Everyone’s using AI, yet almost nobody’s getting value from it. 🤷♂️
That sounds deliberately provocative. It is. But the data backs it up.
McKinsey’s 2025 global AI survey found that 88% of organisations now use AI in at least one business function. Meanwhile, EY’s 2025 Work Reimagined Survey of 15,000 employees across 29 countries shows that only 5% use AI in advanced ways that actually transform their work. The rest? Basic search. Document summaries. The digital equivalent of using a Formula 1 car to drive to the shops.
The AI training gap sits at the centre of this disconnect. Organisations keep buying tools and rolling out licences. They keep announcing AI strategies in quarterly earnings calls. What they don’t do is teach their people how to use any of it properly. And the research from every major consulting firm, academic institution and workforce survey tells the same story: that gap between adoption and results costs real money, real productivity and real competitive advantage.
The Adoption Illusion
Let’s start with the numbers that look impressive on paper. AI adoption has climbed from 55% in 2023 to 78% in 2024 to 88% in 2025, according to McKinsey’s surveys (previous link). Generative AI usage specifically went from 33% to 71% in a single year. BCG reported that 98% of companies are at least experimenting with AI. These are massive numbers. Headlines love them.
Now look at what happens when you ask about results.
McKinsey’s own data shows that only 39% of organisations report any measurable EBIT effect from AI. Most of those attribute less than 5% of their EBIT to it. A mere 6% of respondents qualify as “AI high performers” with meaningful bottom-line impact. Nearly two-thirds say their organisations have not yet begun scaling AI across the enterprise. They’re still experimenting. Still running pilots. Still testing AI in isolated pockets without deep integration into how work actually gets done.
BCG’s September 2025 research is even more stark. Only 5% of companies are achieving AI value at scale, while 60% report minimal revenue and cost gains despite substantial investment. Their October 2024 survey of 1,000 CxOs found that just 4% have developed AI capabilities that consistently generate significant value. That leaves 74% of companies showing no tangible return from AI. 😬
Here’s the bit that should make every CFO pay attention: It’s not a technology problem. These companies have the tools, the infrastructure, and the budgets. What they don’t have is people who know how to use AI strategically within their specific workflows. They bought the gym membership but never learned how to use the equipment.
I should know. I bought a gym membership, and before I invested in a personal trainer, I wasn’t getting any gains and was actually hurting myself. 🤦♂️
When people know what they’re doing, AI actually works
Before you dismiss AI as overhyped nonsense (tempting, I know), the controlled research on AI productivity gains is genuinely compelling. The problem is that these gains come with a massive asterisk that most vendors conveniently ignore.
The landmark study here is the Harvard/BCG “Jagged Frontier” experiment. Researchers tested 758 BCG consultants on 18 realistic consulting tasks. Those using GPT-4 completed 12.2% more tasks, finished 25.1% faster and produced over 40% higher quality output compared to the control group. The lowest-performing consultants saw a 43% improvement. That’s a significant skill-levelling effect.
But here’s the asterisk. On tasks that fell outside AI’s capability boundary, consultants who relied on AI without critical oversight were 19 percentage points less likely to produce correct answers than those working without AI at all. Same people. Same tool. Same organisation. The difference was whether they understood when and how to apply it. [READ THIS PARAGRAPH AGAIN] 👈
The researchers identified two archetypes of successful AI users. “Centaurs“ strategically divided labour between human and AI based on each one’s strengths. “Cyborgs“ deeply integrated AI into subtask-level workflows with constant interaction. Both dramatically outperformed a third group: those “sleeping at the wheel” who passively delegated to AI without thinking about it.
This pattern repeats across studies. EY’s 2025 survey found that employees receiving more than 81 hours of AI training per year report productivity gains averaging 14 hours per week. That’s nearly two full working days, every single week. Compared with a median of eight hours for the average employee. The difference between “using AI” and “using AI well” is enormous. And it comes down to training, not technology.
