Playtime Is Over: The AI ROI Paradox HR Cannot Avoid

The AI investment cycle is producing a lot of activity, very little verifiable productivity, and almost no honest accounting of where the value is supposed to come from.

The math: Adoption × Augmentation × Alignment

At People Analytics World in London, Dana Minbaeva (King’s Collage) opened with MIT’s recent research that found that despite the billions invested in generative AI globally, roughly 95% of organisations report no measurable return. Only about 5% of integrated AI pilots are extracting meaningful P&L impact. EY’s data tells a parallel story: nearly 9 in 10 employees now use AI at work, yet only 28% of organisations are positioned to turn that deployment into high-value outcomes.

They describe the same gap.

Dana Minbaeva proposed a formula that I truly love:

AI ROI = Adoption × Augmentation × Alignment

It is multiplicative, not additive. If any one of the three is zero, the whole thing is zero. That is the entire diagnosis in one line.

Adoption is straightforward — are people using the tools? This is the easy variable. The technology acceptance literature has decades of evidence that adoption depends on perceived usefulness and perceived ease of use, which together produce attitude, intention, and behaviour. Most companies measure this, more or less.

Augmentation is the hard one, and the one almost nobody is measuring. Dana framed it as a 2×2: the decision (human-led or technology-led) on one axis, and the content (human-created or technology-created) on the other. Pure human decision plus human content is business as usual. Pure tech decision plus tech content is full automation. The interesting middle is guided augmentation (human decides, tech generates) and preconditioned augmentation (tech decides, human creates). Most current AI use is sitting on the BAU edge of that matrix, not the augmentation centre.

Alignment is whether AI initiatives connect to actual business strategy and value pools, or whether they are pilots that were sold to the CTO, not commissioned by the business with a real problem in mind. This is, in my experience, the variable that fails most often and most quietly. Pilots are often picked because a vendor pitched well, not because the company had a defined problem AI was uniquely placed to solve.

Multiply Adoption by zero Augmentation by zero Alignment and you get the 95% number. Multiply all three by something meaningful and you get the 5%.

But beyond this overall ROI, there is another big question: AI is sold with a huge productivity promise, but where is this?

The leak: where does the productivity actually go?

If task-level AI gains are real — and the evidence suggests they are, with documented improvements of 10–70% on well-defined, text-intensive tasks — then where does that productivity disappear at the enterprise and macro level?

This was Ekkehard Ernst’s (ILO) question, and the framing he offered from the ILO is the single best macroeconomic explanation I have heard. He calls it the aggregation paradox: massive task-level gains shrinking, leaking, or evaporating entirely as you scale up.

Five leaks specifically:

  1. Diffusion is restricted. Effective access to AI is stuck at around 7% of firms globally. The gap between large enterprises and SMEs is enormous: 45% of large enterprises have dedicated AI policies versus 14% of SMEs. 65% of large enterprises provide AI training; only 29% of SMEs do. Most strikingly, 26% of large enterprises can verify AI gains. Zero percent of SMEs can.
  2. Within-firm scaling is expensive. The fixed costs of system-level redesign are high. AI does not become valuable when bolted onto existing workflows; it becomes valuable when workflows are re-engineered around it (AUGMENTATION). That re-engineering requires investment most firms have not made.
  3. Reallocation is sticky. Moving labour and capital to the productive frontier — the firms actually generating returns from AI — has friction. People do not move easily; capital moves more easily, but unevenly.
  4. Task restructuring is outpaced by automation. New productive roles are not being created at the speed at which existing tasks are being automated. This is the classic technology-adoption mismatch, and it has happened with every general-purpose technology.
  5. Price effects shift expenditure. AI lowers costs in specific sectors, which shifts spending toward less AI-affected, lower-productivity services — Baumol’s cost disease in 2026.

The point Ekkehard Ernst made most forcefully — and the one that should make every CFO uncomfortable — is the measurement gap: up to 74% of perceived AI productivity improvements are unverified. Companies are committing massive resources to AI without the measurement discipline required to move from pilots to scalable performance. Where verification exists, the gains are real and large: 60–98% reductions in task time on targeted processes. But verification is rare.

If you cannot verify it, you cannot bank it. And if you cannot bank it, you do not have ROI. You have a feeling that things are faster.

The 70/20/10 problem: faster, not smarter

There is a classic innovation portfolio framework — the 70/20/10 rule — that has been used for years to think about where companies should invest to balance core operations against transformation. Roughly 70% of investment goes to known problems with known solutions (low risk, low time). About 20% goes to known problems with unknown solutions (medium risk, longer time). And 10% goes to unknown problems with unknown solutions (high risk, longest horizon).

I have been carrying that diagram around mentally since PAW because it explains exactly what is happening with AI right now.

Most current AI deployment is sitting in the 70% box. Known problem (we already know what an email is, what a meeting summary is, what a piece of marketing copy is, what a code review looks like), known solution (use a generative tool to do it faster). This is what @Alexis Saussinan (Merck) called “Everyday AI” and what most of the conference quietly admitted is dominating the actual usage statistics.

The problem is that the 70% box is also where the value is hardest to measure and easiest to leak. Writing emails faster, producing decks faster, generating boilerplate code faster — these are real efficiency gains at the individual level. They almost never aggregate into a P&L line. They show up as “I have more time” and disappear into more meetings, more emails, and more decks. Time saved is not money saved unless someone removes the headcount or redeploys it. Most companies do neither.

