Executive Summary
Almost every company is spending on AI, yet the returns are pooling in a few hands. PwC's AI Performance Study, released in April 2026, found that the top 20% of companies capture 74% of the economic value AI creates. The finding draws on a survey of 1,217 executives at large, publicly listed companies worldwide. The interesting part is that the gap has little to do with how much AI a company has deployed. This piece follows the study's counterintuitive answer to a familiar question: "We adopted AI, so why aren't we seeing results?"
What separated the front-runners from the rest was what they pointed AI at. Leaders aimed it not at trimming labor costs but at new revenue and reinventing their business models, and in doing so earned roughly 7.2 times the AI-driven financial gains of their peers, with operating margins four points higher. At the same time, instead of bolting on one more tool, they redesigned how work flows, and they expanded autonomous decision-making while tightening governance in step. The majority that chased efficiency alone are still stuck in pilot mode.
For practitioners, the message is plain. If AI isn't paying off, check what you're using it for before you swap out the model. And what underpinned the leaders' edge was not a flashier model but the AI foundations beneath it: clean data and reusable components you can pull off the shelf instead of rebuilding every time. Good results come from good foundations.
The figures below are the ones the study brought into focus. Together they show the size of the gap and the difference in behavior that created it.
Share of AI's economic value held by the top 20% of companies
Revenue and efficiency gains for AI leaders vs. their peers
Likelihood of reporting business model reinvention
Executives surveyed (25 sectors, global)
Drawn at its simplest, the gap fits on a single chart. Among companies running the same technology and comparable investment, nearly three-quarters of the economic value AI creates flows to the top fifth. The difference in area is where "We adopted AI, so why aren't we seeing results?" begins.
The top 20% of companies capture 74% of AI's economic value. The remaining 80% split the other 26%. (PwC 2026 AI Performance Study)
74:20 — Why It Matters
On April 13, 2026, PwC published its 2026 AI Performance Study, based on a survey of 1,217 executives at large, publicly listed companies. The sample spans 25 sectors, and most respondents come from companies that have already invested heavily in AI. So the study isn't about whether to use AI; it's about how outcomes diverge among the companies already using it.
The first number that stands out is how lopsided the distribution is. The top 20% of companies capture 74% of the economic value AI generates. Same markets, same technology, comparable levels of investment, yet the returns concentrate in a few hands. Companies classified as AI leaders pulled ahead of their peers by roughly 7.2x on AI-driven revenue and efficiency gains, and their operating margins ran four percentage points higher. The study's central warning is that this gap is widening, not closing.
There is a mechanism behind that widening. Leaders scale proven use cases faster, expand autonomous decision-making safely, and learn faster as they go. Differences in the speed of learning compound over time. That's why PwC is blunt that "the gap will widen further unless companies change their approach." Even from a similar starting line, if one side learns a little faster every year, the distance after a few years becomes hard to close.
How it's measured: the AI fitness index
It matters that this gap is a measurement, not an impression. PwC adjusted AI-driven revenue and efficiency gains against the industry median, then converted each company's AI capability into a single score it calls the "AI fitness index." Because every industry starts from a different baseline, the index tracks relative position within a sector rather than raw performance. That adjustment is also what brings the gap into view when efficiency-only gains would otherwise vanish into the industry median.
The index rests on two axes and nine factors. One is AI use: how broadly, deeply, and skillfully a company applies AI to real work and decisions, and whether that firepower is aimed at growth opportunities. The other is AI foundations: the base that makes such use trustworthy and scalable. Foundations break down further into six areas, namely strategy, investment, data and technology, workforce, governance and risk, and innovation. Taken together, the nine factors fan out into dozens of concrete management and investment practices. Leaders scored high on both axes at once; companies strong on only one axis never reached the top of the gap.
The crucial point is that the two axes don't act in isolation; they multiply. PwC's analysis shows that the stronger the foundations, the larger the performance lift you get from increasing use. With weak foundations, raising use only nudges performance upward; with solid foundations, the same increase in use produces a far steeper climb. Foundations hold up the floor of performance and, at the same time, act as an amplifier that steepens the slope of use.
Key takeaway: The 74:20 gap isn't about who adopted AI; it's about who did things differently among the companies that already did. That difference is measured along two axes and nine factors: the depth of use and the strength of the foundations. The two axes don't add; they multiply.
