The value gap

Chapter oneThree numbers that describe the whole phenomenon

The cleanest picture comes from McKinsey's global survey: 1,719 executives across 97 countries, fielded between 4 May and 8 June 2026 and published on 25 August 2026.¹

Eighty percent of users report higher individual productivity. Thirty-seven percent of organizations attribute any EBIT impact to AI, and that number did not move year over year despite growing budgets. Six percent qualify as AI high performers: they attribute at least 5% of EBIT to AI and describe the impact as significant. That share is flat too.

These numbers do not say AI fails to work. They say it works at the level of the individual and stops before the level of the firm. Between the two sits a gap, and the gap is the work.

The louder figure - 95% of generative AI deployments with no measurable P&L impact, from the MIT NANDA report of August 2025 - makes the same point more forcefully on weaker evidence: 52 interviews, 153 survey responses, more than 300 initiative reviews, no peer review, and a definition of success narrowed to a return within roughly six months.² Treat it as a directional signal, not as proof. The McKinsey numbers are enough to carry the argument.

Chapter twoWhat happens to the recovered hour

Assume an employee saves one hour a day thanks to AI. That assumption has support. In a study of 5,179 customer support agents, access to a generative assistant raised productivity by 14% on average, with 34% for the least experienced staff and close to nothing for the best.³ In a field experiment with 758 BCG consultants, tasks inside the model's capability frontier were completed 12.2% more often and 25.1% faster.⁴

The question is what happens to that hour next.

The hour can be captured: converted into higher volume, shorter response times, lower unit cost, better quality, or people moved to higher-value work. It can also dissipate: absorbed as the task expands to fill the available time, spent by colleagues verifying and fixing output, converted into work nobody asked for, or simply invisible because nobody measured the starting point.

By default, the second happens.

This does not mean raising pressure or cutting headcount. Gartner estimates that around 80% of organizations report workforce reductions, and that those reductions do not translate into returns.⁵ Cutting cost without changing the process buys one-off budget relief, not durable advantage. A capture decision sounds like "we will handle 30% more cases with the same team" or "we will cut response time from 26 hours to 8". It only has to be made, written down and measured - that is the entire difference.

Chapter threeWhy "it feels faster" is not a measurement

METR ran a randomized controlled trial with 16 experienced open-source developers on 246 real tasks in their own repositories, where they averaged five years of prior experience.

Before the study, participants forecast a 24% speed-up. After doing the work, they judged themselves 20% faster. The measurement showed they were 19% slower.⁷ The distance between perception and fact was 39 percentage points - among competent, motivated people working on their own code.

This is one domain and a small sample. It should not be generalized to all office work, and the claim here is not that AI slows people down. The conclusion is methodological, and it is hard: an employee's self-report is not productivity data. If the only source of knowledge about AI's effect is a user survey, the company is not measuring the effect. It is measuring the mood.

The BCG consultant experiment showed the other side of the same coin. On a task deliberately placed beyond the model's reach, consultants using AI were 19 percentage points less likely to arrive at the correct answer - while producing answers that were more coherent and more persuasive.⁴ Confidence rising faster than accuracy is the most expensive effect a deployment can produce, because the cost surfaces at the customer.

Chapter fourThree archetypes of value capture

Companies differ less in which models they use than in which of three levels they stopped at.

Level one is the most common scenario: the company grants access to tools, sometimes runs a training session, and stops there. What is missing is usage measurement, knowledge of the real saving, a link between the tool and a specific process, and a mechanism for translating savings into KPIs. The cost is certain; the benefit is asserted.

Level two begins when AI stops being an employee's tool and becomes part of a process: the company measures handling time, volume, unit cost and quality before and after, and defines a new performance standard. This is where value shows up in metrics rather than in opinions.

Level three means changing the operating model: redesigning processes, changing roles, dividing work differently between people and agents. It is not mandatory and rarely a sensible starting point. But it is the level that correlates with financial results: among companies reporting material EBIT impact from AI, close to three-quarters have redesigned their workflows, against one-quarter of the rest.¹

The order matters, and it runs against instinct. Not "we will deploy AI and then see what can be improved", but "we will decide what the process should look like, and only then put AI into it".

Chapter fiveThe J-curve: why year one looks like failure

Brynjolfsson, Rock and Syverson described the pattern in which a new general purpose technology first depresses measured productivity and only later raises it as the productivity J-curve.

The mechanism is simple. Spending on technology is visible at once: the license invoice, the integration cost. Complementary spending - redesigning processes, new roles, cleaning up data, quality control, training - is incurred early, is several times larger, and is not booked as an asset. For the earlier wave of computerization, the authors estimated that the organizational capital accompanying an IT investment can be up to ten times the investment itself.⁸

The practical consequence: in the first period the measured effect is often negative, and that is normal rather than proof of failure. The second consequence is less comfortable. A company building organizational capital and a company that bought licenses and is waiting look identical at this point. Both show cost without effect. The only thing separating them is whether the complementary investment is being made at all.

Which is why the six-month checkpoint question is not "how much have we saved". It is: what exactly did we change in the workflow, and do we have a baseline to compute against.

Chapter sixThe programme: one process, four moves

There is no need to rebuild the company. There is a need to take one process all the way through, measurement included, and only then scale what held up.

Chapter sevenArithmetic you can put in front of a CFO

The starting model does not need to be complicated:

And this is where the calculation stops, because PLN 180,000 is so far a number on a slide. It turns into money only once the company decides what the 5,200 recovered hours are actually spent on. That is the eighth row of Table 1, and it is where most AI projects end without a result: not because the saving failed to materialize, but because nobody claimed it.

Chapter eightWho should act now

Not every company needs to rebuild its operating model today. Pressure rises fastest where a large share of the work is processing information at a computer: documents, customer queries, offers, data, content, repetitive actions in business systems. The smaller the physical component, the sooner AI moves cost, time and volume.

Pressure is also rising from the other direction - from costs and from disappointment. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value and inadequate risk controls.⁹ At the same time, one company in five already limits its use of AI because of operating costs.¹ The window in which "we are experimenting" justifies the spend is closing.

Companies that learn to measure and capture this value earlier will run faster and cheaper. Those that stop at buying licenses will eventually conclude that AI does not work. The problem will not be the technology. It will be the absence of a process, a metric, and a decision about what is supposed to change.

ConclusionStart with the process and the metric, not the license

Return on AI is not a function of model quality. It is a function of three decisions taken before deployment: which process, what baseline, and what happens to the recovered time. Companies that took those decisions are today in the six percent. Companies that did not have tools, costs and satisfied users - and a P&L in which none of it appears.

AI is not magic. It creates value when an organization knows where to apply it, how to measure the effect, and how to turn an individual's productivity into the performance of the whole.