KPMG Finds Enterprise AI Adoption Still Outrunning Measurable Impact

KPMG’s latest enterprise AI findings reached Google News with a stark conflict: 95 percent of surveyed organizations have an AI strategy, but only 8 percent report established returns.

That difference matters more than another announcement about faster models or larger infrastructure investments. Companies have acquired AI tools, launched pilots, and encouraged employees to experiment. Most have not rebuilt the workflows, accountability systems, or measurement practices needed to turn that activity into enterprise performance.

KPMG’s Q1 2026 Global AI Pulse places 54 percent of organizations in the early stages of research, experimentation, or strategic planning. Another 39 percent are scaling AI or driving adoption. Only a small group has crossed from broad activity into measurable, repeatable returns.

The resulting divide is no longer between organizations that use AI and those that do not. It separates companies treating AI as accessible software from those operating it as an integrated business capability.

That is the real story behind the latest Google News headline. Adoption keeps spreading, but impact remains concentrated among organizations that redesign work around AI instead of placing AI beside existing work.

Google News Highlights an AI Adoption Gap, Not an Access Gap

Enterprise AI has become common enough that basic adoption no longer proves maturity.

KPMG based its Q1 findings on a survey of 2,110 senior executives across 20 countries, territories, and jurisdictions. The respondents represented eight sectors and included C-suite leaders and their direct reports at large organizations.

The headline numbers appear contradictory at first. According to the Global AI Pulse, 95 percent of organizations have an AI strategy. Sixty-four percent report meaningful business value, yet only 8 percent say they have established ROI.

Those answers measure different levels of progress. A team can save time on document review or software development without producing a result visible in enterprise financial statements. It can also report useful outcomes before proving that those outcomes exceed deployment, integration, oversight, and computing costs.

KPMG’s maturity breakdown makes that distinction clearer. Eleven percent of organizations remain focused on research and development. Twenty-two percent are experimenting through proofs of concept or pilots, while 21 percent are building strategies, performance indicators, and data infrastructure.

Together, those groups account for the 54 percent still classified as early-stage adopters. Twenty-six percent are scaling AI across the enterprise, and 13 percent are driving broader organizational adoption. The remaining 8 percent report established returns tied to meaningful business outcomes.

This is not evidence that 92 percent of AI programs have failed. The survey does not make that claim. It shows that most organizations have not yet reached the point where returns are both established and measurable.

A separate McKinsey survey points in the same direction. It found that 88 percent of respondents reported regular AI use in at least one business function during 2025. However, only about one-third said their organizations had begun scaling AI programs.

The gap exists because using AI in one function is a low threshold. An approved chatbot, a coding assistant, or an automated summarization tool can satisfy that definition. Enterprise impact requires connected changes across processes, data, incentives, controls, and decisions.

This distinction also explains why apparently conflicting surveys can all be accurate. One study may measure whether employees have access to AI. Another may ask whether a company has deployed agents. A third may require verified financial returns.

Google News readers therefore need to inspect the denominator and maturity definition behind each headline. “Using AI,” “scaling AI,” and “earning ROI from AI” describe different states, even when coverage treats them as interchangeable.

KPMG’s survey is useful because it places those states on one progression. The progression begins with research and pilots, continues through enterprise scaling, and ends with established returns. Most organizations remain somewhere in the middle.

Spending Is Rising Faster Than Measurable Returns

Companies are increasing AI commitments before they have settled how value should be measured.

KPMG says surveyed organizations planned an average of $186 million in AI investment over the following 12 months. The average was weighted to reflect company size and regional representation, so it should not be read as a universal corporate budget.

Regional plans also varied. Organizations in Asia-Pacific reported the highest average planned investment at $245 million. The Americas followed at $178 million, while Europe, the Middle East, and Africa averaged $157 million.

Those figures show commitment, but spending is an input. It says little about whether a company has selected valuable workflows, achieved consistent adoption, or captured enough benefit to cover operating costs.