Surprise No. 1: Growth, Not Efficiency
For many companies adopting AI, the first goal that comes to mind is cost reduction. Automate the work people used to do, cut labor costs, speed up throughput, lift productivity. It's intuitive, and the short-term payoff is easy to measure. PwC's analysis turns that instinct on its head. The single strongest driver of AI financial performance wasn't efficiency but the ability to capture growth opportunities, especially where industry boundaries are dissolving and convergence opens up new markets. On its own, that capability outweighed the pursuit of efficiency.
That's why PwC titled the study "Want ROI from AI? Go for growth." Leaders were 2.6x more likely than their peers to report an improved ability to reinvent the business model itself through AI. And they were two to three times as likely to identify and actually pursue growth opportunities arising from industry convergence. They treated AI not as a tool for cutting costs but as an engine for opening new sources of revenue.
Why efficiency alone never closes the gap
None of this means efficiency is pointless. The problem is that efficiency is an advantage competitors quickly match. A tool that automates a given task is one rivals soon adopt too, and the savings become the industry's new baseline. So the ROI of efficiency-only players converges toward the industry median, and the gap on the AI fitness index barely moves. Growth works the other way: it's a game of creating new markets or claiming them before anyone else, and once you're ahead, the lead accumulates.
It helps to notice that the two advantages run on different clocks. Efficiency is an advantage that levels out. The benefit of cost savings is realized once and then stops, and the moment a competitor brings in the same tool, the difference converges to zero. Growth is an advantage that accumulates. The company that opens a new revenue stream first banks data, customers, and learning in that market ahead of everyone, and those assets become the springboard for the next round of growth. Efficiency is a game of subtraction, so it has a floor at zero cost; growth is a game of addition, so it has no ceiling. That is why, running the same AI, the company betting on efficiency converges while the one betting on growth diverges.
The growth opportunity leaders caught best sat at the points of industry convergence. Where boundaries dissolve, as in healthcare and finance, or manufacturing and software, the advantages of incumbents waver and new rules get written. Because AI is strong at quickly connecting heterogeneous data and domains, the ability to design new products and services first at these convergence points becomes the ability to claim them first. The figure that leaders were two to three times as likely to identify and pursue convergence growth means they moved into places where no one yet knew the rules.
So here is the study's first answer to "We adopted AI, so why aren't we seeing results?" If efficiency savings were the only goal, the results aren't absent; they're buried in the industry average and therefore invisible. To produce measurable gains, you have to aim AI not at what you can reduce but at what you can open up.
Key takeaway: The strongest single driver of AI financial performance was capturing growth opportunities, not efficiency. Efficiency is a leveled advantage that everyone catches up to; growth is a first-mover advantage that accumulates the gap.
Surprise No. 2: Redesign, Not Tools
The intent to use AI for growth isn't enough on its own. The second difference PwC identified is how companies bring AI into the organization. Where laggards stopped at bolting one more AI tool onto existing work, leaders reworked the flow of work itself. They were roughly 2x more likely than their peers to redesign workflows rather than simply add a tool. The tools may be the same; what differs is how deeply they settle into the way work gets done.
3.1 Scaling autonomy and governance together
At the heart of that redesign is widening the range of decisions AI makes without a human in the loop. Leaders were 2.8x more likely to increase the number of decisions made without human intervention. What's striking is that they tightened AI governance at the same time. They didn't loosen control as autonomy rose; they tightened it precisely because autonomy rose. They were also about 2x more likely to run AI at a higher level of sophistication, such as performing multiple tasks within guardrails or improving on its own.
Strengthening governance showed up not as a slogan but as concrete machinery. Leaders were 1.7x more likely than their peers to have a Responsible AI framework in place, and 1.5x more likely to run a cross-functional AI governance committee. The effect of that machinery shows up in people. Employees at leader companies were 2x more likely to trust the output AI produced. Trust leads straight to adoption. Employees have to believe the results before they'll use AI in real work, and only then can the range of autonomous decisions widen. Autonomy, governance, and trust link into a single loop.
What that combination means is clear. Autonomous AI is dangerous without governance, and governance without autonomy produces no results. Leaders treated the two not as a trade-off but as a pair that has to move together. PwC's companion piece from the same period, "No more pyramids," makes the same point about redesigning workforce structure for the agentic AI era: the division of labor between people and AI has to be redrawn from scratch.
