NEW YORK — Corporate executives in the U.S. and Europe are hitting pause on broad generative AI rollouts as return on investment remains stubbornly hard to prove, Dec. 17, 2025.
That doesn’t mean spending is stopping. It means the “ship it everywhere” phase is ending, replaced by a tighter enterprise playbook: fewer pilots, clearer success metrics, stronger data controls, and deployments that live inside core workflows rather than in standalone chatbots.
Why generative AI ROI is stalling
The first wave of deployments was built on the idea that large language models would act like an “easy button” for knowledge work. But real-world use has exposed a messy middle: models can be dazzling in narrow tasks and unreliable in everyday business processes, forcing leaders to slow down and reset expectations.
In a Reuters report, Forrester found just 15% of executives surveyed in the second quarter said profit margins improved due to AI over the last year, while BCG found only 5% of executives surveyed saw “widespread value” from AI. The same report describes projects shelved after models produced inconsistent outputs and struggled with long, technical documents — prompting one executive to say, “We all thought it’d be the easy button.”
Analysts have been warning that a chunk of pilots wouldn’t survive contact with production. Gartner has predicted at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value.
The gap shows up even in organizations that believe they are “on track.” In Bain & Company’s executive survey, leaders reported 80% of generative AI use cases met or exceeded expectations — but only 23% said they can tie initiatives directly to new revenue or lower costs. In other words: teams are busy, demos look better, but finance-grade proof is still rare.
The pause is tactical: from experiments to enterprise accountability
Boards and CFOs are applying the same scrutiny they applied to cloud migrations and digital transformations: measurable outcomes, defensible risk management and a credible path from “pilot” to “default.” The near-term focus is shifting toward:
Workflow integration: AI that lives inside ticketing, CRM, ERP, call-center platforms and developer tools — not a separate chat window no one trusts.
Data readiness: Clean inputs, permissioned access and “source of truth” systems that reduce hallucinations and audit headaches.
Governance by design: Logging, evaluation, red-teaming, privacy controls, model-risk policies and clear accountability for outputs.
Use-case discipline: Narrower scope, fewer stakeholders, tighter definitions of “done,” and a hard stop for projects that can’t clear ROI hurdles.
What changes in 2026: the enterprise becomes the center of gravity
As the hype around general-purpose chatbots fades, the next phase is expected to look more like enterprise software evolution: embedded assistants, task-specific agents and domain tools built around internal data and business rules.
Gartner forecasts that 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. The message for buyers is blunt: the winning deployments won’t be “AI everywhere,” but “AI where it can complete an end-to-end task with guardrails.”
That shift also helps explain why the ROI conversation is getting more intense even as spending continues. According to Network World’s summary of Gartner spending estimates, global AI spending is expected to exceed $2 trillion in 2026, driven largely by AI embedded in products and the infrastructure underneath it. In other words, enterprises may buy more AI — but they will demand it behave more like enterprise tech: predictable, governable and measurable.
A practical 2026 enterprise playbook
Start where measurement is easiest: software engineering, IT service desks, knowledge management, and back-office document workflows.
Design for “human-in-the-loop” from day one: escalation paths, approval steps and clear boundaries for what agents can’t do.
Budget for operations, not just experiments: model monitoring, evaluation, security reviews, data curation and change management.
Build a vendor strategy: decide what you buy, what you build and what you partner on — and avoid “agentwashing” labels that don’t match real autonomy.
The long view: this cycle has played out before
Today’s disappointment looks less like failure and more like a familiar adoption curve: big promises, uneven pilots, then slower organizational change that ultimately determines value.
In a 2017 McKinsey Global Institute discussion paper, researchers pointed to a gap between AI investment and commercial application that is typical in early technology curves — a warning that scale would take time, not quarters.
Harvard Business Review’s 2019 guidance on building an AI-powered organization made a similar point in different terms: AI would reshape business, but “not at the blistering pace many assume,” because the hard part is operating model change — data, talent, processes and decision rights.
Then generative AI arrived and reset expectations upward. McKinsey’s 2023 estimate of generative AI’s economic potential put the upside in the trillions of dollars annually across dozens of use cases. Two years later, executives are learning the downside of that framing: the prize may be massive, but the path is slower, more expensive and more enterprise-specific than the early hype implied.
Bottom line
The painful reality of 2025 is that most organizations can’t yet prove generative AI is improving margins at scale. The decisive shift coming into 2026 is not retreat — it’s consolidation: fewer “cool” pilots, more enterprise deployments engineered for reliability, governance and measurable outcomes.