AI expenses also behave differently from many traditional software expenses. Model usage can grow with tokens, requests, automated steps, and the number of agents involved. A successful pilot can therefore become more expensive when it reaches more workers or handles longer processes.

KPMG’s Q2 update made this problem more visible. The survey found that 42 percent of leaders had only partial visibility into AI spending. Thirty-three percent cited limited understanding of cost structures, including tokens, as a challenge when deploying agents.

That issue pressures chief financial officers, technology leaders, and business-unit owners at the same time. Finance teams want defensible returns. Technology teams need dependable infrastructure and security. Operating leaders must show that a deployment improves a real process rather than creating more activity.

The pressure becomes sharper when a company automates multi-step work. An AI agent is a system that can plan and execute several actions toward a goal. Unlike a single chatbot response, an agent may call tools, retrieve company data, revise its output, and pass work to another system.

Every additional step can create cost, latency, and another opportunity for failure. A process that looks impressive in a controlled demonstration may become difficult to supervise when thousands of employees use it with live data.

This is why established ROI remains rare even when reported business value is much more common. Business value can include faster completion, improved employee experience, or better access to information. ROI requires a defined financial comparison between benefits and the full cost of producing them.

The calculation should include more than model subscriptions. Integration work, data preparation, employee training, evaluation, security reviews, incident response, and human supervision all consume resources.

Some benefits also resist immediate financial measurement. An assistant may help an analyst find relevant context faster, but the company still needs to connect that saved time with capacity, revenue, quality, or avoided risk.

Organizations often stop at the first metric. They count licenses, active users, prompts, or generated documents. Those measures show activity, not business impact.

A stronger measurement chain begins with a specific workflow. It then tracks whether AI changes cycle time, error rates, conversion, service quality, or another operating result. Only after that connection is established can leaders make a credible financial claim.

KPMG’s later Q2 findings support this approach. Organizations with strong cost visibility were five times more likely to report established ROI, at 15 percent compared with 3 percent.

That relationship does not prove that dashboards create returns. More capable organizations may be better at both cost control and AI deployment. Still, the pattern reinforces a practical point: teams cannot manage an investment they cannot see.

The Real Divide Is Orchestration Versus Tool Deployment

The organizations pulling ahead are coordinating AI across workflows, data, governance, and human decisions.

KPMG identifies about 11 percent of respondents as AI leaders. Their advantage does not come solely from spending more or deploying more tools. It comes from treating AI as part of an operating model.

The report calls this orchestration. In this context, orchestration means coordinating models, agents, data, controls, employees, and business processes as one managed system.

That is a higher standard than offering employees a chatbot. A general assistant can help with drafting, summarization, or exploration while leaving the surrounding workflow unchanged. Orchestrated AI changes how work enters a process, how context follows it, who reviews decisions, and how outcomes return as feedback.

Consider a customer-support workflow. A basic deployment lets an agent draft a response. A more mature system retrieves the correct account history, applies policy, flags uncertainty, routes sensitive cases to a person, records the resolution, and measures whether service improved.

The second system requires much more than a capable model. It needs trusted data, system permissions, clear escalation rules, monitoring, and an accountable process owner.

The same pattern applies to product development. An isolated assistant can summarize interviews or generate requirements. An integrated workflow connects research evidence, customer feedback, product decisions, technical constraints, and later results.

That connection explains why a searchable AI knowledge base can matter more than another generic interface. The model needs relevant organizational context, while employees need a way to inspect and reuse the material behind its output.

McKinsey’s earlier research found that workflow redesign had one of the strongest relationships with reported earnings impact from generative AI. Yet fewer than one-third of respondents said their organizations followed most of the surveyed adoption and scaling practices.

Fewer than one in five reported tracking clearly defined performance indicators for generative AI solutions. That finding exposes the weakness behind many adoption claims. A company cannot learn which deployments work if it has not defined what success means.

