of that reaches the income statement. This gap has persisted for a second year.
Key takeaways
- Individuals are working faster: Most notice faster work and better decisions. The contribution to company earnings stays flat.
- Large companies scale agents while the rest stand still: Large organisations are pulling ahead on agents. Smaller ones remain at last year’s level.
- The key lies in the workflow: High performers redesign work. Companies that simply buy seats pay operating costs.
Related:Copilot in the Mittelstand: What seats really cost / Agents cost providers five times as much as chat
What is an AI high performer? McKinsey includes only organisations that attribute at least five percent of their earnings before interest and taxes (EBIT) to AI and describe the impact as significant. In the current survey, around six percent of respondents qualify.
The gap shows up in the income statement
Between early May and early June 2026, McKinsey QuantumBlack surveyed around 1,700 professionals and managers in 97 countries. Almost nine in ten regularly use AI in at least one function. The personal impact is tangible: Eight in ten feel more productive, and half make better decisions.
The company as a whole tells a different story. Only 37 percent attribute any contribution to earnings before interest and taxes to AI. That share has barely changed since 2025. High performers under McKinsey’s strict definition remain a small minority.
The gap is no measurement error. Employees resolve tickets faster. Approvals, handovers and invoicing continue as before. The minutes saved disappear wherever the next bottleneck remains unchanged.
Copilot sits on top of the old workflow
McKinsey draws a clear distinction: High performers use AI for growth and innovation rather than efficiency alone. Almost three quarters fundamentally redesign workflows. Only about a quarter of the rest do so. The difference lies in the work itself rather than the model.
Adding chat alongside the old process manual speeds up the individual while leaving the chain intact. Quotation, approval, booking, dispatch: Every step has an owner, a waiting time and an exception. Without redesign, AI merely shifts the queue to the next stage.
Those handling daily operations feel the strain. Almost half of middle managers and professionals without management responsibilities report burdens caused by AI. Among senior leaders, the figure is just under a third. Speed at the top, friction in the middle: This is precisely where potential gains go unrealised.
“
AI raises employee productivity. The financial impact fails to keep pace.
– Editorial team, based on McKinsey QuantumBlack, State of AI 2026
Four checks before buying more seats
A business needs a path it can follow within a quarter. The survey provides the lesson rather than a software shopping list.
- Choose a workflow that moves money. Quotations, order clarification, invoicing or procurement. A chat tool that merely replaces internal emails remains invisible in the income statement.
- Measure first. Cycle time, error rate, rework and days of tied-up capital. Without a baseline, every Copilot success story is just a feeling.
- Redesign the workflow. Remove steps, consolidate approvals and reduce handovers. Integrate the model into the redesigned workflow rather than placing it alongside.
- Assign an owner and cap costs. One person is accountable for the metric. Operating costs, including tokens, get a limit before seat numbers grow. Around a fifth of respondents are already reducing AI use because of these costs.
High performers are twice as likely as the rest to have leadership backing AI adoption and to measure its impact.
Source: editorial analysis based on McKinsey QuantumBlack, The state of AI in 2026.
The process comes before the seat
Large organisations with revenues in the billions are scaling agents much faster: four in ten, up from just over a quarter the previous year. Smaller organisations remain at around 22 percent. Companies that distribute seats without taking ownership of a workflow buy into precisely this gap.
Expectations of job cuts are rising. Almost four in ten expect an AI-driven reduction in the coming year. Actual reductions over the past year were far below what the previous year’s survey had predicted. Cutting staff before a process delivers visible results in the income statement is speculation.
The decision is simple: One workflow, one metric, one cap. If no impact linked to EBIT emerges after 90 days, Copilot remains a tool for individuals. Fewer seats are then the right answer.
Frequently asked questions
Is Copilot enough to turn AI into earnings?
No. The survey shows that eight in ten respondents work faster, while AI’s EBIT contribution stays flat. Without a redesigned workflow, the gains remain confined to individuals.
Which metric matters after 90 days?
A measure that captures costs or tied-up capital alongside cycle time. Tickets per hour alone are insufficient because they leave the bottleneck in the chain invisible.
What sets high performers apart?
They redesign workflows, measure the impact and secure leadership backing for adoption. Only around six percent of respondents meet McKinsey’s definition.
Should we adopt agents because large companies are scaling them?
Only if the chosen workflow already has a baseline and an owner. Agent scaling in smaller organisations has barely advanced over the past year.
Why doesn’t faster work show up in the income statement?
The minutes saved disappear into unchanged approvals and handovers. The next bottleneck remains in place; the individual simply reaches it faster.
Read more on MyBusinessFuture
MyBusinessFutureDocument AI: Where bureaucracy costs really fallMyBusinessFutureAI success without an expert: Who owns the outcomeMyBusinessFuture95 percent of AI pilots deliver nothing, 5 percent doMore from the MBF Media Network
cloudmagazinKubeflow: CNCF declares AI platform matureDigital ChiefsStripe acquires OpenRouter, Ramp opens Router.comSecurityTodayZimbra: One email is all it takes via SNMPImage source: AI-generated (August 2026)
Translated from the German original using artificial intelligence. The German version is authoritative.
Also available in
