Every team at the 2026 World Cup gets the same AI. Every business today gets the same tools too. Here is what actually separates the winners.
Quick answer: Most larger businesses now use AI, but only about 5 to 6% capture large financial value from it, according to BCG and McKinsey. The winners are not using better tools. They redesign workflows, fix the data feeding AI, make fewer and bigger bets in core functions, and decide clearly where humans must review.
Every team at the 2026 World Cup had the same AI. FIFA and Lenovo built Football AI Pro, a generative assistant for match analysis, and made it available to all 48 participating teams. Lenovo's executive vice president described the aim as "providing the same cutting-edge resources and analytical tools to all the teams". Nobody expected that to make the 48 teams equal, and it did not. Business is now in the same position: almost every competitor can buy the same models and tools, yet the financial returns from AI are landing with a small minority. This article looks at how concentrated that value really is, and what the winners do differently.
Among larger companies, nearly everyone does. Across all businesses, far fewer do, and the gap between big and small firms is the first sign that "having AI" means very different things.
McKinsey's 2025 global survey of 1,993 respondents found that 88% of organisations use AI regularly in at least one business function. In Singapore, IMDA reported that 62.5% of non-SMEs had adopted AI in 2024, against 14.5% of SMEs, although SME adoption had more than tripled from 4.2% the year before. In the United States, the Census Bureau's business survey shows overall AI use hovering between 17% and 20% of firms from December 2025 to May 2026, with 37% of firms of 250 or more employees using it and under 20% of the smallest firms.
Individuals are further ahead than their employers. A KPMG and University of Melbourne study of more than 48,000 people in 47 countries found that 66% of people use AI regularly. MIT's NANDA research found workers using personal AI tools at over 90% of companies, while only 40% of companies had official LLM subscriptions.
| Measure | Figure | Source |
|---|---|---|
| Organisations using AI in at least one function (large-firm survey) | 88% | McKinsey State of AI, 2025 |
| Singapore non-SMEs vs SMEs that adopted AI (2024) | 62.5% vs 14.5% | IMDA, October 2025 |
| US firms using AI (Dec 2025 to May 2026) | 17% to 20% | US Census Bureau, May 2026 |
| People who use AI regularly | 66% | KPMG and University of Melbourne, 2025 |
| Companies that are "future-built" and generating substantial value | 5% | BCG, September 2025 |
| AI high performers (5% or more of EBIT from AI) | About 6% | McKinsey State of AI 2026, via FM Magazine |
So the headline is true for the companies most surveys reach, and the tools themselves are cheap and widely available. That is exactly why access has stopped being an advantage.
Very. The major studies use different methods, but they converge on the same shape: roughly one company in twenty is capturing large, measurable value, a larger middle group sees some gains, and most see little.
BCG's September 2025 survey of 1,250 senior executives found that 5% of companies qualify as "future-built" for AI, 35% are scaling, and 60% report minimal revenue and cost gains. Compared with laggards, the future-built group reported 1.7 times the revenue growth, 3.6 times the three-year total shareholder return and 1.6 times the EBIT margin. BCG's 2024 study had found a similar split, with 74% of companies yet to show tangible value from AI and only 4% generating significant value consistently.
In McKinsey's 2025 survey, 39% of respondents attributed any EBIT impact to AI, most of them less than 5%, and about 6% qualified as high performers. The 2026 edition, The State of AI in 2026: On the Road to ROI, found that 80% of respondents said AI had improved individual productivity, but only 37% attributed at least some EBIT impact to it, and the high-performer share stayed at about 6%. McKinsey's own site timed out when we tried to open the report, so these figures are cited from reputable coverage of it.
MIT's Project NANDA report, published in 2025, estimated $30 billion to $40 billion of enterprise spending on generative AI and concluded that 95% of organisations were seeing no business return. It found that 60% of organisations had evaluated enterprise-grade AI systems, 20% reached pilot and only 5% reached production.
Treat the 95% with care. Coverage describes the method differently: Virtualization Review lists 300+ public AI initiatives, 52 interviews and 153 survey responses, while Fortune reported 150 interviews, a survey of 350 employees and 300 public deployments. Either way it is a small sample, and it measures P&L impact rather than softer gains. Read the three studies together and the honest range is that 5% to 6% of companies capture large value, while somewhere between a quarter and two-fifths see some measurable return.
Because the tool is the cheapest, most replaceable part of the system. Results depend on the data feeding it, the workflow built around it and the judgement of the people using it, and those differ enormously between companies.
MuleSoft's 2025 benchmark of more than 1,050 IT leaders found the average enterprise runs 897 applications, of which only 29% are integrated, and 80% cite data integration as their biggest obstacle to AI. Validity's 2025 survey of CRM users found that 76% say less than half of their CRM data is accurate and complete. A model asked to score leads from a half-empty CRM gives confident answers built on gaps. The same model connected to clean, joined-up data looks like a different product.
