AI value creation is decided by the choices companies make around the technology, because every company can buy the same models. The companies that make those choices well are still few, and the distance between them and the rest is wide. PwC’s 2026 study finds that 20% of companies capture 74% of AI-driven returns.[8] In McKinsey’s 2026 global survey, only 37% of respondents say AI has contributed positively to their organisation’s EBIT, a share McKinsey describes as “essentially unchanged from 2025”.[1] The first of those choices is recognising that AI is not always the answer.
Exhibit 1. Every firm ties AI value to measuring results, and five of the six also name redesigning the work and building the technology and data underneath it.[1-12]
Text version: what each firm’s research says drives value from AI. McKinsey: redesign workflows, track each initiative’s impact, build technology and data foundations, have senior leaders own AI, set growth goals, invest in talent and change. BCG: reshape workflows end to end, track AI value in the P&L, build an agent-ready platform on governed data, reinvent the workforce, grow and reinvest, make the big choices in one C-suite programme. Bain: redesign the work, measure outcomes, treat data as a prerequisite, lead AI as a CEO-owned transformation, fund next use cases from realised returns, widen autonomy step by step. PwC: embed AI broadly and deeply, keep a road map with owners and tracking, build focused foundations, build employee trust, aim at growth, fund like an investor. IBM: redesign the highest-value workflows, measure and fund on evidence, reduce technical debt, rewire leadership and organisation, reinvest productivity gains in growth, govern outcomes. Accenture: measure each workload’s outcome in money, make use and results visible, name an owner, require a value case, bill teams for use, govern alongside access.
Those choices run through the people who use AI, the process it enters, the data it depends on and the way it is funded and run. Organisations in the GCC face them now, because adoption and budgets in the region are rising while measurable value remains rare.[2] That makes the approval of the next AI budget, more than the choice of the next model, the moment that decides whether the spending pays.
AI is not always the answer to a performance problem
Heavy use of AI does not protect a company whose business model is failing. In BCG Institute’s study of US public companies, a small minority of the adoption leaders have declining margins and growth.[4] Their trouble comes from “fundamental business model challenges that AI cannot overcome”, such as commoditised core offerings or legacy models displaced by digital disruptors.[4] In BCG’s words, “AI amplifies a strong strategy; it does not substitute for one.”[4] For a company in that position, more AI spending leaves the underlying problem in place and adds cost on top of it.
A board can test for that position before it funds anything by asking whether customers would still pay enough for the offer if it were delivered perfectly and at lower cost. When the answer is no, the repair lies in the offer, its pricing or the business model, and AI spending can wait until that repair is under way. When the answer is yes, AI has a sound business to amplify, and the harder questions move to the organisation around it.
A large share of AI value is lost inside the organisation, where a better model does not reach. IBM’s 2026 study of AI returns names fragmented processes, inconsistent measurement, poor visibility and technical debt as the factors that erode returns, and estimates that a fifth of AI value is lost to friction inside the enterprise.[11] It also warns that organisations “often try to solve process, data, and architecture problems with more powerful AI, when fixing the underlying environment would do more to improve performance.”[11] The fragmented process and the missing measures survive the upgrade and keep holding the return down.
Talent separates the AI leaders far more than technology does
Most chief executives still expect their edge to come from the model. In IBM’s 2026 CEO study, 63% expect their competitive advantage in 2030 to come mainly from the sophistication of their AI models.[10] The evidence from the companies already ahead points somewhere else. In BCG Institute’s study, the leading tier scores barely higher than the active tier below it on technology and deployment, while its talent score is nearly three times as high.[4] BCG’s reading is that “tools have commoditized; every company can buy the same AI products from the same vendors.”[4] A plan judged mainly on its model is judged on the part a competitor can copy most easily.
Talent is the constraint that more spending on models cannot lift. Jacek Olczak, Group CEO of Philip Morris International, told IBM’s CEO study that “It’s not the technology that is the limiting factor. We are the limiting factor.”[10] BCG Institute’s leaders spread AI skills well beyond the specialists, and BCG argues that distributed fluency lets an organisation develop “a richer pipeline of use cases than any central team could generate alone.”[4] A central team can build only what it hears about, while people across the business see the problems worth solving in their own work. For a board, the reach of a plan’s skills budget says more about its chances than the name of its model.
