In PwC’s 2026 CEO survey, 56% of respondents reported neither higher revenue nor lower costs from AI over the previous year. That is a gap in financial outcomes, not a count of failed projects.
My proposal is to give people the ability to lead improvements within a shared, governed AI ecosystem. Two common implementation patterns can leave one of those requirements missing.
Two ways to lose alignment
Delivery without employee ownership. Technical teams receive a backlog; business teams receive the solution. People who know the work have too little influence over priorities, design or success measures. Small improvements can stay outside the backlog, and changes imposed on employees can create resistance.
RAND’s interviews with experienced AI practitioners identified misunderstood business problems and a preference for new technology over useful solutions among recurring failure causes. The issue is the delivery arrangement: technical expertise remains essential.
Initiative without shared foundations. Employees build agents, apps and automations using different tools and assumptions. Without common priorities, connected data and clear ownership, local progress can produce duplication and incompatible solutions.
The 2024 Microsoft/LinkedIn Work Trend Index found that 78% of AI users brought their own tools to work. This shows demand extending beyond employer-provided tools, rather than proving unauthorized use.
Direction and participation matter
BCG’s 2026 AI at Work survey found a larger difference in reported business impact between strategy groups than between tool-access groups. With limited strategic clarity, 55% of regular users with limited tool access and 60% with strong access reported measurable impact. With strong clarity, those figures were 80% and 83%.
Sourced · BCG AI at Work, 2026, slide 22. Regular AI users, n = 9,923. Survey association.
Strategic clarity here includes an AI priority, a clear strategy and guidance on using time saved. The outcome is a respondent’s assessment of improvement in business metrics. The comparison does not establish causation or measure centralized versus decentralized delivery.
Participation adds another part of the picture. In OECD’s worker survey, AI users in firms consulting workers or their representatives about new technologies were 18 percentage points more likely in finance and 19 points more likely in manufacturing to report that AI improved their own performance. BCG also associated involving employees in AI ideas with better job enjoyment and business impact. These are associations with perceived outcomes, not causal tests of employee leadership.
A shared system, distributed leadership
People-Led Enterprise AI combines shared company direction with leadership by the people closest to the work:
| Who | What they own |
|---|---|
| Business leaders | Shared outcomes, investment priorities and value measures. |
| Technical and platform teams | Connected tools, data access, controls and support for complex solutions. |
| Employees and their teams | Problems to solve, suitable improvements, checks and lessons to share. |
Self-serve capabilities should let employees create skills, agents and apps. Shared identity, approved connections and reusable components keep those solutions within a governed ecosystem. That ecosystem can contain several tools or vendors.
McKinsey’s research describes a similar division: risk and data governance often sit centrally, while talent and adoption commonly use hybrid arrangements. Shared control can support local ownership without putting every improvement into one delivery queue.
The model in practice
HEINEKEN reports more than 7,500 makers building apps and automations, supported by a five-person platform product team. Employees start with personal solutions; broader use brings stronger production requirements. AI agents follow the same governance approach. This company-reported case, published by Microsoft, illustrates shared rules and reusable capabilities supporting local building.
Start with a workflow people want to improve. Agree who owns the outcome, give the team suitable tools, and compare quality, effort and cost before and after. Escalate complexity when needed, then share what works. The same evidence can improve local solutions and guide company priorities.
Sources and study limits
- PwC, 2026: 4,454 CEOs in 95 countries and territories; fieldwork September–November 2025. Reported revenue and cost changes are not audited ROI or a project-failure rate.
- BCG, 2026: 11,749 white-collar respondents in 14 markets; the chart covers 9,923 regular AI users. Group sizes and exact survey wording are not published in the slide. These are self-reported associations.
- OECD, 2023: surveys conducted in 2022 in finance and manufacturing across seven countries. Consultation covers new technologies generally. The performance comparison includes 1,131 finance and 842 manufacturing AI users and does not establish causation.
- Microsoft/LinkedIn, 2024: 31,000 knowledge workers in 31 markets. Bringing a tool to work does not itself demonstrate a policy breach or failed transformation.
- RAND, 2024: 65 practitioner interviews identify recurring mechanisms. They do not measure a universal AI project-failure rate.
Prepared with AI assistance. Sources and methods are documented; the editorial conclusions are the author's.
Keep reading
Faster agents do not shorten a review queue
If changes arrive faster than your reviewers clear them, the queue grows no matter how fast the agent wrote them. Changing who reviews what shortens it.
The 4D Framework, drawn on one migration ticket
Delegation, Description, Discernment and Diligence, one at a time, on a single database migration that goes from "should an agent touch this?" to a signed change.