HFS survey: only a third of AI-active enterprises hit their AI business case

Only 33% of AI-active enterprises say their most significant AI program has met or beaten its business case. That is the headline number from a survey of 502 senior AI decision makers that HFS Research conducted with Rackspace. The results were published November 12 in a report whose title tells enterprises to run their AI as one rather than keep paying for parts that never add up. Almost every respondent, 96%, also said they have significant governance gaps.

The more useful number for a CIO sits one layer down. Success shifts with deployment stage, and the way it shifts points at one specific problem: who runs AI after it goes live. HFS's reading is that companies have bought models, platforms, infrastructure and services, yet often nobody is responsible for running those pieces as one system.

The sample shapes how far the findings travel. Respondents worked at companies with at least $100 million in annual revenue, spread across eight industries and four regions, and were surveyed in August and September 2026. All were running AI pilots or more, and 83% were operating AI in production or beyond. The survey was blind, naming no vendor, and HFS says plainly that the results describe AI-active enterprises, not adoption across the whole market.

The payoff thins as AI spreads

Among enterprises using AI in a single function, 37% met or exceeded their business case. For those deploying across several functions, the figure drops to 28%. Where AI is embedded in standard operations, it climbs to 56%.

HFS adds its own caution: these are different groups of companies, and the data doesn't show that operating maturity causes better returns. As a pattern, though, it has a recognizable shape. Business-case success falls from 37% to 28% as AI moves from one function to several, which is where HFS says ownership of the whole system tends to go missing. The 56% group is made up of companies where AI has become part of how ordinary work gets done.

The kind of value being reported tells a similar story. Of respondents, 58% cite workforce productivity gains and 42% cite measurable cost reductions. Only 28% report better customer experience, and 21% report measurable revenue growth. AI is paying off first in efficiency. Revenue gains remain the minority result.

HFS's own case-study library points the same way from a different angle. Across more than 300 banking AI case studies the firm reviewed, productivity made up the biggest chunk of use cases at 44%, with nearly 20% built around risk and decision-making. A survey and a case library measure different things, but together they suggest enterprise AI portfolios are weighted toward efficiency before anything else is attempted.

HFS's banking analysts pressed on the same gap in a webinar the firm posted the same day, with HFS president Saurabh Gupta joined by practice leaders and senior advisors. Gupta noted that the US bank efficiency ratio was 56.9 in 2018 and is 55.6 today, barely moved despite heavy technology spending. He attributed the swings over the past 15 years to interest rates and the macroeconomic environment, not productivity, even as technology spending at global banks has grown close to 9% a year on average.

The panel described AI-created capacity that rarely reaches the P&L and a sector stuck in what Gupta called pilot purgatory. The session's description says only 13% of banks and capital markets firms have scaled AI across the business, even though AI spending is growing nearly 25% a year. The panel put a figure on what stands in the way: an estimated $1.5 trillion in enterprise debt across technology, data, process and talent, which HFS recommends tackling with a two-velocity approach that fixes foundations while capturing value. If a business case was built mostly on hours saved, the CFO's harder question is where those hours went.

Governance worries outrank technical ones

Respondents worry more about the controls than the code. For 78%, governance failures pose a bigger threat than technology failures to scaling AI. And 44% lack a single, consistent way to run AI in production. The release groups the barriers to AI returns into three: fragmented accountability, unpredictable costs and governance gaps.

Set beside the stage data, that 44% helps explain why multi-function rollouts fall behind. Each new function can bring its own model, platform and provider, along with its own way of tracking cost and behavior. HFS identifies the central obstacle as split accountability: no one clearly owns how AI performs after deployment, what it costs, or how it is controlled. Fragmentation is easy to tolerate in one function. Spread across five, it becomes an operating model.

HFS founder and CEO Phil Fersht did not soften his view in the release. He said the industry has built a large ecosystem selling models, platforms, infrastructure and consulting, and that when results fall short, each party blames another while the enterprise is left to make the pieces fit. Technology and talent are plentiful, he argued. What is scarce is people and partners willing to own how the whole system runs after go-live. In his words, the one-third result should make the industry face how hooked it has become on selling AI, as opposed to accepting responsibility for getting it to work.

For an IT leader, that turns governance into a concrete staffing question. When performance or cost goes the wrong way in a function that went live last quarter, which person or provider is expected to answer, and against what measure?

Three ways to fill the AI operator role

HFS proposes a fix it calls a Services-as-Software Operator. This could be an internal team, a set of existing providers brought under one remit, or a managed services partner. Whatever form it takes, it would hold day-to-day responsibility for how live AI performs, what it costs and how it is governed. The operator coordinates the technology and partners already in place. The enterprise keeps its business context, policy, decision rights and legal accountability.

In the release, Rackspace CEO Gajen Kandiah said that moving from pilots to production requires cost, security, governance and infrastructure readiness to be handled as one. He said customers need a single accountable operator that integrates, runs and keeps improving the environment, spanning infrastructure, applications, data and AI. He added that customers should keep control of their data and their technology choices, measure performance and cost clearly, and match the pace of deployments to business needs.

The finding fits how Rackspace presents itself: as the operator of the full enterprise AI stack, from governed private cloud to inference and agents in production, delivered through what it calls an Outcomes-as-a-Service model for regulated and mission-critical industries. In the company's own description, that model rests on secure infrastructure, data foundations and forward-deployed engineering, aimed at sectors "where governance, sovereignty, and uptime are non-negotiable." The boilerplate on the study release adds a Build | Operate | Innovate delivery model, more than 25 years of operating enterprise technology, and vendor-neutral, model-agnostic capabilities.

The survey also landed in a busy week for the company. Rackspace reported third-quarter 2026 results after the market close on November 9, with revenue of $609.1 million, down 9.25% from the same period a year earlier, according to Quiver Quantitative. The next day it announced an initial multiyear agreement with Together AI to support enterprise AI inference. Taken together, those announcements show the operator role is where Rackspace is placing its commercial bet, which makes the study's design safeguards worth checking.

That alignment deserves weight, and several things in the study help with it. Respondents never saw a vendor name. HFS states that it has no outside investors and that companies cannot pay to influence its research or ratings. And the report's own definition doesn't require outsourcing, since an internal team qualifies. HFS's banking panel went further on the day of the release, recommending that banks own their enterprise AI harness layer instead of renting it. For a company that already has a strong platform team, that points toward naming an internal operator before shopping for an external one.

What buyers say they would pay for

Appetite for a single owner is strongest among companies that have already felt the gap. The buying signals in the survey are unusually direct for a market that usually talks about AI in terms of features.

The report attaches a condition to the premium: providers have to show clear operational responsibilities and measurable outcomes to earn it. That changes what is being bought. A buyer paying above infrastructure cost is paying for accountability, so the contract has to define what accountability means in numbers.

The banking panel offered two tools for writing that definition. One is a tiered pricing model that moves beyond effort-based contracts. The other is a balanced business case built on four measures: productivity, prediction, personalization and performance. For procurement teams whose AI cases rest almost entirely on productivity, the second gives a way to tie fees to more than one kind of result.

The panel also devoted a segment to whether outcome-based pricing is still a pipe dream, and audience questions pressed for real examples of it. That suggests the contracting model the 73% say they would pay for is still being worked out, even among the analysts recommending it, so buyers may need to write the measures themselves.

The companies this reaches first are those about to take an AI program that worked in one function and extend it to a second or third. The drop from 37% to 28% happens at exactly that step. HFS's advice amounts to deciding who runs the combined system before that expansion goes live.