Aramco looks to its AI-trained employees, not its data analysts, for its use cases. Its President and CEO, Amin H. Nasser, told the World Economic Forum in Davos in January 2026 that the 6,000 employees the company trained on AI “are the subject matter experts that come up with the use cases”.[13] Use cases that come from the people who run the work start out tied to a real business problem. They also come with someone who will notice whether the problem gets solved, and every later measure of value depends on that person.
The process and its data decide how much potential becomes value
Generative AI has so far turned less of its potential into value than older automation did. By Bain’s estimate, generative AI has realised 38% of its potential value across functions, against 52% for traditional automation after two decades of process mapping, integration work and change management.[6] The gap suggests that much generative AI is being deployed without comparable groundwork. Bain calls automating a broken process “the single most costly mistake in AI deployment”, because AI leaves the faults in place, makes them run faster and raises the cost of undoing them later.[5]
The return depends more on how the work is redesigned than on which model is bought. Among the companies Bain works with, technology is the smaller part of the bill, since “for every dollar spent on technology, four are spent on people and process”.[7] McKinsey’s 2026 high performers “fundamentally redesign workflows that are enabled by AI” instead of inserting AI into existing ones.[1] A multinational consumer company in BCG’s report on agentic AI mapped the work and the tacit knowledge behind it, redesigned each process to balance AI and human tasks and only then scaled the best ideas.[3] In marketing, the time employees spent on routine work fell by up to 70% while output quality doubled.[3] For a board, the order of the work is the lesson, because the gains followed the mapping and redesign that came before any scaling.
Imperfect data is a weak reason to delay AI, because the first uses can run where the data is already sound. Bain calls imperfect data infrastructure “the most cited reason to defer AI investment, and the least valid one”.[5] Its advice is to start with bounded workflows where the data is accessible, use AI on the data problem itself and keep large-scale modernisation for the uses that need it, letting early wins help pay for it.[5] A company that waits for clean data everywhere pays for modernisation before it knows which uses will need it. The useful question for a board is whether the first use of AI can run on data the company already trusts.
Data quality still matters, and Nasser told the Davos audience that it was the most important thing in Aramco’s AI work.[13] Raising it can start with repairing the process that degrades it. At a global resources company in BCG’s agentic AI report, master-data change requests went through a manual process with a long backlog.[3] It was so slow that “some employees gave up, degrading data quality over time.”[3] Cleaner data could not have arrived while that process kept running. The company rebuilt the request so that an agent drafts it from a plain-language description, which cut the process to a small fraction of its former steps and made it more accurate.[3] Employees now fix data errors when they spot them.[3]
The leaders use AI to grow as well as to save
Efficiency is where most AI programmes begin, but the leaders do not stay there. About 80% of McKinsey’s 2026 high performers pursue efficiency gains from AI, the same share as other respondents, yet most of the high performers also use AI for growth or innovation.[1] Among BCG Institute’s leaders, growth is the dominant path to value, and only one in ten takes efficiency as its main path.[4] Growth marks the leaders alongside talent and redesigned work, while the pursuit of savings is common to most organisations. A board that approves only savings cases funds what about four in five of McKinsey’s respondents already pursue.[1]
Savings from AI are capped by the cost they remove, while growth is limited by the market a company can reach. A programme framed entirely around savings therefore sets a ceiling on its own return. BCG adds a competitive reason, noting that doing the same things faster “is unlikely to yield durable competitive advantage when the underlying technology is widely available.”[4]
Growth can also come from changing the offer itself, and PwC finds its AI leaders more than twice as likely as other companies to use AI to reinvent their business model.[8] John Deere’s See & Spray is one of the cases PwC describes. The camera-based system identifies weeds and sprays herbicide only where it is needed, and Deere packaged it in a service-like model that let farmers pay for verified outcomes.[8] In the 2024 growing season it saved farmers an average of 59% of their herbicide, according to the Deere figures PwC cites.[8] In PwC’s account, the service model positions Deere to capture more value from a scalable services revenue stream than from a one-time hardware advantage.[8] A business case that counted only the farmers’ savings would have missed where Deere’s own value lies.
Money for AI rarely follows return, and running cost is often left unmanaged
Organisations “tend to spread AI budgets across departments somewhat evenly, rather than focusing on the functions that deliver the strongest returns”, according to IBM.[11] Only 6% of organisations direct their five largest AI investments to the five functions generating their highest AI returns, and that group achieves 22% higher AI ROI than the rest.[11] Once returns are measured, an even spread has nothing left to justify it, and a budget left that way keeps paying for the weakest uses.
