The End of the Big Bet: Why 2026 AI Budgets are Going Small

Enterprises spent two years funding sweeping AI transformations. However, they are now breaking those programs into smaller deployments, asking for results sooner and cutting the ones that cannot show a clear business case.

In 2023 and 2024, the enterprise was still operating on the assumption that AI would change the way companies worked, and the budgets reflected that conviction. Boards approved multiyear transformation programs, CIOs signed platform-wide contracts and organizations spent ahead of evidence about which applications would deliver a return. AI was moving quickly enough that doing nothing began to look like the riskier decision, so executives were prepared to spend heavily on experiments rather than explain later why they had been the ones to sit out the biggest technology shift in a generation.

After two years of pilots and deployments, companies are no longer making those decisions in the dark; they have seen which projects made it into production, which stalled and which failed to produce a return. This experience is finally showing up in the way new projects are funded. Rather than committing years of spending on the assumption that the value will eventually emerge, executives are asking what a project can prove early, how much it will cost to find out and whether there is a sensible reason to keep funding it if the answer is disappointing.

The bill came due

Deloitte’s State of AI in the Enterprise 2026 report, based on a survey of 3,235 director-to-C-suite leaders across 24 countries and six industries, found that only about one-quarter of organizations had moved 40% or more of their AI pilots into production. Another 37% were using AI at what Deloitte calls a surface level, adding it to existing processes without fundamentally changing them. And 84% had not redesigned jobs around what the technology can actually do.

The numbers point to a gap between how enterprise AI was initially sold and how it has been deployed. Companies spent heavily on technology that was expected to change how work was done, but in most cases the work itself remained largely intact.

ALSO READ: How to Plan L&D Budget in the Age of AI?

MIT’s Project NANDA found a similar gap between experimentation and measurable business value. Its July 2025 report, The GenAI Divide: State of AI in Business 2025, drew on a review of more than 300 publicly disclosed AI initiatives, interviews with representatives from 52 organizations and survey responses from 153 senior leaders. Roughly 95% of the enterprise GenAI projects examined produced no measurable impact on the profit and loss statement, while only 5% of integrated AI pilots were generating millions of dollars in value.

The 95% figure has since been widely described as a failure rate, but the report’s own qualification is more precise. Its researchers say the figures are directional, based partly on interviews rather than company reporting, with sample sizes and definitions of success varying across categories. The finding is therefore better understood as a measure of projects that failed to demonstrate measurable financial impact than as a count of projects that companies formally abandoned.

The report also found a substantial difference in deployment rates depending on how companies built their systems. Among its sample, external partnerships with vendors reached deployment roughly 67% of the time, compared with about 33% for internal builds. The authors caution that the comparison does not establish causation and may reflect differences in organizational capability, procurement and risk tolerance. Still, the gap points to one of the more consequential choices companies are making as they move from AI experimentation toward production.

S&P Global Market Intelligence captured the problem from another direction. Its 2025 survey of more than 1,000 respondents across North America and Europe found that organizations scrapped an average of 46% of their AI proofs of concept before they reached production. The share of companies abandoning most of their AI initiatives rose from 17% to 42% in a year, while cost, data privacy and security risks emerged as the leading obstacles.

ALSO READ: Customer Experience After AI: A Field Report

RAND’s research offers a different view of what happens inside those failures. The organization interviewed 65 data scientists and engineers, each with at least five years of experience building AI and machine-learning systems. Among them, 84% identified leadership decisions as the primary cause of failure. The problems they described were often managerial rather than technical: teams were sent after the wrong problems, executives expected more from models than they could deliver, AI was applied where it was not necessary, and priorities changed before projects had time to mature.

There is a fairly mundane lesson in those findings. Companies made large commitments before they had enough experience to know which applications were likely to work, then gave those commitments enough money and organizational weight to make abandoning them difficult.

That is one reason the smaller project has become more attractive. It gives a company a chance to find out what works before it has built an entire strategy around the answer.

What a smaller bet looks like

Ecosystm’s 2026 analysis of enterprise AI describes organizations moving away from large, open-ended initiatives and toward small and medium-sized deployments tied to a tangible business outcome. The question on repeat, being asked, is what a project can achieve by the end of the quarter. Pilots that drift without a clear result are being cut, while the projects continuing to attract investment tend to address specific operational problems, including compliance reporting, customer and supply-chain processes and cyber threat intelligence.

There is an obvious financial advantage to working this way, but the more important benefit may be that it changes the consequences of being wrong. A company can test an application against a defined problem, measure the result and decide whether to expand it. If the experiment goes nowhere, the loss is limited to the money and time committed to that particular deployment rather than a multiyear transformation program whose success has become difficult to separate from the careers of the people who approved it.

The approach also allows companies to run several experiments at once and move money toward the ones that demonstrate value. That is a more familiar way of managing a technology portfolio than betting on a single sweeping transformation and hoping the benefits arrive on schedule.

The spending data suggests that this shift is not being driven by a retreat from AI.

KPMG’s AI Quarterly Pulse Survey, which polls 130 U.S. C-suite and business leaders at companies with at least $1 billion in annual revenue, found that 67% would maintain AI spending even if a recession arrived within 12 months. The survey projected $124 million in AI deployment over the coming year, while 59% expected measurable returns within that same period.

Those expectations are easier to reconcile with a portfolio of smaller deployments than with a single transformation program whose benefits may take years to appear. The shorter the project and the narrower the objective, the easier it is to ask a simple question: Did it work?

Where the money is actually coming from

The growth in AI budgets also looks different once you ask where the money is coming from.