Workflow redesign also creates political and organizational difficulty. It changes responsibilities, approval paths, staffing assumptions, and sometimes the boundaries between departments. Installing a tool is often easier than negotiating those changes.

Employees may use an assistant frequently while avoiding it for consequential decisions. Managers may encourage experimentation without changing output expectations. Security teams may approve one interface but block the data connections needed for deeper integration.

Those conditions produce visible usage without structural impact. Employees generate more drafts and summaries, but decisions still move through the same bottlenecks. The organization becomes busier without becoming materially faster.

KPMG’s leader group suggests another route. These organizations align AI with growth rather than focusing only on cost reduction. They also report stronger workforce readiness and greater confidence in measuring business impact.

That matters because cost reduction encourages a narrow implementation question: which existing tasks can a model perform? A growth-oriented question asks how AI changes the product, customer experience, or speed of learning.

The first approach can deliver useful efficiency. The second requires more coordination, but it also creates outcomes that competitors may find harder to copy.

Microsoft describes a similar destination through its “Frontier Firm” concept. Its Work Trend Index analyzed 31,000 workers across 31 countries, LinkedIn labor data, and Microsoft 365 activity.

Microsoft says these firms organize work around human-agent teams and intelligence available on demand. That framing reflects a vendor’s perspective, so its performance claims deserve careful scrutiny. Still, it captures the organizational direction highlighted by KPMG.

The competitive split is therefore not Microsoft versus Google, or one foundation model versus another. It is tool deployment versus operational orchestration.

Organizations on the first route accumulate licenses and pilots. Organizations on the second route select workflows, connect trusted context, assign ownership, measure outcomes, and revise the process.

Model choice still matters for quality, cost, security, and latency. However, changing models cannot repair missing accountability or an undefined business objective.

What AI Adoption Numbers Still Do Not Prove

Survey momentum is real, but self-reported maturity and value remain imperfect measures of operational performance.

KPMG’s research captures executive perceptions at large organizations. Those leaders can describe strategy, investment, and organizational direction, but they may not observe everyday use across every team.

The survey also covers organizations that differ by region, industry, regulation, and technical capacity. A bank deploying AI into controlled review processes faces different constraints from a software company adding an assistant to development tools.

Aggregated percentages can hide those differences. They can also conceal whether respondents use the same definition of “meaningful business value.”

One executive may interpret value as faster employee output. Another may require increased revenue or reduced expense. A third may count risk avoidance, even when the avoided loss cannot be observed directly.

Established ROI sounds more precise, but survey respondents can still apply different methods. Some may include implementation and supervision costs, while others compare benefits only with software fees.

KPMG’s numbers should therefore be treated as directional evidence about maturity, not audited proof of economic returns.

The same caution applies to cross-report comparisons. Stanford’s 2026 AI Index reports that organizational AI adoption reached 88 percent in 2025. It also notes that agent deployment remained in the single digits across nearly every business function.

Those facts are compatible because broad access arrived before deep operational use. They also show why a single adoption percentage cannot answer whether AI has changed the organization.

The AI Index points to stronger productivity gains in structured, measurable work. It cites studies showing gains in customer support, software development, and marketing output, while noting weaker results in tasks requiring deeper reasoning.

That pattern offers an important constraint. AI performs best where teams can define inputs, inspect outputs, and measure results. Open-ended strategic work has fewer reliable feedback signals.

A company can still use AI for complex work, but it needs stronger human review and a clearer way to detect mistakes. Without those safeguards, faster output may simply accelerate weak decisions.

Security and privacy add another source of friction. KPMG found that 75 percent of respondents expressed concern about AI-related risk and security. Only 20 percent of early-stage organizations reported confidence in managing AI risk, compared with 47 percent of AI leaders.