Moving money towards returns needs someone who answers for each return. IBM’s CEO study advises chief executives to “Name a business leader for every competitive AI investment and hold them accountable for adoption, outcomes, and value realization”, with each initiative tied to a revenue, margin or growth metric.[10] PwC’s leading companies are more likely than others to track business impact systematically and to hold senior leaders directly accountable for AI outcomes.[8] Without a named owner, a system that goes live but changes nothing in how the work is done can still be counted a success, because nobody is measured on the difference.
Most token spending cannot yet be tied to a financial return. Accenture estimates that only one dollar in five of token spend shows up as a quantified financial outcome.[12] The rest sits “in a space companies believe is productive but cannot prove.”[12] An outcome that was never defined at approval is hard to establish once a use case is running. ABN AMRO avoids the problem at the start, according to Accenture, by requiring every AI use case to contribute to a defined value outcome before it reaches production.[12] By May 2026, every use case it had in production was mapped to one.[12]
At DBS, the tracking of business outcomes decides where capital goes. Its Chief Executive Officer, Tan Su Shan, says the bank’s commitment to that tracking “ensures that capital is allocated only to initiatives with proven traction, while allowing for course-correction for those that fall short.”[14] The second half of that sentence matters as much as the first, because a portfolio that can only add initiatives keeps funding the ones that have stopped earning.
Running cost is a part of the return that management controls directly, and one lever within it is which model handles each task. Frontier models cost 10 to 20 times more per token than capable mid-tier alternatives, by Accenture’s estimate, yet 54% of AI requests already go to a more expensive tier than the task needs.[12] The technology chief of one restaurant chain told Accenture that users had defaulted to the most expensive model even for menial tasks, because they had not yet learned when cheaper ones were enough.[12] Overspending of this kind is rarely a decision anyone makes, and in Accenture’s cost scenarios the cost curve is “a management choice”.[12]
Exhibit 2. The firms name six ways to control what AI costs to build and run, and their research finds wide gaps in applying them.[3][8][11][12]
Text version: six controls on what AI costs, with one finding each. Send each task to the model it needs: 54% of AI requests go to a costlier tier than needed (Accenture). Manage token spending: 50% of lagging companies give access without tracking spending, or track it without setting a limit (BCG). Make AI spend visible: 85% of technology leaders lack real-time visibility (IBM). Judge cost per outcome: 35% can calculate it for their largest use case (Accenture). Make teams pay: 7% have any chargeback (Accenture). Reuse AI components: 51% of leading companies do so to a large or very large extent, against 21% of others (PwC).
Accenture recommends a routing policy enforced at a central gateway, which keeps frontier models for the small share of work that needs complex reasoning and planning and sends the rest to mid-tier or open-weight models.[12] Routing by price alone would be its own mistake, and BCG cautions that in many cases “the cheapest model is not the right answer”.[3] By BCG’s reasoning, a more capable model that completes a task first time can cost less than a weaker one that retries or produces work a person has to redo.[3] The unit to manage is therefore the cost of each successful outcome at the required quality, which a token price alone does not show.
The leading companies treat token spending as a decision in its own right. Nearly three-quarters of BCG’s future-built companies have made an explicit call on token spending and manage it towards a return, while half of the laggards either grant access without tracking it or monitor spending without setting any limit.[3] Accenture finds a common sequence in which “access came first” and governance “came when the bill arrived”, and warns that reversing the order later, under financial pressure, is harder and more expensive.[12] A spending limit set at approval lets a board see early when a use of AI begins to cost more than it returns.
Cutting the technology underneath AI to pay for it lowers the return it was meant to raise. More than half of the executives IBM surveyed say they face pressure to cut non-AI technology spending to free up resources for AI.[11] IBM calls this “a dangerous contradiction”, because organisations are scaling AI while reducing investment in the foundation it depends on.[11] Organisations that count modernisation and technical-debt remediation within their AI business cases achieve 29% higher AI ROI than those that treat modernisation as a separate conversation.[11] The systems that would otherwise be cut are the ones every use of AI reads from and writes to.
