The Four Percent Promise

On March 23, 2026, the ECB’s chief economist stood up in Frankfurt and quantified the prize. Over a decade, artificial intelligence could lift euro area productivity by more than four percentage points. That is not a marginal gain. Four percentage points compounded over ten years is the difference between a stagnating bloc and a growing one. It is the size of the gap that Mario Draghi spent a year documenting in his competitiveness report.

The prize comes with a condition. The four-point figure assumes adoption reaches at least half the economy. Today, adoption among EU enterprises with ten or more employees is twenty percent. For small firms, it is seventeen.

The ECB is not selling optimism. It is publishing a conditional. The condition is adoption. Adoption is architecture, not aspiration.

The Speech

The keynote was delivered at the ECB-SAFE-RCEA International Conference on the Climate-Macro-Finance Interface (3CMFI) in Frankfurt. The speaker was Philip R. Lane, member of the ECB Executive Board and the bank’s chief economist. The title — “AI and the euro area economy” — is on the ECB website. The text is public. Read it, not the commentary about it.

Lane structured the speech around three questions. What is AI doing to the macroeconomy now? What might it do over the next decade? And what would have to be true for Europe to capture the gains? The first answer is: not very much yet. The aggregate effects of AI on productivity, employment, and inflation remain limited and uncertain. The diffusion is rapid. The investment is rising. The measured macroeconomic impact, so far, is small.

The second answer is where the four percent appears — and where it must be read with care.

Lane reviewed external forecasts. Goldman Sachs Research, in March 2023, projected that widespread AI adoption could raise annual labour productivity growth by around 1.5 percentage points over a decade. McKinsey, in June 2023, suggested that AI combined with broader workplace automation could add as much as 3.4 percentage points per year through 2040. These are not Lane’s forecasts. They are the range of plausible scenarios that Lane was working from.

Then Reuters did the synthesis. In coverage published the same day, the wire service summarised Lane’s framing: a take-up rate in line with previous innovations such as the internet would deliver at least 1.5 percentage points of additional productivity growth over ten years; if adoption continued at today’s pace and reached at least half the economy, the gain could exceed four percentage points. That summary is the source of the “four percent promise” headline. It is reporter paraphrase, not a Lane direct quote. The substance is in the speech. The synthesis is in the reporting.

The direct quote — the one Lane actually delivered — is this: the greatest impact will be achieved if AI materially boosts the pace of innovation, as rather than just boosting the level of productivity, this could increase the long-run potential growth rate. That sentence is the speech in miniature. The four-point figure is conditional on innovation acceleration, not on tool adoption alone.

The Denominator

The four-point scenario requires AI to reach at least half the economy. The current state is documented.

Eurostat published its 2025 figures on December 11, 2025. Twenty percent of EU enterprises with ten or more employees used AI technologies that year, up from 13.5 percent in 2024. The growth — 6.5 percentage points in a single year — is the fastest rate the survey has ever recorded. The level is still twenty percent.

Disaggregate by firm size and the structure becomes visible. Seventeen percent of small enterprises (10–49 employees) used AI. Thirty point four percent of medium enterprises (50–249 employees). Fifty-five percent of large enterprises (250+ employees). The gap between small and large is thirty-eight percentage points. The denominator that Lane’s scenario requires — half the economy — is not present in any size class except large firms, and even there it is barely there.

Geography compounds the structure. Denmark leads at forty-two percent. Finland is at 37.8. Sweden at 35. At the other end, Romania reports 5.2 percent. Poland 8.4. Bulgaria 8.5. The euro area average sits between these poles. The countries that will pull the four-point scenario into reality are not the countries where the gap is widest. They are the countries where the gap is already smallest.

This is the denominator problem. The ECB’s four-point figure is a function of a fraction. The numerator is plausible — AI does generate productivity gains in the firms that use it well. The denominator is the share of the economy where those firms exist. Today, the denominator is one in five. The promise requires one in two.

What the SAFE Survey Found

The other primary source is the ECB’s own Survey on the Access to Finance of Enterprises (SAFE). The fourth-quarter 2025 wave included ad-hoc questions on AI adoption. Five thousand firms across the euro area responded. The results were published in the Economic Bulletin focus piece in February 2026.

