Enterprise Agentic AI August 2026: McKinsey’s ROI Gap, AT&T’s Full Deployment, and Gartner’s Sobering Warning

August 2026 produced the most comprehensive picture of enterprise agentic AI adoption to date — and it revealed a fundamental contradiction that every organization deploying AI agents needs to understand.

On one side: 80 percent of individual workers using AI tools report improved personal productivity. Gartner is tracking what could be the largest single-year expansion in the history of enterprise software, with 40 percent of enterprise applications predicted to embed task-specific AI agents by year end, up from under 5 percent in 2025. Global agentic AI spending is projected to reach $201.9 billion in 2026, a 141 percent increase over 2025. The AT&T deployment shows what a fully committed enterprise AI program looks like in production across five distinct business functions simultaneously.

On the other side: only 6 percent of organizations qualify as “AI high performers” in McKinsey’s definition, a figure that has not moved in a year despite record spending. Only 37 percent of organizations report any measurable EBIT impact from AI at all — flat versus 2025. KPMG found that the average enterprise AI budget has reached $186 million, yet only 8 percent of those companies report tangible return on investment. Gartner is predicting that more than 40 percent of agentic AI projects will be canceled before the end of 2027.

Both sides of this picture are accurate. Understanding what separates the enterprises capturing real returns from those accumulating costs without measurable earnings impact is the central practical question for enterprise AI leadership in the second half of 2026.

McKinsey August 25 Report: The Full Data Picture

McKinsey published “The State of AI in 2026: On the Road to ROI” on August 25, 2026. The report was based on an online survey of 1,719 respondents across 97 nations, with data collected between May 4 and June 8, 2026. The methodology weighted respondents by each nation’s contribution to global GDP, making it one of the most geographically representative AI adoption surveys conducted to date.

The headline numbers established several important benchmarks. Eighty-nine percent of organizations regularly use AI tools in at least one business function. Forty-four percent have already scaled AI enterprise-wide. Eighty percent of individual AI users report improved personal productivity — a striking and consistent finding across geographies and industries.

The EBIT data told a different story. Only 37 percent of respondents attributed any measurable positive impact on earnings before interest and taxes to AI use, a figure unchanged from McKinsey’s 2025 survey despite significant increases in both adoption and spending. Only 6 percent of organizations qualified as what McKinsey defined as “AI high performers,” meaning organizations attributing 5 percent or more of organizational EBIT to AI with significant impact. That figure was also flat year over year.

This gap between individual productivity gains and organizational earnings impact is the central finding of the 2026 report. McKinsey’s analysis identified the cause: individual productivity gains do not automatically aggregate into enterprise financial benefit without deliberate workflow redesign. When AI tools make individual workers faster at tasks that were already being performed, the organization captures the efficiency gain. When those tasks are not connected to measurable business outputs through redesigned workflows, the efficiency gain remains invisible on the income statement.

The implication is specific: organizations that have deployed AI broadly but not redesigned the workflows those tools operate within are unlikely to close the gap between productivity gains and earnings impact simply by adding more AI. The lever is workflow architecture, not more software.

The Build vs. Buy Shift: 32 Percent Chose Not to Purchase Software

One of the most practically significant findings in McKinsey’s August 2026 survey was the build versus buy shift. Nearly one third of all respondents said their organizations had opted to build at least one software product or feature in-house using agentic coding tools rather than purchasing it from a vendor.

This represents a meaningful change in enterprise software economics. The traditional model — buy specialized SaaS for each function, pay ongoing subscription fees, and accept the feature roadmap the vendor delivers — is being challenged by agentic coding tools that make in-house development faster and cheaper than it previously was.

McKinsey’s Michael Chui described the dynamic directly: agentic coding tools are creating a more capable option for bringing some software development in-house. The practical effect is that enterprises with competent engineering teams are increasingly asking whether a needed capability should be purchased or built. For many narrow, function-specific features, the build option is now economically competitive in a way it was not before agentic coding tools existed.

For the SaaS industry, this shift represents a structural headwind. Not a catastrophic one — most enterprise software provides value far beyond what organizations could replicate in-house with current tools. But for point solutions addressing narrow, well-defined functions, the competitive position of incumbent vendors is weaker than it was two years ago.

For enterprise technology buyers, the implication is an additional evaluation step. Before purchasing a software feature or point solution, the question now has three answers rather than two: buy from an existing vendor, evaluate alternative vendors, or build in-house using agentic coding tools. The third option warrants an honest assessment in every procurement decision where the required capability is specific and well-defined.

