KPMG's Global AI Pulse shows a wide gap between AI strategy and AI ROI. The practical answer is not more pilots. It is redesigning how work gets done.
The uncomfortable number in the KPMG Global AI Pulse Q1 2026 and full PDF report is not that 95% of organizations have an AI strategy.
It is that only 8% report established return on investment.
That gap tells the real story. Most companies are not failing because they lack AI ambition. They are failing because they are putting AI on top of workflows, handoffs, dashboards, permissions, and management routines that were not designed for AI-enabled execution. KPMG's research surveyed 2,110 senior executives across 20 countries, territories, and jurisdictions, and found that organizations plan to invest an average of US$186 million in AI over the next 12 months. Yet only a small group, about 11%, is pulling ahead as AI leaders.
Here is the problem.
A pilot is not a strategy. Tool adoption is not business transformation. AI transformation only matters when it changes how work gets done.
Short answer
AI ROI shows up when AI changes the operating model around the work: workflow design, ownership, source-of-truth access, decision rights, review points, governance, adoption, and measurement cadence.
For enterprise AI transformation, the unit of change is the workflow, not the software license.
The hard part is not the model. It is the operating model around it.
Executive takeaways
1. Stop counting pilots. Count measurable business outcomes. 2. Start with the workflow. Then choose the AI system. 3. Make ownership explicit. If no one owns the workflow, no one owns the outcome. 4. Build governance into the design. Governance is not cleanup after launch. 5. Train managers, not only specialists. Adoption is a management rhythm. 6. Measure before and after. Baselines matter more than demo quality.
What KPMG's data really says
KPMG's central point is useful: enterprise AI has moved from adoption to orchestration. In the report, 39% of organizations are scaling AI or driving adoption across the enterprise, while 64% report meaningful business value. But only 8% report established ROI. That means activity is rising faster than value measurement.
The report also identifies three shifts separating AI leaders from the rest:
- Moving from pilots to AI-enabled operating models.
- Embedding governance as a prerequisite for scale.
- Distributing workforce capability across the enterprise.
That matches what I see in real operating environments.
The companies getting value are not just buying better tools. They are changing the way decisions move, the way handoffs work, the way data is accessed, and the way managers inspect performance.
The companies struggling are usually doing the opposite. They add copilots, agents, or automation into the same old operating model and then wonder why the dashboard does not move.
Why AI programs stall after the pilot
Pilots are forgiving. Real workflows are not.
In a pilot, the scope is small. The users are motivated. The data is curated. Exceptions are handled manually. The executive sponsor is paying attention. Everyone knows it is a test.
In production, the workflow has dependencies.
Sales qualification touches marketing data, CRM hygiene, account ownership, routing logic, manager review, compensation, and forecast inspection.
Customer onboarding touches sales handoff, customer success ownership, implementation capacity, product configuration, support readiness, and renewal risk.
Support triage touches ticket classification, entitlement data, knowledge quality, escalation rules, SLA commitments, and customer communication.
Forecasting touches pipeline definitions, stage discipline, rep behavior, manager inspection, finance assumptions, and executive decision-making.
AI can improve all of this. But not if it is dropped in as a separate layer.
If the process is unclear, AI accelerates confusion. If the data is bad, AI distributes bad assumptions faster. If decision rights are fuzzy, AI creates more debate. If governance is late, the team either blocks deployment or ships risk into the business.
That is why AI ROI is an operating model issue.
What an AI operating model means
An AI operating model is the set of rules, roles, workflows, systems, controls, and review cadences that determine how AI is used to produce business outcomes.
It answers practical questions:
- Who owns the workflow?
- What business outcome is being improved?
- What data can the AI system access?
- What system is the source of truth?
- What decisions can AI recommend?
- What decisions can AI execute?
- Where does a human review the output?
- What exceptions need escalation?
- What gets measured weekly?
- Who has authority to change the process?
Without those answers, AI becomes another disconnected tool.
The AI operating model checklist
Use this before approving another AI pilot:
1. Workflow
Name the workflow in plain language.
Bad: "Use AI in sales."
Better: "Reduce manual sales qualification time while improving routing accuracy for inbound enterprise leads."
The workflow should be specific enough that a manager can inspect it.
2. Outcome
Define the business metric before selecting the tool.
Examples:
- Reduce average support triage time.
- Increase lead-to-opportunity conversion.
- Improve forecast accuracy.
- Reduce onboarding delays.
- Increase renewal risk detection.
- Reduce finance close rework.
- Improve CRM data completeness.
License usage is not ROI. Prompt volume is not ROI. User activity is not ROI unless it connects to a business outcome.
3. Ownership
Assign one accountable workflow owner.
This is where many AI programs get weak. IT owns the tool. Data owns the pipeline. RevOps owns the CRM. Sales owns the number. Legal owns the risk. Nobody owns the workflow end to end.
That structure does not scale.
4. Data and permissions
Define the source of truth, access rules, and audit trail.
AI systems need data access, but not unlimited access. Leaders need to know which systems are being used, what data is retrieved, what is excluded, and how outputs can be traced.
KPMG identifies data privacy, cybersecurity, data quality, and regulatory uncertainty as major barriers to AI strategy execution. These are not side issues. They are execution issues.
5. Decision rights
Separate recommendations from actions.
For example:
- AI can summarize renewal risk, but the account owner decides the save plan.
- AI can draft a customer response, but support approves high-risk replies.
- AI can flag forecast risk, but the manager owns the inspection.
- AI can recommend lead routing, but RevOps controls the routing policy.
The higher the risk, the more explicit the human-in-the-loop design needs to be.