The headline number sounds encouraging. Two-thirds of respondent firms reported that their employees use AI. Read further and the structure inverts. Twenty-seven percent of firms do not use AI at all. Thirty-three percent use it very infrequently or experimentally. Thirty-one percent use it moderately. Seven percent use it significantly.

Seven percent. That is the share of euro area firms where AI is genuinely embedded in operations. Not opened occasionally. Not piloted in one department. Significantly used. The figure is the same for SMEs and large firms — which is the only piece of good news in the data. When firms commit to AI, size does not determine depth of use.

But the commitment is rare. Among non-users, thirty percent cite lack of usefulness as the primary barrier. Twenty percent cite system incompatibility. Twenty percent cite a skills shortage. The barriers are not curiosity or budget. They are deployment methodology, data readiness, and the inability to identify where the tool earns its keep. These are architectural problems. They are not solved by buying licences.

Investment intent follows the same pattern. Firms expect to allocate nine percent of total investment to AI on average. Non-users plan four percent. Moderate users plan eleven. Significant users plan twenty. The most-committed firms invest five times the rate of the least-committed firms. The gap is widening, not closing.

The SAFE data also segments by firm characteristics. Forty-five percent of large firms and listed or venture-backed companies are at an advanced stage of adoption — significant or moderate use. The figure rises to fifty-six percent for young firms. Founded-since-2020 companies adopt AI at nearly twice the rate of established mid-cap firms in the same sector. The pattern is consistent across industries: the firms that are most operationally flexible adopt fastest, and the firms with the most institutional weight to move adopt slowest.

The most-cited barrier among non-users is not technical. Thirty percent report lack of usefulness. That is not a complaint about AI. It is a complaint about deployment. A firm that cannot identify where AI earns its keep is not a firm with a tools problem. It is a firm without an architecture problem statement. The tool is downstream. The problem statement is the bottleneck.

This is what “widespread but infrequent” looks like in the numbers. AI is everywhere. Adoption is somewhere else.

The Worker Layer

Lane added a third data point that is easy to misread. The ECB’s Consumer Expectations Survey shows that the share of employed euro area workers using AI rose from twenty-six percent in 2024 to forty percent in 2025. The uptake, Lane noted, is outpacing the historical diffusion of the internet or personal computers.

Forty percent of workers are using AI tools. Seven percent of firms are using AI significantly. These two numbers describe the same economy.

The implication is direct. Workers are adopting AI faster than their employers are deploying it. The tools are in the browser. The integration is not in the workflow. An employee using ChatGPT to summarise a meeting is using AI. A company whose operations have been redesigned around AI capabilities is deploying AI. The first generates a productivity benefit that is captured by the individual and largely invisible to the firm. The second generates a productivity gain that shows up in output, costs, and competitive position.

The forty-percent worker figure is what the ECB calls “rapid diffusion.” The seven-percent firm figure is what an economist would call adoption. Diffusion without adoption produces noise — measurable activity that does not translate into measurable productivity. The macroeconomic data is consistent. The aggregate effects of AI remain, in Lane’s words, limited and uncertain.

The Energy Risk

Lane added a warning that most coverage skipped. The optimistic scenario assumes that the infrastructure required to scale AI — data centres, compute, electricity — is available. It may not be.

The speech notes that further acceleration of AI-related and digital investment could be hampered by insufficient energy supply, shortages of skilled staff, and overregulation. The energy constraint is the most material of the three. AI workloads are energy-intensive. Persistently high fuel costs raise the marginal cost of training new models and running inference at scale. They also reduce the rate at which firms can adopt — every euro spent on power is a euro not spent on integration.

This is the asymmetric risk in the four-point scenario. The upside requires adoption to double from twenty percent to fifty. The downside requires only one variable to move against the trend — energy prices, a hyperscaler reaching capacity in one region, a regulatory tightening on data centre permits — and the optimistic scenario compresses toward the baseline. The 1.5-point trajectory is more robust than the 4-point trajectory. The latter is a tail outcome that requires several things to go right simultaneously.

What This Means for a 200-Person Manufacturer

Translate the macro into the operational. A 200-person manufacturer in Portugal, Germany, or the Netherlands is, statistically, in the medium-enterprise bracket. That bracket reports thirty percent AI adoption. Most peers are not using AI. A meaningful minority are.