A quick reality check, though. The Harvard/BCG study tested elite consultants at one of the world’s top firms. These are people already in the top percentile of knowledge workers. Extrapolating those productivity gains to every role in every organisation would be dishonest. Your warehouse operations team and your mid-market sales department will see different results. The principle holds (trained users dramatically outperform untrained ones), but the magnitude of the gains depends heavily on the nature of the work and the starting baseline of the team.
The training gap in hard numbers
So if training makes this much difference, surely organisations are investing heavily in it?
They are not.
EY’s survey of 15,000 employees is probably the most damning data set on this topic. 88% of employees use AI at work. Only 12% receive sufficient AI training to unlock the full productivity benefits. That’s not a gap. That’s a chasm. The vast majority of AI users are self-teaching, guessing, or doing what everyone else does: typing vague prompts into ChatGPT and hoping for the best.
The consequences show up in multiple ways. Bright Horizons’ 2025 Education Index found that 42% of workers expect their role to change significantly due to AI, yet only 17% use it frequently. 34% feel unprepared for AI-driven changes. And 42% say their employer expects them to learn AI on their own. That’s the corporate equivalent of handing someone a scalpel and saying, “you’ll figure it out.”
The shadow AI problem makes this worse. EY reports that between 23% and 58% of employees (depending on the sector) are bringing their own unsanctioned AI tools to work, bypassing governance entirely. When organisations don’t train people, people train themselves. Often badly. With tools that haven’t been vetted for security, accuracy or compliance. Every organisation worried about data governance should be far more worried about the army of untrained employees copy-pasting sensitive data into random AI tools.
Perhaps the most telling statistic comes from Bright Horizons: AI adoption jumps from 25% to 76% when employers actually provide AI training. That’s a threefold increase from simply offering structured guidance. The demand is there, as is the willingness, but the support is not.
SHRM’s 2025 survey of learning and development executives confirms the pattern. While 53% of organisations say they prioritise employee upskilling, only 21% believe they’re doing it effectively. Employees themselves say the training they do receive is generic and misaligned with their actual roles. 39% want role-specific skills training. 33% find current programmes irrelevant, and 50% say they simply lack the time to participate. The training that exists isn’t working because it’s designed for a checkbox, not for capability building.
What the 5% do differently
The companies extracting real value from AI share specific characteristics, and the research is remarkably consistent on what those are.
BCG’s 10/20/70 rule has become the clearest framework for understanding this. AI success depends 10% on algorithms, 20% on technology and data, and 70% on people and processes. That’s not a marketing slogan. It’s derived from BCG’s analysis of their client transformations and their survey of 1,000 CxOs across 59 countries. The biggest roadblocks to AI value are people-related and process-related. The algorithms are the easy part.
AI success depends 10% on algorithms, 20% on technology and data, and 70% on people and processes
BCG’s Artificial Intelligence Capabilities
McKinsey’s 2025 analysis of 25 organisational attributes reinforces this. They found that workflow redesign has the single largest effect on EBIT impact from generative AI. Organisations reporting significant financial returns are twice as likely to have redesigned end-to-end workflows before selecting tools or models. Read that again. They redesigned how work gets done first. Then they picked the technology. Most organisations do the exact opposite: buy the tool, distribute licences, pray for results.
BCG’s “future-built” companies (the 5% succeeding at scale) deliver 1.7x more revenue growth and 3.6x higher total shareholder return compared to laggards. And here’s what they do that others don’t: they pursue half as many AI opportunities but achieve twice the ROI. Focus over breadth. They don’t try to “AI-ify” everything. They pick the highest-value use cases, invest deeply in those, train the relevant teams specifically for those workflows, and measure outcomes obsessively.
McKinsey identified 12 best practices distinguishing high performers. These include establishing dedicated teams, providing role-based capability training, senior leader role-modelling, embedding solutions into existing business processes, and tracking KPIs. Organisations adopting most of these practices were 3 times more likely to report an EBIT impact of 5%+ or more. 🤯
Also on that report, McKinsey includes a line that should be printed and stuck on every executive’s wall:
“Decisions that seem bold on paper (such as buying hundreds of gen AI tool licences for developers) are made without a clear understanding of the potential gains and with insufficient training of developers. The result: predictably poor outcomes.”
