NextFin News - Fabio Panetta has reframed the central-bank debate on artificial intelligence: the question is not how large the productivity shock will be, but who captures its gains. The European Central Bank Governing Council member argued on Monday that the distribution of AI-driven income will determine whether the technology prolongs inflation or accelerates disinflation, a warning delivered as the ECB holds interest rates at restrictive levels with inflation still well above its 2% target.
Panetta, who is also governor of the Bank of Italy, told an event hosted by the National Bank of Ukraine that central banks cannot remain on the sidelines. "Central banks need to understand who benefits from artificial intelligence-driven gains, as their distribution will help shape aggregate demand and inflation," he said. The remark cuts against the market's dominant one-way bet that AI is a purely disinflationary supply shock, and it puts a distributional question at the centre of monetary policy.
The Two-Way Inflation Channel
The transmission mechanism Panetta describes runs through demand, not supply. If AI mainly creates new tasks and raises expected labour income, households may spend more before the productivity benefits have fully materialised. Demand outruns supply, and inflationary pressures persist. Conversely, if automation dominates and displaces workers, weaker consumption could bring the disinflationary effects of AI forward.
"If AI mainly creates new tasks and raises expected labour income, demand could increase before AI's full productivity benefits materialise, prolonging inflationary pressures," Panetta said. "If automation dominates, weaker consumption could cause the disinflationary effects of AI to emerge sooner."
The distinction matters because it splits the AI-inflation question into two mutually exclusive outcomes that require opposite policy responses. A labour-augmenting AI that lifts expected incomes argues for a tighter-for-longer stance; a labour-displacing AI argues for earlier easing. Central banks that price only the disinflationary supply shock risk being caught on the wrong side of the demand response.
This is a cyclical demand mechanism, not a structural one. The demand channel operates over the near to medium term - as expectations about income adjust faster than measured productivity - and it is mean-reverting in the sense that once productivity catches up with expectations, the inflation impulse fades. The structural question is different and slower: whether AI transforms the innovation process itself and permanently raises potential growth. Panetta's warning is about the cyclical leg, and it is the leg that matters for the next policy cycle.
Why Distribution Is the Transmission Mechanism
The mechanism is straightforward but often overlooked. Productivity gains do not translate into aggregate demand automatically; they travel through the income distribution. Gains that accrue to capital owners carry a lower marginal propensity to consume than gains that accrue to labour. When AI raises profits faster than wages, the extra income tends to be saved or deployed in financial assets rather than spent on goods and services - dampening the demand-side inflation impulse even as supply capacity expands.
Recent empirical work supports Panetta's framing. An IMF working paper published in July 2026, using five waves of the Anthropic Economic Index covering more than 100 countries, found that AI's realised gains flow disproportionately to high-wage occupations. The paper values the time currently saved by AI at $2.7 trillion annually, or 3.4% of global GDP, but shows the concentration is far more extreme in developing economies - where virtually all AI usage-based value is generated within a small professional enclave - than in high-income economies, where gains are spreading into education, sales, and office work.
That concentration gradient has direct inflation implications. If AI adoption in advanced economies broadens beyond its software-engineering core and begins to raise expected incomes across a wider share of the workforce, the demand channel Panetta warns about strengthens. If adoption remains confined to high earners, the disinflationary channel dominates. The concentration metric is declining in many high-income countries, which is precisely the direction that keeps the inflation risk alive.
The second-order implication is what the market has not fully priced. Investors have largely treated AI as a deflationary technology shock: more output per worker, lower unit costs, lower prices, lower rates. Panetta's point is that the first-order supply effect can be overwhelmed by the second-order demand effect if income expectations adjust first. The sequence matters more than the size: expected income rising before measured productivity is the exact condition under which a productivity story becomes an inflation story.
The euro area enters this debate with a high profit share and a labour share that never fully recovered its pre-crisis level. That starting point tilts the near-term distribution of AI gains toward capital, which is why the disinflationary channel dominates today. But the same high profit share means firms have the margin capacity to fund rapid AI deployment - and deployment is what broadens adoption. The mechanism is self-undermining: the capital-biased distribution that suppresses demand today also finances the diffusion that will spread income gains tomorrow. Panetta's warning is about that transition, not about the current snapshot.
The Productivity Debate Is Not Settled
Estimates of AI's macroeconomic impact range from the modest to the transformative, and the dispersion itself is a policy problem. Goldman Sachs Research projected in March 2023 that widespread AI adoption could drive a 7% increase in global GDP over a decade, lifting annual labour productivity growth by around 1.5 percentage points. A June 2023 McKinsey study suggested AI combined with broader automation could add as much as 3.4 percentage points per year to productivity growth through 2040.
At the lower end, Acemoglu (2025) concludes that aggregate total factor productivity gains over the next ten years are unlikely to exceed 0.66% in total. An OECD study published in 2025 projects AI could add between 0.4 and 1.3 percentage points to annual aggregate labour productivity growth over the next decade in countries with high AI exposure. Aghion and Brunel (2024) generate a median estimate of 0.68 percentage point of additional annual TFP growth, while Bergaud (2024) suggests an annual productivity boost for the euro area of 0.29%.