3.2 Don't reinvent it every time — reusable components
The third difference, and the one that leads into foundations, is reuse. Leaders were 2.4x more likely to build reusable, centrally cataloged AI components. Rather than rebuilding data pipelines, models, and validation steps from scratch on every project, they register a once-validated asset and pull it off the shelf when needed. This isn't just about speed; it builds consistency of quality. Build from scratch every time and you introduce fresh defects every time; reuse a validated component and quality accumulates.
Stand the behavioral differences seen so far on a single bar and you can see where leaders pulled furthest ahead of their peers. The chart below sets "peers = 1" and shows how many times more leaders earned, or how much more often they acted, in each area. The 7.2x in total financial performance on the far left is the result; the multiples to its right are the behaviors that produced it.
Peers = 1. The 7.2x in total financial performance is the result; the rest are the multiples of the behaviors that produced it. (PwC 2026 AI Performance Study)
The table below pulls together the leader-versus-laggard behavioral differences covered in this section, governance machinery included. Every figure is the leaders' relative likelihood of exhibiting that behavior compared with their peers.
| What leaders did differently | Likelihood vs. peers |
|---|---|
| Report improved business model reinvention | 2.6x |
| Identify and pursue industry-convergence growth | 2–3x |
| Redesign workflows rather than add tools | 2x |
| Expand decisions made without human intervention | 2.8x |
| Run AI at higher sophistication (multi-task and autonomy within guardrails) | ~2x |
| Build reusable, centrally cataloged AI components | 2.4x |
| Have a Responsible AI framework | 1.7x |
| Run a cross-functional AI governance committee | 1.5x |
| Employees trust the output AI produces | 2x |
Key takeaway: Leaders didn't add AI as a tool; they wove it into how work gets done. They scaled autonomy and governance together, and they reused validated assets so quality accumulated. That last difference, reusable components, leads straight into the discussion of foundations in the next section.
The Laggard's Traps — Pilot Mode and the Efficiency Trap
Having seen what the leaders did, it's time to look at where the majority stalled. PwC's portrait of the laggard comes down to two traps: getting stuck in pilot mode, and the efficiency trap. Each one reinforces the other.
4.1 Stuck in pilot mode
Plenty of companies produce impressive demos and small wins and then fail to move past them. Experiments scattered across departments never scale into operations, and a validated pilot never gets absorbed into standard work. This kind of stall, often called "stuck in pilot," is usually a foundations problem more than a technology one. When data sits in different formats across departments and there's no way to reuse an asset validated in one place somewhere else, every pilot has to start over from scratch.
4.2 The efficiency trap
The second trap is the consequence of the efficiency bias from Section 2. When efficiency savings become the only goal because they're easy to measure, the gains are real but they sink into the industry median and never surface as a competitive gap. Executives ask, "We invested in AI, so why aren't we the ones pulling ahead?" The truth is that no one pulls ahead because everyone is chasing the same efficiency. The efficiency trap is a problem of direction, not of effort.
4.3 Why they can't escape — scattered data and the absence of reuse
There's a common structural cause behind why both traps persist: data sits in different formats and different systems across departments. Data that one team painstakingly cleaned and validated starts again from zero in the team next door. With no asset to reuse, every project pays the same cost of collecting and cleaning data all over again. Because that cost recurs every time, the barrier to a big bet rises, and the company retreats to small, safe efficiency projects. The real cause of falling behind is neither the model nor the will, but the absence of assets you can pull off the shelf.
When the two traps interlock, they become a vicious cycle. Pilots don't scale, so the company can't place the big bets that growth requires; so it retreats to easily measured efficiency; and because efficiency gains get leveled, it falls back into small pilots. And underneath every step of that loop lies the same deficit: no data to reuse. The point where this loop breaks is exactly the foundation that separated leaders from laggards.
Key takeaway: Falling behind is a structural problem, not a lack of effort. At the bottom of the vicious cycle — pilots that won't scale, retreat into efficiency — lies a common cause: the absence of clean data and assets you can pull off the shelf.
It Comes Down to Foundations — AI Foundations
The four preceding sections converge on one conclusion. The leaders' edge didn't come from a better model; it came from the foundation that model stood on. Recall that PwC's AI fitness index is built on two axes, AI use and AI foundations. Half the index is foundations. And the foundation isn't anything glamorous. Data, governance, and reusable components (the catalog) are what a foundation actually consists of.