Predictably poor outcomes. Said by the world’s largest management consultancy.
That’s about as direct as it gets.
What this actually means for your organisation
I’m not going to give you a 47-step implementation roadmap. If you need one of those, your problem is bigger than AI training. But the research points to a handful of principles that consistently separate the organisations getting value from those burning money.
Start with the work, not the tool.
This sounds obvious, but… it isn’t, for most companies. 😅
Before you train anyone on anything, map the workflows where AI could genuinely change outcomes. Not “improve” them in a vague, hand-wavy sense. Change them measurably, like fewer hours on a specific task, or higher quality on a specific output, or even faster turnaround on a specific process. If you can’t define the outcome you want, you’re not ready to train anyone.
Be honest about where AI fits your actual operations.
The Harvard/BCG study showed that AI makes knowledge workers dramatically better at some tasks and measurably worse at others. Your organisation has both types of tasks. Identifying which is which requires domain expertise, not so much an AI vendor’s demo. This is where your experienced team members become essential. They know the nuances of the work, where judgment matters, and where pattern-matching is enough. Involve them early.
Make training role-specific and workflow-integrated.
Generic “intro to AI” courses produce generic results. The EY data is unambiguous: employees who receive deep, role-specific AI training (81+ hours annually) gain nearly double the productivity of those with basic training. That doesn’t mean everyone needs 81 hours. It means the training that works is training designed around the specific tasks people actually do, embedded into their daily workflows, not bolted on as an afterthought.
Measure outcomes instead of adoption.
Only 35% of organisations track ROI from AI. Virtually all high performers do. Stop counting how many people logged into the AI tool this month. Start measuring whether the work got better, faster, or both. If you can’t measure it, you can’t improve it. And if nobody’s measuring, nobody’s accountable.
Accept that fewer, deeper bets beat broad, shallow experimentation.
BCG’s highest-performing companies pursue half as many AI opportunities as their peers. They go deep on a small number of high-value use cases rather than sprinkling AI across fifty different departments with no clear priority. Pick three workflows where AI could make a genuine difference. Invest properly in those. Get the results. Then expand. The “let a thousand flowers bloom” approach is how you end up with a thousand dead flowers and a very expensive garden. 😉
The AI training gap is widening
There’s a compounding effect happening that makes the AI training gap more urgent than it might seem.
BCG’s research shows that future-built companies are reinvesting their AI gains into further AI development. They plan to spend 26% more on IT and dedicate 64% more of their IT budget to AI in 2025. That results in an overall AI investment 120% higher than their slower competitors. The leaders are pulling away, and the gap is accelerating.
Meanwhile, EY flagged a retention problem that creates a genuine dilemma. Employees who receive extensive AI training are 55% more likely to leave their organisation, because their skills become highly valuable in a competitive talent market. So organisations face a choice: under-invest in training and get no value from AI, or invest heavily and risk losing the people you’ve trained.
The answer, according to EY, is pairing deep training with retention strategies: clear career paths, total rewards packages and internal opportunities that compete with external offers. Easier said than done, but the alternative is worse.
The bottom line is straightforward.
The 88% of organisations “using AI” that see minimal results are not suffering from bad technology. They’re suffering from an AI training gap that nobody seems willing to close with real investment. The Harvard/BCG study demonstrated with experimental precision that the same AI tool in the same organisation produced a 40% quality increase for strategic users and a 19-point accuracy decrease for passive ones.
The tools are everywhere, but the training is nowhere to be found. And the 5% who figured this out are already compounding their advantage, while the other 95% keep buying more licences and wondering why nothing changed.
Close the AI training gap or watch the results gap widen.
Those are the two options.
Leave a Reply