6. Governance
Governance should be built into the workflow, not added after deployment.
NIST's AI Risk Management Framework is a useful reference because it treats AI risk management as part of design, development, use, and evaluation. ISO/IEC 42001 also provides requirements for establishing and improving an AI management system.
For European organizations, this is not only good operating discipline. The EU AI Act has entered its main implementation window, with obligations applying in phases. That raises the cost of vague ownership, weak documentation, and uncontrolled AI deployment.
Governance is not the enemy of speed. Bad governance is. Good governance lets teams scale without guessing every time a new use case appears.
7. Cadence
Define the management rhythm.
This is the part executives underestimate.
AI adoption does not happen because a launch email went out. It happens because managers inspect the workflow, review exceptions, compare before-and-after metrics, coach behavior, and remove friction.
Every AI workflow needs a review cadence:
- What changed this week?
- Where did the AI system help?
- Where did it fail?
- Which exceptions repeated?
- Which users avoided the workflow?
- Which outputs needed human correction?
- What business metric moved?
If nobody reviews it, it will drift.
AI leaders orchestrate. Everyone else deploys.
KPMG's report draws a useful distinction between organizations that deploy AI and organizations that orchestrate AI.
Deployment means putting AI into use cases.
Orchestration means coordinating AI across workflows, systems, governance, people, and decisions.
That distinction matters more as agentic AI grows. Agents are not just content generators. They can route work, trigger actions, coordinate handoffs, retrieve data, and manage multi-step processes. That makes them more useful, but also more operationally sensitive.
An agent that drafts a proposal is one thing.
An agent that updates CRM, changes opportunity fields, triggers a discount approval, alerts finance, and creates a customer follow-up task is a different operating problem.
The second case requires ownership, permissions, exception handling, auditability, and manager inspection. Otherwise, the company has not automated a workflow. It has automated ambiguity.
The mistakes executives should avoid
Mistake 1: Treating AI as an IT rollout
IT matters. Security matters. Architecture matters.
But AI transformation is not only an IT program. It changes how work moves across the business. That means the executive owner needs authority over process, people, systems, and outcomes.
Mistake 2: Measuring activity instead of impact
A dashboard showing AI usage does not prove business value.
Executives need before-and-after metrics tied to cost, revenue, speed, quality, risk, or capacity. If the metric cannot be tied to a business decision, it is probably not the right metric.
Mistake 3: Centralizing all capability in a specialist team
A central AI team can build standards and reusable patterns. But execution has to move into the functions.
Managers need enough AI fluency to inspect work. Operators need enough fluency to redesign workflows. Technical teams need clear business outcomes, not vague AI ambition.
Mistake 4: Ignoring the handoffs
Most AI value gets trapped at handoffs.
Marketing to sales. Sales to customer success. Customer success to support. Support to product. RevOps to finance. Finance to leadership.
If AI improves one team's task but leaves the handoff broken, the enterprise outcome will not move much.
Mistake 5: Scaling before the workflow is stable
Do not scale a broken workflow just because AI makes it faster.
Fix the process. Clarify the owner. Clean the data. Define the review points. Then automate.
What leaders should do next
Start with three workflows, not 30 use cases.
Pick workflows with real economic impact and visible operating friction. Good candidates include:
- Sales qualification.
- Customer onboarding.
- Support triage.
- Renewal risk.
- Forecast inspection.
- Proposal generation.
- Finance close.
- CRM hygiene.
- Executive reporting.
For each workflow, complete a one-page operating model:
1. Business outcome. 2. Current baseline. 3. Workflow owner. 4. Systems involved. 5. Source of truth. 6. AI role. 7. Human review point. 8. Decision rights. 9. Risk controls. 10. Weekly review cadence. 11. Scale or stop criteria.
Then run the workflow for 30 to 60 days and inspect it like an operator.
Not a demo. Not a steering committee theater. A real operating review.
What moved? What broke? What did users avoid? What did managers trust? What should be automated further? What needs to stay human?
That is where AI transformation becomes real.
FAQ
Why are companies struggling to get AI ROI?
Most companies are adding AI to existing structures without changing workflows, ownership, data access, decision rights, governance, or measurement. That creates AI activity, but not necessarily measurable business value.
What is the difference between AI adoption and AI transformation?
AI adoption means people are using AI tools. AI transformation means AI changes how work gets done and improves a measurable business outcome.
What is an AI-enabled operating model?
An AI-enabled operating model defines how workflows, systems, people, governance, data, and decision rights work together so AI can produce consistent business outcomes.
How should executives measure AI ROI?
Measure AI ROI against business outcomes: revenue, margin, cycle time, quality, customer experience, risk reduction, capacity, forecast accuracy, or cost-to-serve. Do not rely on tool usage alone.
Where should an enterprise start with AI transformation?
Start with a workflow that has measurable friction and a clear owner. Map the process, set a baseline, define the AI role, build governance into the workflow, and inspect results weekly.
Is governance slowing down AI?
Poor governance slows AI down. Good governance makes scale possible by clarifying permissions, accountability, risk controls, auditability, and human review.
Final takeaway
The next phase of enterprise AI will not be won by the company with the longest list of pilots.
It will be won by the company that can redesign how work gets done.
AI ROI requires more than tools. It requires workflow ownership, trusted data, decision rights, governance, workforce capability, and operating cadence.
That is less glamorous than a demo.
It is also where the value is.
About Bryan Barrett
Bryan Barrett helps companies turn AI pilots into practical workflows, stronger accountability, better handoffs, and measurable execution. Learn more on the About page or get in touch.