BCG's 2024 work found that roughly 70% of AI implementation challenges stem from people and process, 20% from technology and only 10% from algorithms, yet companies spend disproportionately on the algorithm. MIT's researchers identified the same gap from another angle: "Most GenAI systems do not retain feedback, adapt to context, or improve over time". A chatbot dropped onto a website, with no change to how leads are routed or followed up, produces a demo rather than a result.
KPMG's study found that 66% of people rely on AI output without evaluating its accuracy, 56% have made mistakes in their work because of AI, and only 46% are willing to trust AI systems. Winners neither trust blindly nor refuse to trust. They decide in advance which decisions AI makes alone, which it drafts for a person and which stay human.
The same judgement applies to the numbers AI systems report. On an Indian D2C food and handicrafts store we manage, Meta's reported purchases ran roughly 30% above the store's actual orders. Every team using that ad platform sees the same dashboard; the ones who reconcile it against real orders make better decisions.
They treat AI as a change to how the business runs, not as a purchase. Across the BCG, McKinsey and MIT research, six behaviours stand out:
None of these is about the model. They are about data, process, ownership and money, which is why the gap persists even as the tools get cheaper. If you are deciding which processes deserve that investment first, our guide to AI automation use cases ranks them by effort and return.
Ad platforms are the clearest case. Two businesses can run the same Google or Meta campaigns with the same automated bidding and get very different results, because the algorithm only optimises toward the signals it is given.
If the conversion tag is broken, the landing page is slow or the account sends form fills instead of real sales, the AI optimises confidently toward the wrong thing. The difference between accounts is rarely the bidding model; it is whether someone built clean signal and watches what the system does. That is why we treat paid advertising as a measurement and feedback system first, and an AI tool second.
Across several client accounts we have found the same pattern: a website redesign or relaunch removed the Google Ads conversion tag, and the account kept spending while reporting zero or wrong conversions. The bidding AI was working exactly as designed, on data that had silently stopped being true.
The clearest lesson from our own builds is that the advantage sits in the system around the model, not in the model itself.
When we built our multi-tenant AI chat platform, we designed it to host five specialised AI agents, so a new client's chatbot is configuration rather than a new build. Every client uses the same underlying models; what differs is the knowledge, rules and handoffs configured for each business, and that is where the results come from. On a separate published project, our support automation platform, the same approach of designing the system around the model delivered 70% fewer support tickets.
The second lesson is about plumbing. One hosted chatbot we run went dark, and the cause was not the model. It was a CDN security rule and an expired credential. We now monitor DNS and CDN rules, tokens and database quotas as closely as the AI's answers.
The third came from our WhatsApp CRM platform. Meta throttles some marketing sends (error 131049), and our frequency cap originally counted those failed sends as "already messaged", which locked real customers out of other messages for 24 hours. The fix was small: count only delivered or sent messages. Finding it required someone to look at what the automation was actually doing, not just whether it was switched on.
Stop measuring AI progress by the number of tools you have. Everyone can buy the same tools, and the research is consistent that this alone produces little. Pick one or two processes in a core function where a better decision is worth real money, fix the data feeding them, redesign the workflow end to end and name an owner who answers for the result. Decide in writing what AI does alone, what it drafts and what stays human. Buy proven components where you can and spend your effort on integration and measurement. That is unglamorous work, and it is the reason a small group of companies keeps pulling ahead. If you want an outside view of where to start, this is the core of what an AI automation agency should do for you.
No. The research suggests the opposite: most companies, large and small, have not yet captured meaningful value, so the field is still open. Singapore SME adoption tripled to 14.5% in 2024, and IMDA reported that SMEs using AI-enabled solutions under the Productivity Solutions Grant achieved average cost savings of 52%. Starting narrow, with one well-chosen process, suits a small business better than a broad rollout.
For most businesses, buy or partner first. MIT NANDA's 2025 research found that tools bought from specialist vendors or built with partners succeeded about 67% of the time, against roughly a third for internal builds. Build only where your process or data is genuinely unique, and spend the saved effort on integrating the tool with your data and redesigning the workflow around it.
Tie each AI use case to one business metric before you start, such as cost per resolved ticket, lead-to-sale conversion or hours of manual work removed, and record the baseline. McKinsey's surveys measure value as EBIT impact, which is a useful discipline: if a tool improves individual productivity but no business metric moves, it has not yet paid off.
Because time saved by individuals is often absorbed rather than redeployed. McKinsey's 2026 survey found 80% of respondents reporting better individual productivity but only 37% seeing any EBIT impact. Profit moves when the workflow changes: fewer handoffs, less outsourcing, faster sales cycles. That requires redesigning the process, which is why high performers are far more likely to have done it.
Integration with real data and workflows. MuleSoft's 2025 benchmark found that 95% of organisations face challenges integrating AI into existing processes and 80% cite data integration as their biggest obstacle. MIT's researchers made a similar point: most generative AI systems do not retain feedback or adapt to context, so they stay stuck as pilots that never become part of daily work.