If the four-point scenario materialises, the manufacturer will face competitors operating with measurably higher productivity within a decade. Not in a vague “AI will change everything” sense. In the specific sense that a competitor with AI embedded in production planning, quality control, customer service, and design iteration produces more output per worker than a competitor without. The productivity gap will be visible in margins, in pricing, in lead times, and ultimately in market share.

If the 1.5-point scenario materialises — the more likely outcome given current adoption rates — the gap is smaller but it still favours the firms that adopted early. The cumulative effect of a 1.5-point annual productivity differential over ten years is approximately a sixteen percent total productivity advantage. For a manufacturer competing on margin, that is not a rounding error. It is the difference between sustainable operations and structural decline.

In either scenario, the firms that capture the gain are the ones that moved from worker-level diffusion to firm-level adoption. Letting employees use ChatGPT individually is not a strategy. It is what happens by default. Strategy is the specific decision about which workflows AI will be embedded in, which data feeds it requires, who maintains it, how its output is checked, and how its use is measured.

A 200-person manufacturer that wants to be on the productive side of the four-point promise needs to make six decisions in the next eighteen months. Which two operational workflows benefit most from AI integration. Which data sources those workflows require and whether they exist in usable form. Who in the organisation is responsible for the integration — not for evaluating it, for owning it. How the output of the AI system is validated before it affects a customer or a financial outcome. How adoption is measured — daily use, weekly use, what kind of use. And how the system is updated as the underlying models change.

None of these decisions require a PhD in machine learning. All of them require someone in the company to own them. The barrier is not technology. It is allocation of accountability.

The Eurostat country breakdown sharpens the point. A manufacturer competing in Danish or Finnish markets faces a peer set where forty percent of firms are already running AI. A manufacturer in Romanian or Polish markets faces a peer set where five-to-eight percent are. The competitive landscape is not uniform across the single market. The same product sold into two member states encounters two different productivity baselines from the firms next door. Build for the Danish baseline and the Romanian market remains accessible. Build for the Romanian baseline and the Danish market closes off — quietly, through margin pressure, before anyone announces it.

The trap, in either market, is to mistake worker-level diffusion for firm-level capability. A factory floor where the planning team uses ChatGPT to draft emails is not a factory floor with AI in production. A maintenance team that occasionally asks a chatbot to interpret a sensor log is not a maintenance team running predictive maintenance. The Consumer Expectations Survey will register both of these as AI use. The productivity data will not. The Eurostat enterprise figure is the harder measure, and it is the one that maps to the four-point scenario. Twenty percent. Not forty.

Where Bluewaves Comes In

Bluewaves builds in three-week waves. Each engagement produces a Gizmo — a specific AI tool deployed in a specific workflow, owned by a specific team. We do not run pilots. We do not produce strategy decks. We ship working AI into daily use.

The choice is deliberate. The ECB data tells you why. Seven percent of euro area firms use AI significantly. The other ninety-three do not. The gap is not awareness. It is the institutional gap between knowing about a technology and embedding it in operations. Closing that gap is not a matter of more presentations. It is a matter of building one specific thing, getting it adopted by the team that will use it daily, and measuring its use as the proof.

A 200-person manufacturer that finishes its first Wave has one Gizmo in daily use by one team. That is more than ninety-three percent of euro area firms have achieved.

The Position

The four-point promise is real. The conditional is also real. The ECB published both. Lane’s speech is honest about the asymmetry: the upside is large, the conditions are demanding, and the current denominator is one-fifth of what the optimistic scenario requires.

European productivity will rise if European firms deploy AI. It will not rise if European workers individually use AI tools while European firms continue to debate adoption strategies. The first state is widespread. The second state is rare. Lane published the gap between them. The gap is the article.

A 1.5-point productivity lift over a decade is achievable on current trajectories. It is also insufficient to close the EU-US per-capita GDP gap that Draghi estimated at seventy percent attributable to lower productivity. The four-point lift would close meaningful ground. It requires adoption to reach half the economy. Half the economy means small firms adopting at three times the current rate. Small firms adopting at three times the current rate means small firms doing something they have not yet done: embedding AI in operations, not just opening it in a browser.

The architecture is not optional. It is the variable. The ECB measured it. The number is twenty percent. The promise is four. The distance between them is the work.

The work has not started.