Enterprise AI spending is expected to reach $2.7 trillion in 2026 and $3.6 trillion by the end of 2027, according to Gartner. But despite unprecedented spending, AI implementations have largely failed to deliver worthwhile returns for enterprise companies.
Eliassen's Q3 survey of U.S. technology leaders uncovered that, when it comes to AI, enterprise companies:
These findings pose a very real threat to technology leaders: As C-suites begin to scrutinize AI spending more closely, being able to demonstrate real ROI is becoming increasingly urgent. Not only will tech leaders struggle to get the same AI budgets they’ve received in the past, but they’ll also likely be asked some difficult questions about where all that money went — and what the business got in return.
While AI ROI may be a problem for technology leaders, it isn’t actually a technology problem — and it doesn’t have a technology-based solution (at least, not yet).
To understand how enterprise tech leaders can begin to demonstrate ROI on their AI implementations, it’s important to understand why they’re so far behind other market segments.
In "The State of AI in Business 2025: The GenAI Divide," MIT's Project NANDA found that while enterprise organizations lead in pilot volume, those pilots often fail to scale. This failure, the report concluded, is largely because:
Smaller companies, meanwhile, generally run leaner, more unified tech stacks, so there's less legacy integration debt standing between a pilot and a production workflow.
MIT’s Project NANDA also found that large enterprises take an average of nine months to scale AI pilots to full implementation, compared to just 90 days for mid-market firms. Every extra month of governance review, security sign-off, and stakeholder alignment is a month where the investment isn't yet generating a return.
Note, however, that this isn’t a bug, but a feature, and it’s not something that inherently needs to be solved. In fact, the many layers of approval and governance checks necessary in these cases are good things — as long as they’re the right governance checks and approvals. More on that front later.
McKinsey's State of AI in 2026 survey found 80% of individual workers report productivity gains from AI. However, only 37% of organizations reported any EBIT gains from AI — and just 6% of companies attributed significant earnings impact to it.
Despite monumental investments in AI among enterprise organizations, this pattern has stayed flat year over year. Like the lag in the enterprise pilot-to-production pipeline, this may be more of an issue of scale than anything else: A few hours saved per employee per week is a rounding error against a multibillion-dollar revenue base, but the same relative gains may have a much more visible effect on a smaller company's books.
MIT’s NANDA research found that enterprise AI budgets skew heavily toward “visible,” front-office functions, like sales and marketing — think GenAI tools like image and content generation tools — even though functions like operations and finance tend to show better ROI.
Smaller organizations, on the other hand, have fewer dollars to spread around and often have no choice but to concentrate their AI spending on the one or two use cases that can impact the bottom line.
With so many factors seemingly stacked against them, how can large enterprises start seeing real ROI from their already substantial AI investments?
The good news is that they don’t need to spend even more on AI. They just need to know what they’re spending — and where.
By now, most have heard usage-cost horror stories like Uber’s, in which the company blew through its entire annual token budget in just four months. This can only happen when there’s no reporting in place to monitor spending and no guardrails to govern it.
To avoid learning the same costly lesson as Uber and countless others, put guardrails and robust consumption reporting in place before attempting to scale any AI solution. These include:
Instead of defaulting every task to the most capable (and expensive) model, route routine work — summarization, classification, transcription — to smaller, cheaper models and reserve frontier models for complex reasoning or agentic workflows. Given the massive price spreads between the cheapest and most expensive models, this tactic alone can cut costs substantially without a noticeable hit to output quality.
Rather than relying on reactive spending alerts, mature AI programs enforce budgets in real time, often by blocking or rerouting a request before it exceeds a given budget. These can and should be applied at the team, application, environment, individual user, and model level.
Unified dashboards can combine usage costs across providers, with automated alerts firing when spend crosses or approaches a defined threshold. Requests that reach or exceed a given budget can trigger approval workflows, ensuring that leadership has visibility into and control over any excess spending.
Functions like sales and marketing are low-hanging fruit for AI tools, since their capabilities can easily translate to time savings in those areas. But the real, needle-moving impact of AI comes from reimagining workflows, processes, and even entire functions around the benefits AI can offer.
Eliassen’s own research found that:
For enterprise organizations, bridging the gap between easy wins and AI that’s truly integrated into workflows and processes often starts with a structured framework that identifies use cases where AI can deliver meaningful impact, implements a formal AI adoption strategy, and even guides AI operating model design. This kind of framework can help formalize AI use, reduce or eliminate unsanctioned — and often costly — AI licenses, and focus AI spending on high-utility use cases.
To understand the ROI of any AI implementation, organizations must know what they’re spending across the enterprise. That includes licenses, infrastructure, and any FTEs responsible for those implementations, as well as ongoing consumption costs.
Remarkably, large enterprises fall short in this area. Almost half of respondents from $1B+ organizations told Eliassen that they don’t track spending at the consumption level — notably fewer than respondents from smaller organizations.
This is due, in part, to the fact that AI consumption pricing is complex, varies from one solution provider to the next, and seems to change from month to month. For these reasons, traditional financial teams aren’t equipped to monitor these costs, and they’re even less equipped to attribute costs like token usage to teams and use cases.
Creating an AI FinOps discipline empowered with the right tools and the necessary authority to monitor, approve, and attribute usage costs is a vital first step to realizing real ROI from AI investments.
Demonstrating ROI on AI spend is becoming increasingly critical, and tech leaders’ budgets hang in the balance. But doing so doesn’t require additional devs, new solutions, or still more spending. Instead, it requires decidedly non-tech solutions, like:
These may not be technology solutions, but that doesn’t mean they’ll be easy or quick to implement. They will require buy-in and assistance from the C-suite and other functional leaders, and they may slow down the pilot-to-production pipeline even further. But when implemented correctly, these solutions can give tech leaders control over their AI spending, visibility into which solutions and use cases are demonstrating the best returns, and a much more defensible position for those upcoming budget meetings.
Get a risk-adjusted ROI forecast for your AI implementations with our AI Value Navigator today, and visit our resources page today to discover even more real-world insights on AI strategy, AI implementations, and more.