GCC organisations have the ambition for AI but lag on what turns it into value
GCC organisations have adopted AI widely, yet few earn measurable value from it. In McKinsey’s 2025 survey of executives and board directors across the six GCC countries, 84% say their organisations have adopted AI in at least one business function.[2] Only 11% qualify as what McKinsey calls value realisers, organisations scaling AI that can attribute a measurable share of their earnings to it.[2] Appetite for spending is high, since nearly nine in ten plan to raise AI budgets in the coming year.[2] For a GCC board, a larger budget on its own is therefore unlikely to close the gap between use and value.
Exhibit 3. In McKinsey’s 2025 GCC survey, 84% of organisations use AI in at least one business function, 31% are scaling it or have fully deployed it, and 11% also attribute at least 5% of their earnings to it.[2]
Text version: share of GCC organisations (McKinsey, 2025). 84% use AI in at least one business function; 31% are scaling AI or have fully deployed it across the organisation; 11% are scaling or have scaled AI and attribute at least 5% of their earnings to it; 37% of those not yet getting value from AI rate their technology and data as strong.
The shortfall sits in the capabilities that deliver AI more than in strategy. Among respondents outside the value-realiser group, nearly three-quarters rate their organisation strong on aligning AI with strategy.[2] Fewer than half say the same of talent and operating model, change management, or technology and data.[2] Most of the value realisers rate themselves strong on all three, although the group is small.[2] For a GCC board, another AI strategy document adds to what the organisation already does well. The next budget is better aimed at the capabilities the realisers have and the rest lack.
In Saudi Arabia, early gains from AI have yet to show up consistently in reported financial returns. Saudi organisations report stronger intermediate outcomes from AI than the global average, according to PwC, but its study notes that “these gains are not yet being reported as financial returns as consistently as they need to.”[9] PwC treats this as a transition still under way from adoption to enterprise-wide value.[9] Until the gains reach the accounts consistently, a board has little evidence with which to tell the uses of AI that pay from those that only add activity.
What boards should do, and avoid, with the next AI budget
Funding approval is the last point at which a board can still put process, data and measurement ahead of the choice of model. Bain suggests asking how a process would be designed from scratch before any discussion of technology, and two further questions belong at the same meeting.[5] Asked before approval, they decide what the first money buys, whereas after launch they can only explain why a system is not paying back.
If the process were designed from scratch today, what would it look like?
Does the data the decision depends on exist, and can it be trusted?
Can the result be tied to a defined business outcome before anything goes live?
A process that would look very different if designed today should be redesigned before AI is added, since automating it as it stands is the mistake Bain ranks as the costliest.[5] Missing or untrusted data points the first money at a bounded workflow where the data is usable, or at the process that degrades it. And if the redesign question leads back to what the company sells and how it charges, the repair lies outside AI altogether.
The approval should also carry conditions that make a running system easy to judge a year later. BCG’s view is that every material workflow needs “a named owner, an intended outcome, and explicit budget guardrails.”[3] A routing plan and a spending limit fix what the system may cost. Each condition turns a promise in the proposal into something the board can check at the first budget review.
GCC organisations already have the executive commitment and the budgets that most AI programmes ask for first. The next round of approvals will decide whether that spending joins the minority that earns a return. The model in each proposal is the part every competitor can buy, while the value will come from the choices a board insists on around it before the money is released.
Frequently asked questions
How are major companies using AI?
The companies getting most from AI redesign the workflows it enters, spread AI skills beyond specialists and use it for growth as well as efficiency. McKinsey’s high performers redesign workflows enabled by AI, and most also pursue growth or innovation.[1] Aramco’s chief executive, Amin Nasser, says the 6,000 employees it trained on AI come up with its use cases.[13]
Is AI ROI real?
Yes, AI ROI is real for a minority and thinner for most. PwC’s 2026 study found that a fifth of companies capture nearly three-quarters of AI-driven returns,[8] and fewer than four in ten McKinsey respondents report AI adding to EBIT.[1] IBM finds that organisations directing their largest AI investments to their highest-return functions earn higher AI ROI.[11]
What is AI value creation?
AI value creation is the gain from AI that shows up in a company’s earnings as a quantified financial outcome. Research from McKinsey, BCG, Bain, Accenture, PwC and IBM shows it depends on the work around the technology, from the process it runs in and the data it uses to the people using it, its running cost and its controls.
How is AI implemented in business?
Successful AI implementation starts with the business problem, redesigns the process and defines the business outcome before AI is added, and can use AI to fix weak data. Companies Bain works with spend four dollars on people and process for every dollar on technology.[7] ABN AMRO requires every use case to link to a defined value outcome before production.[12]