5.1 A clean foundation you can pull off the shelf
The reusable components from Section 3 (2.4x) and the pilot stall from Section 4 are really two sides of the same coin. When validated data and components are pooled centrally and ready to pull off the shelf, a new project starts not at zero but at 70. Without that, you start from zero every time and wear yourself out repeating the same data cleaning and validation. This is also why leaders could place big bets on growth: a solid foundation lowers the cost and risk of each new attempt, which lets you be bolder.
Scaling autonomy and governance together is impossible without a foundation too. To increase the share of decisions made without a human in the loop, you have to trust the quality of the data those decisions rest on. When data is inaccurate or of unclear provenance, autonomous AI rapidly mass-produces wrong decisions. So the more leaders raised autonomy, the deeper they dug into the foundation of data and governance. The observation from Section 1 that "the two axes multiply" takes concrete shape here. Only when the foundation is solid does raising use produce a steeper climb in performance, and only then does raising autonomy add speed rather than risk.
5.2 Good results come from good foundations
The most practical conclusion this study leaves for practitioners is a reordering of priorities. When AI isn't paying off, the first thing you want to reach for is a better model, but the real variable this study points to sits beneath it. What to use it for (growth), how to weave it in (redesign), and what to build it on (foundations). Swap out the model without answering those three questions and the gap won't close. The next section turns these three questions into a working sequence and lays out what to tackle first.
Editor's note. The foundation layer this study points to — clean data, reusable validated assets, and the data governance that holds up autonomous AI — is exactly the territory the Pebblous Data Communication Team works in day to day. AI-Ready Data ultimately comes down to building "a clean foundation you can pull off the shelf," and this study shows, across 1,217 companies, that this foundation is what divides the performance gap. That's why this report caught our attention.
Key takeaway: Half of the AI fitness index is foundations. The leaders' edge came not from the model but from the foundation of data, governance, and reusable components. Good results come from good foundations.
A Practitioner's Playbook — What to Tackle First
Even a precise diagnosis is hard to act on without an order of operations. Translate the study's findings into a working sequence and a clear priority emerges. Swapping out the model is not at the top of that list. Trace the path leaders walked in reverse and three steps remain.
First, identify the growth use cases. Bring in tools before deciding where to use AI and the effort almost always drifts toward efficiency projects, because they're easy to measure. So the starting point has to be "what can we open up," not "what can we cut." Look in particular at the seams where industry boundaries dissolve, where adjacent markets meet your own data, and frame new revenue hypotheses there. The output of this step is not a tool but a list of growth questions worth solving with AI.
Second, redesign workflows around those use cases. Bolting one more AI button onto existing work doesn't create a gap. Redraw which of the steps a person used to handle now goes to AI, which a person reviews, and which is left to autonomy. Widen the range of autonomous decisions, but tie that range to governance as you go. Growing Responsible AI principles and cross-functional review at the same pace as autonomy was the leaders' way.
Third, invest in the foundation. For the two steps above to be repeatable, you can't start from zero each time. Register validated data, models, and checks in a central catalog so the next project can pull them off the shelf. Pulling data out of departmental silos and into a consistent format with clear provenance belongs here. This step is the least visible, but it's the engine that keeps the first two running again and again rather than once.
The three steps form a loop, not a straight line. The stronger the foundation, the lower the cost and risk of each new growth use case, which makes bolder redesign possible, and the results of that redesign accumulate back into the foundation as assets. That loop is how leaders learned faster every year. So when AI isn't paying off, the first question to ask is not "what's the better model" but "what are we doing this work on top of." Good results, in the end, come from good foundations.
References
Core study — PwC 2026 AI Performance Study
- 1.PwC. (2026). "Three-quarters of AI's economic gains are being captured by just 20% of companies — with the leading companies focused on growth, not just productivity." PwC Press Release (2026-04-13). pwc.com
- 2.PwC. (2026). "Want ROI from AI? Go for growth." PwC 2026 AI Performance Study. pwc.com
- 3.PwC. (2026). "Want ROI from AI? Go for growth" (full report PDF). PwC. roi-from-ai.pdf
- 4.PwC. "Decoding ROI from AI." PwC (AI Performance hub). pwc.com