The next wave of AI winners will not be the companies that buy the most AI. They will be the companies that deploy it best. This report explains why the bottleneck has moved, what the data now shows, and what it means for operators and investors.
Illustrative. Spend trajectory per reported hyperscaler capital expenditure. Returns per PwC Global CEO Survey, January 2026, n = 4,454 chief executives.
For three years, enterprise leaders asked a single question: can AI do this? That question has largely been answered. Frontier models now write code, automate workflows, analyze contracts, process documents, and increasingly act autonomously. The capability debate is over.
What has not been answered is the question that actually matters to a P&L: is any of this creating value? And here the data from the first half of 2026 is brutal. PwC surveyed 4,454 chief executives across 95 countries in January and found that 56 percent report neither increased revenue nor decreased costs from AI in the past twelve months. Only 12 percent report both. MIT's NANDA research found that 95 percent of generative AI pilots produced no measurable P&L impact. S&P Global found that 42 percent of companies abandoned most of their AI projects in 2025, more than double the rate a year earlier. IBM's global CEO study put the share of AI initiatives delivering expected ROI at one in four.
Meanwhile, the money keeps flowing. Hyperscalers are on track to spend roughly $675 billion on AI infrastructure in 2026, up 63 percent from the prior year. Adoption is nearly universal: McKinsey finds 88 percent of organizations now use AI in at least one business function. The inputs have never been larger. The outputs, for most companies, have never been harder to find.
This report makes five arguments. The bottleneck has moved from models to operating models. AI is a change management problem wearing a technology costume. Enterprise software is shifting from products to workflows. Private equity has an emerging AI multiple problem it has not yet priced. And the winners of the next decade will build AI operating systems, not AI projects.
We also make a sixth argument, one the market itself just validated in dramatic fashion: no company should expect to cross this gap alone. In May 2026, the two leading frontier AI labs each stood up multi-billion dollar deployment ventures with some of the largest investors in the world, a tacit admission that even the companies building the models believe deployment is a distinct discipline requiring dedicated partners. We explore what that means for middle market companies and the sponsors who own them.
In 2023, the constraint on enterprise AI was access to frontier capability. In 2026, that constraint has disappeared. What replaced it is worse.
Organizations today can choose from frontier closed models, open source foundation models, agent frameworks, workflow orchestration platforms, low code environments, vertical AI vendors, and AI native application builders. The problem is not a lack of tools. The problem is an abundance of tools with no coherent operating model.
The evidence for this shift is now overwhelming. McKinsey's latest State of AI research finds that fewer than 20 percent of AI pilots ever cross into enterprise scale production. The rest live on indefinitely as proofs of concept, consuming budget and producing slide decks. Gartner estimates that just one in fifty AI investments delivers transformational value, and projects that 30 percent of generative AI initiatives launched in 2024 will be abandoned by the end of this year. Gartner separately predicts that 60 percent of AI projects lacking AI ready data will be abandoned through 2026, a reminder that most of the work of deployment happens below the waterline: data engineering, integration, governance, and measurement.
The common enterprise journey now has a recognizable shape. Pilot one succeeds. Pilot two succeeds. Pilot seven succeeds. Enterprise impact: nothing measurable. Analysts have started calling it pilot purgatory, and it is the default state of enterprise AI in 2026. Roughly 80 percent of the work required to move from pilot to production is not model work at all. It is workflow integration, data plumbing, and the unglamorous instrumentation that makes value visible to a CFO.
Sources: McKinsey State of AI; BCG AI at Scale 2026, n = 1,800 executives; MIT NANDA. Bars scaled to reported percentages.
Notice what is absent from every failure statistic above: the models. Nobody abandons an AI project because the model cannot write an email or summarize a contract. Projects die because the workflow around the model was never redesigned, the data was never made ready, ownership was never assigned, and success was never defined. The technology works. The deployment does not.
The prevailing enterprise AI conversation is still technical. Which model, which vendor, which framework, which agent platform. In practice, none of those choices predicts success. Organizational adaptation does.
The freshest evidence comes from the front lines of adoption. A 2026 survey of 2,400 executives and employees by Writer and Workplace Intelligence found that 79 percent of organizations face challenges adopting AI, a double digit increase over last year, and that 54 percent of C suite executives say AI adoption is actively tearing their company apart. Read that again. The technology is universally deployed, 97 percent of executives say their company rolled out AI agents in the past year, and the dominant executive experience of it is internal conflict. Only 29 percent report significant ROI.
This is not what a technology problem looks like. It is what a change problem looks like. The organizations that do generate outsized returns behave differently in predictable ways. They redesign workflows rather than automate tasks. They embed AI into existing systems of work rather than adding new destinations. They assign a single accountable business owner to every initiative rather than a steering committee. They invest in adoption the way they would invest in a post merger integration. And they measure realized business value, not model performance. PwC's January data confirms the pattern: the CEOs who report financial returns are two to three times more likely to have embedded AI deeply into decision making and demand generation rather than distributing licenses and hoping.