A Sierra Ventures survey of more than 100 enterprise leaders published in July found that 62% of CXOs were increasing AI budgets by more than 10%, while 97% were increasing spending in some form. But only 41% of that spending represented genuinely new budget. Another 32% came from reprioritized IT spending, 15% from cuts to professional services and 12% from cuts to other software vendors. Only 3% came from reducing headcount.

ALSO READ: Why India’s Best AI Talent Keeps Building Abroad?

In other words, some of the money flowing into AI is money that used to belong to something else.

This is an important observation as a new AI project funded from an existing IT budget is not competing against nothing. It is competing against the software, consultants or projects that would otherwise have received that money. The investment case therefore has to do more than establish that AI is strategically important. It has to explain why this particular use of the budget is better than the alternatives.

Forrester put a number on the caution underneath this shift in its 2026 technology and security predictions, forecasting that enterprises would defer a quarter of planned AI spending into 2027 as financial scrutiny slowed production deployments and eliminated weaker proofs of concept. Fewer than one-third of decision-makers, Forrester found, could connect AI’s value to their organization’s financial growth. Without that connection, approval authority has increasingly moved toward the CFO.

The change says something about where AI now sits inside the enterprise. It is still a strategic technology, but it is increasingly being asked to compete on the same terms as every other significant investment.

Governance became the price of admission

There is another reason for the preference for smaller deployments, particularly as companies move from systems that assist employees to agents that can act on their own.

KPMG’s quarterly tracking shows how unsettled that market remains. Reported agent adoption rose from 11% of organizations in the first quarter of last year to 42% by the third, then registered 26% in the fourth. KPMG argues that the fourth-quarter reading reflects a redefinition of what counts as a genuine agent, with companies applying stricter standards and shifting their attention from launching new agents to governing those already in use.

Whatever accounts for the movement in the adoption numbers, the concerns surrounding these systems have become harder to treat as secondary. Cybersecurity now ranks as the single greatest barrier to achieving AI strategy goals, cited by 80% of leaders against 68% in the first quarter. Concern about data privacy climbed across the same span to 77% from 53%, and concern about data quality to 65% from 37%. Those are steep moves for three quarters, and they track the shift from AI that answers questions to AI that takes actions inside live systems.

Three-quarters of leaders now rank security, compliance and auditability as the most critical requirements for deploying agents, placing them ahead of speed. Seventy-two percent intend to deploy agents only from trusted technology providers, while 60% restrict agent access to sensitive data without human oversight. Nearly two-thirds cite the complexity of agentic systems as their leading barrier, a figure unchanged across two consecutive quarters, which points to a structural problem rather than a passing one.

The governance problem becomes harder as an AI system is given access to more data, more applications and more authority to act. A narrowly defined deployment can be monitored and constrained in ways that become more difficult once an agent is woven through multiple business processes.

Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027. The reasons it gives are escalating costs, unclear business value and inadequate risk controls.

None of those problems can be solved simply by choosing a more capable model. They are questions of scope, economics, ownership and governance, which makes the discipline around deployment at least as important as the technology itself.

What this means for the people approving the budget

For executives deciding which projects deserve funding, the research points to a handful of questions that are more useful than another discussion about AI’s potential.

Has the data been assessed for readiness, or merely assumed to be ready?

Gartner predicts organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026, having found that 63% of organizations either lack appropriate data-management practices for AI or cannot say whether they have them. RAND’s interviewees cited data limitations as the second most common cause of failure, behind leadership decisions.

The issue is not simply whether an organization has enough data. The relevant question is whether the data required for the particular application is accessible, reliable and usable.

Is the success metric defined before launch, or afterward?

A project should have a clear measure of success before the first dollar is spent. Otherwise, it becomes remarkably easy to keep redefining the objective as the results come in, particularly when a team has already invested months of work and executives have publicly backed the project.

Is the buy-versus-build decision deliberate?

MIT’s finding that vendor-supported tools succeeded roughly three times as often as systems built entirely in-house argues for treating an internal build as a decision that needs to be justified, rather than as the default expression of technical ambition. There will be cases where proprietary development makes sense, particularly in regulated industries, but the burden should be to explain why.

Does the project have an accountable owner rather than just a sponsor?

RAND’s central finding was that leadership decisions were identified as the primary cause of failure far more often than technical limitations. Someone therefore needs to be responsible not merely for getting the system built, but for making sure the problem it is intended to solve is worth solving in the first place.

The end of the big bet

The evidence does not point to an enterprise market turning away from AI. KPMG and Sierra Ventures both show companies committing more money than they did a year earlier.

What has changed is the way companies are learning what that money should buy.

The first phase of enterprise AI required a certain amount of faith. Companies had to experiment with models, vendors and use cases before there was much evidence about which combinations would work inside a particular business. Large commitments were, in part, the price of finding out.

Two years later, there is considerably more evidence to work with. Companies have seen which pilots made it into production, which ones stalled, where data problems stopped projects and how quickly costs can accumulate once AI systems move from demonstrations to everyday use.

That makes the old model of funding a broad transformation program and waiting for the promised benefits increasingly difficult to defend.

The next phase of enterprise AI will probably be less about persuading companies that the technology matters. Most large organizations have already made that decision. The harder question is which applications deserve to move from experiment to infrastructure, and which should be left behind.

That is a narrower question, but it is also a more consequential one. The companies that become good at answering it will have an advantage over those that simply keep adding AI to the budget. They will know not only where the technology can work, but where it is worth the cost, the organizational disruption and the risk of putting it into production.

For AI, that may be the end of the big bet and the beginning of something more ordinary: a technology that has to earn its place.

DISCLAIMER: This article has been produced by NervNow’s Editorial Team based on research and publicly available sources. While we have made every effort to ensure the accuracy of the information presented, discrepancies or errors may occur. If you identify any discrepancy in the research or data, please contact us at editorial@nervnow.com