The real ROI lives in the 20% and 10% boxes. Augmentation, in Dana Minbaeva’s sense, is the 20% box: known problem, new solution. Take a process you understand and rebuild it around AI from the ground up — not faster execution, but a different operating logic. The 10% box is innovation: new business models, new revenue streams, things you could not do before AI existed.

Alexis Saussinan structured this elegantly into three layers at Merck:

  • Everyday AI — workforce productivity. The 70% box. Real, but small.
  • Operational AI — teamwork and jobs redesign at scale. The 20% box. Real and larger.
  • Advanced AI — competitive advantage. The 10% box. Real and largest, also rare.

His point was that playtime is over. Companies have spent two years experimenting. The next phase has to focus AI investment on measurable P&L impact, with the human kept in the lead, while visibly improving everyday productivity. “No money is printed. You have to prioritise P&L choices and redeploy funds for AI investment.”

AI is not a free addition to the existing budget. Every euro spent on AI infrastructure is a euro not spent on something else. Which means the conversation is not “should we invest in AI” — it is “what are we not doing in order to invest in AI, and what return are we going to demand for that trade?”

AmCham’s AI Readiness Survey

AmCham Hungary ‘s AI Readiness and Maturity Survey, which I co-led this year, found exactly the same pattern at the local level.

The most striking finding is the relationship between AI revenue share and ROI. Among companies where AI contributes 5–20% of revenue, 67% report high, company-wide ROI. Among those at 0% AI revenue, only 8% do. The correlation is direct and unambiguous: the closer AI is to revenue generation, the more measurable its success becomes.

Conversely, organisations using AI primarily for internal efficiency see slower or less visible returns — not necessarily because the impact is smaller, but because efficiency gains rarely translate into accounting cost reductions. Time saved, accuracy improved, decisions made faster — these are tangible benefits that do not appear on a P&L unless someone makes a downstream choice to capture them.

And now the paradox HR cannot avoid

If AI ROI in most companies is going to come primarily from cost reduction — and the AmCham data, the McKinsey data, and the conference evidence all suggest that is where measurable value is showing up first — then by definition that ROI has to come from somewhere.

It cannot come from “we did the same work with the same headcount but it felt easier.” That is not ROI. That is sentiment.

It has to come from one of two places: revenue created that did not exist before, or labour cost removed that did exist before.

The first is the harder, slower, more strategic path. The second is the faster, easier, more measurable one — and it is, by definition, a reduction in force.

I am not saying we are there yet. The AmCham survey actually found that a majority of respondents expect AI to reshape roles, roughly one third expect new roles, and only a few anticipate notable job reductions. That is consistent with global findings. The PwC AI Jobs Barometer, the WEF Future of Jobs Report, and most credible academic work agree that net employment effects can be neutral or positive when organisations invest seriously in reskilling and task reallocation.

When (!!) organisations invest seriously in reskilling and task reallocation. That is a big “when.” And the survey data tells us that 76% of employees want it, only 17% of companies are funding it, and the dominant reported barrier across every study is the skills gap.

Meanwhile, the visible behaviour in the technology sector — the biggest, most aggressive AI adopters — is not “reshape and reskill.” It is large-scale workforce reduction explicitly framed as freeing up capital for AI investment. This is not AI driving ROI, this is RIF driving AI.

So here is the paradox in one sentence: if your AI strategy is funded primarily through cost reduction, and your communication is “AI is not taking your jobs,” at least one of those two things is not true.

I am not saying AI must take jobs. I am saying that the two narratives we are running in parallel — “AI ROI comes from efficiency” and “AI will not affect employment” — are mathematically in tension.

What this means for HR and people analytics

Dana Minbaeva made a claim in her session that I think is the single most important strategic positioning argument the people analytics field has produced in the last five years:

People Analytics is the function uniquely positioned to drive AI strategy — not IT (which is infrastructure-led), not the data team (which is model-led), not the business units (which are use-case-led). PA can measure how value actually emerges from the interaction of people, tasks, and AI. PA can see the verification gap close. PA can attribute the productivity that is currently leaking.

If she is right — and I think she is very much right — then HR has a specific job in the AI ROI conversation, and it is not running prompt engineering courses. It is owning the verification problem. It is being the function that can sit across from the CFO and say: here is what AI is actually delivering, here is where the gains are real, here is where they are noise, here is what we should fund, and here is what we should kill.

That requires a profession that can speak the language of effect sizes and confidence intervals, not engagement scores. It requires people analytics teams that report to the CHRO with the seriousness of FP&A teams reporting to the CFO. It requires the willingness to measure manager effectiveness, hiring quality, and intervention ROI with the same rigour we expect of marketing attribution or revenue forecasting.

It also requires HR to own the conversation of redesigning processes and jobs on the task level (aka augmentation) and everything around it — skill profiles, performance management, compensation, headcount, career paths, organisational models. When we move from Everyday AI to Operational and Advanced, the game is not about writing the same email faster.

It also requires HR to be honest about the workforce side of the AI ROI equation. If a company is building its AI investment case on cost reduction, HR’s job is to surface what that means for the workforce — explicitly, in advance, with a plan — rather than discovering it during the eventual restructuring. Reskilling at scale, redeploying at scale, exiting at scale: these are HR competencies, and they are the difference between an AI-driven workforce transformation that is strategic and one that is reactive.

The companies that win the next decade on AI will be the ones where HR sat in those conversations early, with the data, with the rigour, and with the willingness to be uncomfortable. The companies that lose will be the ones where HR was told about the AI strategy after Finance had already locked the cost case. And as I wrote in my previous article: do not wait until you are invited to the table, take the lead.

Choose AI use cases on business value, with a documented decision trail.

Explore pAIr →Explore gAIn →