Deployly point of view, informed by post deployment reviews and published research on enterprise AI failure modes. Most organizations invert the allocation that success requires.
We believe enterprise AI deployment effort distributes roughly as follows: 20 percent technology, 30 percent data, 25 percent workflow redesign, 25 percent change management. Most organizations spend as if the ratio were the reverse, concentrating budget on licenses and platforms while leaving workflow and adoption unfunded. Then they are surprised when usage dashboards look healthy and the P&L looks unchanged.
UC Berkeley researchers offered a useful reframe of MIT's 95 percent figure: it is not only an AI failure, it is a measurement failure. Companies apply traditional software payback logic to a technology that creates value only when the work itself changes. If you deploy AI without redesigning the work, you have purchased a faster way to do the wrong thing.
Perhaps the most underpriced implication of generative AI is that it changes the economics of enterprise software itself. Value is migrating out of the application and into the workflow.
The historical model was simple. Buy software, train employees, employees execute the process inside the vendor's interface. The emerging model inverts it. Keep the process, deploy AI against it, and let the AI execute inside your own systems of record. When bespoke capability becomes cheap to build and agents can operate across systems, the application layer stops being where advantage lives.
Buyers are already voting with their evaluation criteria. The Futurum Group's survey of 830 IT decision makers, released this spring, documents the shift in real time: direct P&L impact nearly doubled as the primary success metric for AI investments, while productivity gains collapsed as the leading justification. At the same time, autonomous agents surged 31.5 percent year over year as a top technology priority. Enterprises are no longer buying tools that help people work. They are buying systems that do the work, and they expect the results to show up in financial statements.
The source of enterprise value migrates from the application to the workflow itself.
The implications compound. Systems of record become more important, because they hold the data agents run on. User interfaces become less important, because fewer humans sit in them. Bespoke applications become cheap, eroding the premium for packaged features. Workflow orchestration becomes strategically critical. And traditional SaaS categories, particularly seat based tools whose value proposition is a nicer place for a human to do a task, become structurally vulnerable. Sponsors have noticed: PwC reports that dealmakers have grown measurably more cautious on software targets in 2026 precisely because AI clouds their revenue models and long term competitiveness.
Private equity runs on proof. Every thesis is tested before capital deploys, every portfolio company measured against numbers that decide whether to scale, restructure, or exit. With AI, that discipline has not yet arrived. The industry is building conviction faster than it is building evidence.
Consider the gap between what the industry says and what it can show. In EY's latest AI Pulse, 84 percent of PE firms report having appointed a chief AI officer, and two thirds expect to direct more than a quarter of their technology budgets to AI this year. In FTI Consulting's 2026 Private Equity AI Radar, 95 percent of funds say their AI initiatives are meeting or exceeding the original business case. Confidence, in other words, is nearly universal.
Now the other side of the ledger. Grant Thornton's 2026 AI Impact Survey, fielded this spring across 950 business leaders, found that private equity is more confident in its AI strategy than almost any sector surveyed, yet reports below average measurable returns. Just 5 percent of PE respondents say they are fully integrating AI across operations, against 40 percent of technology companies. Only 9 percent of PE leaders are very confident they could pass an AI governance audit within 90 days. Bain estimates that only about 20 percent of portfolio companies have moved a generative AI use case into production with concrete results. And the FTI finding deserves a second look: initiatives are beating their business cases in part because those cases were conservatively scoped. Clearing a low bar is not the same as creating value.
Sources: EY AI Pulse Q4 2025; FTI Consulting 2026 Private Equity AI Radar, n = 200; Bain & Company; Grant Thornton 2026 AI Impact Survey, n = 950 (PE subgroup n = 100). Bars scaled to reported percentages.
This gap matters more now than it would have five years ago, because the industry has less time and less patience than it has had in a generation. PitchBook counts roughly 33,000 companies sitting in PE portfolios globally, and 34 percent have been held for more than five years, up from 28 percent a year ago. LPs are pressing for distributions and rewarding realized returns over paper marks. Every aging portfolio company will face the AI question in its exit process, and buyers are getting better at asking it. Diligence teams increasingly test whether a target's AI story is governed, measured, and documented, or whether it is a narrative built on pilot demos.
That creates an uncomfortable asymmetry. An AI capability narrative might support a valuation premium today. It will not survive a quality of earnings review in 2028. The firms that win the next cycle will be the ones that can walk a buyer through realized AI value creation the same way they walk through pricing actions or procurement savings: baseline, intervention, measured result. Almost nobody can do that today. Whoever builds that muscle first gets paid twice, once in EBITDA and once in multiple.