AT&T and H2O AI: What a Full Enterprise Agentic Deployment Looks Like

AT&T’s July 2026 deployment of the H2O AI Super Agent provides the most detailed public example available of what a committed, multi-function enterprise agentic AI deployment looks like in production at scale. The deployment was announced on July 14, 2026, and represents AT&T’s most advanced implementation of agentic AI to date across its Ask AT&T platform.

The H2O AI Super Agent functions as an orchestrator rather than a single executor. It decomposes high-level objectives into structured sub-tasks, deploys specialized agents in parallel to address different components simultaneously, and dynamically adapts workflows based on context as execution proceeds. This architecture allows the system to handle complex, multi-step enterprise tasks that exceed the capability of any single-purpose agent.

AT&T is using the system across five distinct business functions. In fraud prevention and financial crime detection, the H2O AI Super Agent applies predictive and agentic intelligence to identify anomalies and reduce risk. AT&T has a documented history of severe iPhone fraud exposure, with the problem representing more than $1 billion annually in the US market before its earlier AI-based fraud systems reduced fraudulent activity by more than 80 percent. The agentic deployment extends those capabilities with greater autonomy and real-time adaptation.

In customer experience and field operations, fine-tuned language models improve resolution quality and response speed across AT&T’s large customer base. In enterprise research, document intelligence, and content generation, agent-based retrieval-augmented generation enables what H2O describes as the ability to “talk to documents” — retrieving, synthesizing, and acting on information from large document sets without manual review. In network and tower data analysis, agents provide structured data visualization and operational intelligence from AT&T’s infrastructure. The Ask AT&T platform integrates all of these capabilities for internal users across the organization.

The H2O AI Super Agent’s architecture includes embedded risk controls, audit trails, and governance mechanisms that meet enterprise compliance requirements from the initial design. For regulated environments, H2O supports sovereign AI deployments in air-gapped and on-premises configurations, meaning AT&T can maintain full control over data, models, and execution within its own infrastructure. This matters significantly given AT&T’s regulatory environment across telecommunications, financial services compliance requirements, and the EU AI Act obligations that apply to its European operations.

The AT&T deployment is instructive for enterprise leaders evaluating agentic AI strategy not because every organization can replicate AT&T’s scale or technical sophistication, but because it illustrates the architecture decisions and governance infrastructure that production-grade multi-function agentic AI requires. Organizations that deploy agents function by function without a unified orchestration architecture and without embedded governance will encounter coordination problems and compliance exposure as their deployments expand.

Gartner’s Market Projections and the Cancellation Warning

Gartner’s research on enterprise agentic AI in 2026 presented two findings that should be read together: the market expansion is real and the failure rate will be significant.

On expansion: Gartner forecasts that 40 percent of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. The global AI agents market is projected to reach between $10.9 billion and $12.1 billion in 2026, growing at a compound annual rate of 44 to 46 percent through 2030. Gartner’s long-range projection suggests agentic AI could drive approximately 30 percent of enterprise application software revenue by 2035, potentially surpassing $450 billion. Gartner also projects that by 2027, one third of agentic AI implementations will combine agents with different skills to manage complex tasks across application and data environments.

On failure: Gartner also predicts that more than 40 percent of agentic AI projects will be canceled before the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls. Only 21 percent of organizations currently have a mature AI governance model. Gartner’s analysis identified “agent washing” — the rebranding of existing products such as chatbots, AI assistants, and robotic process automation tools as agentic AI without substantive capability changes — as a significant contributor to failed implementations. Organizations that deploy rebranded automation under the agentic AI label and expect agent-level results will encounter the gap between expectation and reality at the point of business case review.

Gartner’s Anushree Verma framed the risk directly: most agentic AI projects are early-stage experiments or proof of concepts driven primarily by hype and are often misapplied. The real cost and complexity of deploying AI agents at scale is not visible in proof-of-concept phases, and organizations that move from pilot to enterprise rollout without validating the business case against realistic production costs will face the cancellation decision.

The Deployment Gap: 62 Percent Experiment, Under 25 Percent Scale

McKinsey’s data on the deployment gap is consistent with Gartner’s cancellation prediction. While 62 percent of organizations are experimenting with AI agents, fewer than 25 percent have successfully scaled agent deployments to production. The gap between experimentation and production scaling represents the most significant operational challenge in enterprise agentic AI as of mid-2026.