Those concerns are not an argument for stopping deployment. They show why access alone cannot produce maturity. An organization needs policies, technical controls, evaluation routines, and incident ownership before it can expand sensitive use cases.

Workforce readiness creates a similar constraint. KPMG found 68 percent of organizations expressed some confidence in their talent pipeline, but only 22 percent reported high confidence.

Training employees to write prompts will not close that gap by itself. Workers need to know when AI is appropriate, what evidence an output requires, which data may be used, and when a person must take control.

Leaders also need to decide how saved time will be converted into value. If an employee completes a task faster but receives more low-value work, the organization has improved local efficiency without changing performance.

This is where adoption programs often become disconnected from business design. Technology teams manage tools, learning teams manage training, finance teams request returns, and operating teams keep their existing targets. No single owner controls the full outcome.

The skeptical reading of the KPMG survey is therefore not that AI has produced no value. It is that the distance between reported value and established ROI remains too large for simple conclusions.

The optimistic reading is more specific. A small group has begun to connect deployment with measurable enterprise results, suggesting that value is possible when organizational conditions support it.

Both interpretations reject the same mistake: treating access, usage, and impact as synonyms.

Three Signals Will Show Whether the Gap Is Closing

The next phase will be decided by everyday workflow adoption, cost accountability, and measured operating results.

The first signal is movement along KPMG’s maturity curve. Its Q2 survey already showed that the share of organizations where AI forms part of everyday work rose to 22 percent from 13 percent in Q1.

That is meaningful progress because routine use requires more than a temporary pilot. However, the strongest confirmation would be a corresponding rise in established ROI, not only a shift from experimentation into adoption.

KPMG’s Q2 survey found established ROI remained concentrated at 7 percent. The sample and quarterly framing require careful comparison, but the message remains consistent. Everyday use is growing faster than verified returns.

If future surveys show broader adoption alongside rising ROI, the case for enterprise learning will strengthen. If adoption rises while ROI stays near current levels, the value gap will become harder to dismiss as a temporary delay.

The second signal is cost visibility. Leaders should watch whether organizations can attribute model, agent, infrastructure, integration, and human-review expenses to particular workflows.

Cost dashboards alone do not establish value. They become useful when paired with operating measures and accountable owners.

KPMG reported that 53 percent of organizations had AI cost-monitoring dashboards in Q2. Fifty-four percent included cost reviews in AI approval processes.

The stronger test is whether those controls change decisions. Teams should retire weak use cases, redesign expensive ones, and expand deployments that produce defensible results.

If companies continue scaling without understanding usage-based costs, agentic AI may widen the gap. Multi-agent systems can multiply calls, tokens, tools, and failure points inside a single task.

The third signal is workflow-level evidence. Broad employee surveys will remain useful, but decision-makers need measures tied to specific processes.

A support deployment should track resolution quality, cycle time, escalation rates, and customer outcomes. A development deployment should track delivery time, defects, review burden, and maintainability. A research deployment should measure accuracy, source quality, and decision speed.

These measurements should extend beyond short demonstrations. A workflow that succeeds for four weeks may weaken as data changes, employees adapt, or usage expands.

Organizations should also publish clearer distinctions between gross productivity and captured value. Saving ten minutes does not automatically create a financial return. The organization must redirect that capacity or improve an outcome customers value.

Google News will continue surfacing large adoption percentages because they offer simple headlines. The more important evidence will appear in less dramatic operating measures.

Watch the share of organizations reporting established ROI. Watch whether cost visibility changes deployment choices. Then watch whether individual workflows produce durable gains after supervision, integration, and risk costs are included.

For enterprise buyers and knowledge workers, the practical question is no longer whether AI deserves experimentation. The evidence shows that experimentation is already widespread.

The question is whether your organization can name one important workflow, define its baseline, connect the right context, assign an accountable owner, and measure the result over time.

If it cannot, another model or agent will not close the gap. If it can, AI adoption has a path toward impact that survives beyond the next Google News headline.