Across every dataset we reviewed, the companies generating real value share one trait. They do not treat AI as a technology initiative. They treat it as an operating model transformation, and they run it with the same discipline they would apply to an integration or a turnaround.
The middle market is where this discipline matters most, because the middle market has the least room for waste. McKinsey's data shows 83 percent of enterprises with more than 5,000 employees have deployed AI, roughly double the rate of companies with 50 to 499 employees. Large enterprises can afford to burn capital on pilot sprawl and internal AI teams. A $60 million revenue services business cannot. It gets one or two serious attempts at this, which is precisely why the middle market cannot afford to run the experiment driven playbook the Fortune 500 has spent three years proving does not work.
The good news is that the winning sequence is no longer a mystery. It is visible in every successful deployment we have studied, and it looks like this:
None of this is exotic. It is the same operational discipline that has always separated well run companies from the rest, applied to a new class of capability. What is new is the compounding effect. Each workflow deployed teaches the organization something about the next one. Companies that complete three or four cycles develop an institutional capability their competitors cannot buy, because it does not come in a license. The gap between organizations that have this muscle and those that do not is already visible in the data, and it is widening every quarter.
In May 2026, the market delivered the clearest possible verdict on where enterprise AI value gets created. It came from the model makers themselves.
Within the same month, OpenAI announced a dedicated deployment company backed by $4 billion from TPG, SoftBank, Brookfield, Bain Capital and others, and Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to build an AI native enterprise services business focused on deploying its models into enterprise operations. Think about what that means. The two organizations with the deepest understanding of frontier AI on the planet both concluded that models do not deploy themselves, that deployment is a distinct discipline, and that it is valuable enough to justify billions in dedicated capital. The world's most sophisticated investors agreed and wrote the checks.
If the frontier labs need deployment partners to translate capability into enterprise value, it should not be controversial to suggest that a middle market distributor, insurance brokerage, or healthcare services platform does too. Yet many companies still approach AI as a solo endeavor: hire a head of AI, stand up an internal team, run pilots, learn every lesson the hard way at full price. The failure statistics in this report are, in large part, the cost of that approach.
The economics argue for a different posture. Deployment expertise compounds across engagements in a way it cannot compound inside a single company. A partner who has redesigned quote to cash around AI a dozen times arrives knowing where the workflow breaks, which data problems surface in week six, and how adoption fails in a regional sales team. A company doing it for the first time discovers each of these at its own expense, inside its own hold period. FTI's survey identified talent as the single biggest constraint on scaling AI in PE portfolios, cited by 35 percent of fund leaders. That constraint does not resolve by every portfolio company competing for the same scarce operators. It resolves by borrowing pattern recognition that already exists.
To be clear, partnership is not outsourcing. The companies that get this right keep ownership of the outcome internal: the P&L target belongs to a line executive, the workflows belong to the business, and the institutional capability being built belongs to the company. What a deployment partner provides is compression. Compressed learning curves, compressed time to value, and a dramatically compressed list of expensive first time mistakes. In an industry where 34 percent of portfolio companies are past year five of their hold, compression is not a convenience. It is the difference between AI value showing up in this exit or the next owner's.
Stop counting pilots and start counting workflows in production with measured results. Pick one economic bottleneck this quarter, baseline it for 90 days, redesign the workflow around AI, and hold a named executive accountable for moving the number. Kill anything that has been in pilot for more than two quarters without a P&L owner. And be honest about capability: if your team has never taken an AI workflow from design through adoption to measured value, do not make your first attempt an unassisted one.
Treat AI value creation the way you treat every other lever: with a playbook, a baseline, and evidence that survives diligence. Audit the portfolio not for AI activity but for AI proof. How many companies can show a workflow in production, a pre deployment baseline, and a measured delta? That number is probably close to Bain's 20 percent, and it is your real exposure. Build or borrow a repeatable deployment capability that travels across the portfolio, because the alternative is 15 companies each learning the same lessons at 15 separate prices. And start preparing the AI section of your exit narratives now, with documentation a skeptical buyer will accept.
Change the questions. Not "what is our AI strategy" but "which workflows run on AI today, what did they measurably change, and who owns the number." Not "how many employees use AI" but "what would a buyer's diligence team conclude about our AI claims." The gap between those answers is the gap this report describes.
We believe the dominant narrative around enterprise AI remains backwards. The question is not what AI can do. The question is how work needs to change. The organizations that answer it fastest, with discipline and with help, will define the next generation of enterprise value creation. Everyone else will keep running successful pilots.
The AI Value Realization Report is a quarterly Deployly publication examining how enterprises convert AI capability into measurable business value, with a focus on middle market and private equity backed companies. Findings reflect our point of view, informed by deployment work and the third party research cited below. Statistics are drawn from research published between late 2025 and mid 2026 and are attributed to their original publishers.