S&P Global Market Intelligence and McKinsey data from mid-2026 estimated that approximately 31 percent of enterprises currently run at least one AI agent in production, with banking and insurance leading at roughly 47 percent adoption. The median time-to-value on agent deployments is approximately 5.1 months according to BCG and Forrester research, meaning organizations that initiated deployments in early 2026 are reaching the evaluation point where ROI must be demonstrable or the project faces resource reallocation.

The sectors showing the highest production deployment rates share several characteristics: well-defined, high-volume processes where agent performance can be measured against clear success criteria; existing data infrastructure that agents can access without requiring significant pre-deployment data engineering; and organizational tolerance for the governance and oversight investment that production agentic AI requires. Organizations in sectors that lack these characteristics — complex, judgment-intensive processes with limited structured data and weak governance infrastructure — are producing the deployments most likely to be among Gartner’s predicted cancellations.

The Cost Constraint: $186 Million Budgets, 8 Percent Tangible ROI

KPMG’s first-quarter 2026 Global AI Pulse survey found that the average enterprise AI budget had reached $186 million, yet only 8 percent of those companies reported tangible return on investment. The survey provides context for interpreting McKinsey’s finding that 20 percent of organizations cite AI operating costs as a constraint on further AI use.

The cost structure of enterprise agentic AI differs from conventional software in ways that procurement and finance functions are still adapting to. Token costs for language model API calls accumulate at rates that depend on prompt design, conversation length, and usage volume — all of which can change rapidly with agent workflow modifications or traffic increases. The cost monitoring infrastructure required to manage these expenses is a separate investment from the agent applications themselves.

Forrester’s prediction that half of enterprise ERP vendors will launch autonomous governance modules combining explainable AI, automated audit trails, and real-time compliance monitoring in 2026 reflects the recognition that governance infrastructure is becoming a separable and necessary investment layer. Organizations that treat governance as an add-on to existing agent deployments rather than a foundational component of their architecture are accumulating both compliance exposure and operational risk.

What High Performers Do Differently

McKinsey’s data consistently showed that the 6 percent of organizations qualifying as AI high performers shared structural characteristics that distinguished them from the broader population of AI adopters. These characteristics were not primarily about technology selection. They were about deployment architecture, organizational alignment, and the connection between AI capability and business process redesign.

High performers deployed AI agents in workflows where the connection between agent output and measurable business outcomes was explicit and tracked. They did not simply add AI tools to existing processes. They redesigned processes to incorporate AI agents at decision points where agent judgment could be validated against outcomes and improved iteratively.

High performers also invested in governance infrastructure from the start of deployment rather than adding it reactively after incidents or compliance requirements forced their hand. This investment included audit trails, permission boundaries, human review workflows for high-impact decisions, and clear escalation paths when agents encountered situations outside their defined competence. The EU AI Act’s August 2026 enforcement activation created external pressure toward this model, but high performers had already built it for internal quality control reasons before regulatory requirements made it mandatory.

Finally, high performers maintained strict discipline about use case selection. They deployed agents where the task was well-defined, the success criteria were measurable, and the volume was sufficient to justify the deployment investment. They avoided deploying agents in complex, judgment-intensive areas without robust human oversight architectures, not because they lacked ambition, but because they understood that agent reliability in complex domains required a maturation process that short pilot timelines did not provide.

Conclusion: August 2026 Is the Accountability Checkpoint

August 2026 is the point in the enterprise agentic AI adoption cycle where accountability replaces enthusiasm as the primary driver of deployment decisions. The McKinsey data, the Gartner predictions, and the AT&T case study collectively describe a market that is growing rapidly but distributing its returns very unevenly.

The organizations likely to be among Gartner’s predicted cancellations are those that deployed broadly under the assumption that AI capability would automatically translate into business results, without the workflow redesign, governance infrastructure, and disciplined use case selection that turns agent capability into earnings impact.

The organizations likely to join McKinsey’s 6 percent of high performers are those that treat agentic AI deployment as an organizational design challenge as much as a technology challenge — building the measurement frameworks, governance architecture, and process redesign capability that allow agent productivity to aggregate into financial outcomes that appear on income statements.

The technology is available. The question August 2026 is asking is whether the organizations deploying it are ready to use it in the way that generates the returns they are projecting.

Follow our site for continued enterprise agentic AI coverage, analysis, and deployment guidance.