AI can save an individual employee time. Turning that promise into dependable, organization-wide value is much harder.
McKinsey’s 2026 survey captures that tension. Around 80% of participants said AI had made their own work more productive, while only 37% said their employer was receiving a positive EBIT contribution from it.
Adoption also varies with organizational size. In the survey’s largest revenue tier—businesses generating $1 billion or more—40% were expanding their use of AI agents, while 31% were doing the same with coding agents. Among smaller companies, those shares fell to 22% and 17%.
Our analysis: the central challenge is not whether employees use AI, but whether organizations can convert scattered experimentation into owned workflows, reusable capabilities, and measurable business results without losing control.
Why productivity gains haven’t become enterprise value
Individual usefulness is encouraging, but it is not equivalent to organizational performance. An employee may complete a task faster while approvals remain fragmented, data remains inconsistent, or review work shifts elsewhere in the process.
That distinction matters to project teams. Heavy tool usage can create the appearance of progress without improving cycle time, cost, quality, or decision-making. If activity becomes the principal measure, teams may optimize for adoption rather than value.
The survey results point to a conversion problem. Employees perceive benefits that have not yet translated into broad earnings contributions. Closing that gap requires more than additional tools. Leaders need to clarify priorities, assign business accountability, redesign workflows, and measure results against a baseline.
For a project lead, the practical question changes: not “How much did people use the AI tool?” but “Which part of the workflow became faster, safer, cheaper, or better, and what evidence shows it?”
Compare three ways to approach scaling AI responsibly
| Approach | What it enables | Main trade-off | What to watch |
|---|---|---|---|
| Broad experimentation | Builds familiarity, encourages learning, and surfaces unexpected use cases | Tool sprawl, duplicated work, and confusion between activity and value | Actual time savings, usefulness, adoption patterns, and repeated investments |
| Tight central control | Creates consistency and limits uncontrolled deployment | Slow learning and knowledge trapped among specialists | Whether restrictions are preventing useful experiments |
| Managed portfolio | Combines focused learning with business ownership, measurable outcomes, shared foundations, and governance | More coordination and clearer accountability requirements | Pilot selection, sponsor engagement, reusable components, and measured business results |
The managed portfolio offers the strongest balance for most enterprise programs. It preserves room for individual learning while ensuring that major investments answer defined business questions.
Pilot selection should be deliberate. Favor fewer opportunities that combine meaningful business impact with lower risk, adequate data, stable processes, sufficient workflow volume, integration feasibility, and measurable results. The objective is to learn whether AI can improve a specific task and whether the result can work within the surrounding system.
Ownership must accompany that selection. A business sponsor should remain accountable for the outcome, with technology and business functions sharing responsibility. Technical delivery may prove that a system works, but it does not establish that the system benefits the enterprise.
Build governance into the work, not around it
Governance is too late when it operates only as a final approval gate. It should shape the workflow while that workflow is being designed and operated.
That includes decision rights, permitted data, risk tiers, test criteria, monitoring, auditing, escalation, incident response, and eventual retirement. AI may recommend an outcome or take action within defined boundaries, but accountability remains with people in authorized roles.
Automation should generally be reserved for relatively simple, well-understood, low-risk tasks. More complex decisions may require human-AI collaboration. When intervention is expected, employees need the time, skills, and authority to make a meaningful judgment rather than simply approve a recommendation.
Governance therefore has a dual purpose: it limits unacceptable risk while making safe action possible. Teams working this way are more likely to treat controls as part of delivery rather than as an obstacle placed in front of it.
Use pilots to reveal workflow problems
A well-chosen pilot should do more than test model quality. It should reveal whether AI exposes broken handoffs, duplicated controls, unclear decisions, or other process weaknesses that would limit value even if the technology performed well.
Establish a baseline before launch. Measure actual time savings, usefulness, development cost, workflow volume, and the burden created for reviewers. Adoption figures can show reach, but they cannot substitute for evidence that the work improved.
The pilot should also clarify its boundaries. Determine where AI recommends, where people intervene, where escalation occurs, and which decisions remain outside automated action. This makes risk visible before deployment expands.
The lesson should not be limited to the pilot team. Share findings, reusable patterns, and lessons from unsuccessful attempts. Broad experimentation remains useful, but learning only spreads when leaders deliberately preserve and redistribute it.
Build reusable foundations before duplication spreads
Scaling a workflow often reveals a need for shared services, common patterns, reusable data products, and consistent domain governance. Without an architectural view, individual teams may create separate stacks that solve similar problems in incompatible ways.
Reusable foundations let new initiatives move faster because they do not have to rebuild every platform, integration, or control from the beginning. They also make governance more consistent: decision rights and monitoring can become part of the foundation rather than something each project negotiates separately.
This is a trade-off. Common foundations require coordination and may not fit every specialized use case. They should be sound enough to scale, but flexible enough to accommodate changing platforms, standards, feedback loops, decision rights, and employee roles.
What this means for you
If you lead a team considering AI, begin with a valuable workflow and a measurable problem—not a catalogue of available tools.
Rank possible pilots by business impact, risk, data readiness, process stability, workflow volume, and integration effort. Select a smaller number of opportunities, then give each a business sponsor, a clear outcome, shared technology-and-business accountability, and a review point.
Next, examine the workflow around the task. Identify broken handoffs, unclear decisions, duplicate controls, and review work that AI might merely accelerate. If the process needs redesign, specify where human judgment is required and where lower-risk automation is appropriate.
Finally, look beyond the pilot. Determine which components can become reusable services or shared foundations. The organizations likely to convert productivity into enterprise value will be those that stop managing AI as a collection of unrelated tools and start managing it as an accountable portfolio.
Bottom line
AI’s individual productivity gains justify serious investment, but the EBIT results show that value conversion remains incomplete.
Our stance is clear: scale AI responsibly by connecting experimentation to a limited set of owned, measurable workflows. Redesign processes where necessary, embed governance in delivery and operation, and build reusable foundations before duplication becomes expensive.
What to watch is whether organizations move beyond usage metrics and report stronger business outcomes. The decisive shift is from personal assistance to accountable enterprise performance.
Key takeaways
- Establish a baseline and measure actual outcomes; usage and perceived time savings do not prove business value.
- Give each strategic pilot a business sponsor, measurable objective, and shared technology-and-business accountability.
- Treat governance as part of workflow design, including decision rights, controls, monitoring, escalation, and retirement.
- Reuse common platforms, services, and patterns to prevent tool sprawl as the AI portfolio grows.
- Test whether AI improves the workflow itself rather than simply accelerating a broken process.
Frequently asked questions
What does scaling AI responsibly mean in practice?
It means connecting AI initiatives to valuable workflows, measurable results, clear ownership, reusable capabilities, and controls that operate throughout delivery and use.
Why isn’t improved employee productivity translating into EBIT gains?
Personal time savings may not change the wider workflow, its decision structure, or its cost of delivery. Stronger ownership, measurement, process design, and reuse can help close that gap.
Should every team experiment with AI?
Experimentation can generate valuable learning, but strategic pilots should concentrate resources on use cases with meaningful business potential, lower risk, adequate data, and stable processes.
Where should governance sit in an AI initiative?
Governance should shape requirements, testing, decision rights, monitoring, escalation, and retirement rather than appear only as a final approval checkpoint.
How can project teams evaluate an AI pilot?
Establish a baseline, then compare actual results with measures such as time, cost, quality, workflow volume, errors, and reviewer workload.