Early microeconomic evidence is encouraging but uneven. One experiment found access to ChatGPT reduced the time taken on mid-level professional writing tasks by 40% and raised output quality by 18%, with the largest gains accruing to lower-ability workers. A separate study of a generative AI assistant rolled out across more than 5,000 customer support agents found a 15% increase in issues resolved per hour. These micro findings are suggestive, but their macroeconomic significance remains uncertain because not all sectors have the same scope for AI-related process improvements.
The historical analogy is the productivity J-curve described by Brynjolfsson, Rock, and Syverson: general-purpose technologies require complementary investments and organisational co-invention before measured productivity rises, so the aggregate signal lags the investment cycle by years. Electricity and computers both took decades to show up in the productivity data. That lag is exactly the window in which Panetta's demand channel can operate - expectations move on the promise, prices move on the spending, and the productivity that would justify both arrives later, if at all.
The Policy Backdrop
Panetta's remarks land against a restrictive ECB stance. On September 10, the Governing Council raised its three key interest rates by 25 basis points, citing continued inflation pressure from the conflict in the Middle East and inflation expected to remain well above target for an extended period. The deposit facility rate moved to 2.50%, the main refinancing rate to 2.65%, and the marginal lending rate to 2.90%, effective September 16.
The ECB's new staff projections see headline inflation averaging 3.0% in 2026, 2.5% in 2027, and 2.1% in 2028 - still above the 2% target through 2027. Core inflation excluding energy and food is projected at 2.5% in 2026, 2.6% in 2027, and 2.3% in 2028. Growth is expected at 0.9% in 2026, 1.4% in 2027, and 1.5% in 2028. The Governing Council said it would follow a data-dependent, meeting-by-meeting approach and is not pre-committing to a particular rate path.
Wage data give the demand channel a concrete current reading. Euro area hourly labour costs rose 3% year-on-year in the second quarter of 2026, the slowest pace since the fourth quarter of 2021, while negotiated wage growth stood at 2.44% in the second quarter, down from 2.56% in the first. The ECB's wage tracker pointed to stable negotiated wage growth of 2.6% in 2026. For now, wages are moderating - which supports the disinflationary case. But Panetta's argument is forward-looking: the risk is not today's wage print, it is the moment expectations detach from current productivity.
Against that backdrop, Panetta's point is that AI distribution is itself a data point central banks must monitor. Understanding who benefits is no longer an academic question but a credibility issue: "Central banks cannot remain on the sidelines," he said, adding that understanding the changes underway is increasingly essential to central bank credibility.
The Strongest Counter-Thesis
The bear case against Panetta's warning is that it overcomplicates a simple supply story. The mainstream view, reflected in the ECB's own research and in market pricing, is that AI is a positive supply shock that raises potential output and lowers the natural rate of interest. Under that view, distribution is a second-order detail: whether gains accrue to labour or capital, the economy's capacity to produce rises, and the neutral rate falls. Central banks that overreact to hypothetical demand impulses risk tightening into a supply-driven disinflation and unnecessarily damaging growth.
There is force in that argument. The euro area's current wage data show moderation, not acceleration, and inflation is projected to drift back toward target without any AI-driven demand surge. The concentration evidence from the IMF paper shows AI gains remain heavily tilted toward high-wage occupations - exactly the group with the lowest marginal propensity to consume. On today's data, the disinflationary channel dominates.
But the counter-thesis rests on a timing assumption that may not hold. The supply effect requires adoption, diffusion, and complementary investment - the J-curve lag. The demand effect requires only expectations. If financial markets, firms, and households begin pricing AI income gains before the productivity arrives, the demand impulse leads the supply response. Panetta is not arguing that the disinflationary channel is wrong; he is arguing that it is not guaranteed, and that the guarantee the market has priced depends on a distributional outcome - gains staying concentrated at the top - that is already showing signs of broadening.
The falsifying signal is specific and observable. If euro area wage growth stays above 3% year-on-year while labour productivity remains near the 2025 pace of 1.4% and unit labour costs rise for two consecutive quarters, the disinflationary-supply-shock consensus breaks down and Panetta's demand channel is active. Conversely, if productivity accelerates toward the OECD's upper estimate of 1.3 percentage points of annual labour productivity growth while wage growth stays contained, the market's one-way disinflationary pricing is vindicated and the distribution question recedes.
What to Watch
The outlook splits by time horizon. In the short term, the inflation impulse depends on whether expected labour income rises before measured productivity does - a gap that shows up first in confidence surveys, hiring intentions, and wage negotiations in AI-exposed sectors. In the medium term, the breadth of AI adoption across occupations, tracked by concentration metrics such as the IMF's AI concentration index, will determine the demand channel's strength. In the long term, if AI transforms the innovation process itself rather than merely automating existing tasks, potential growth rises structurally and the inflation trade-off changes altogether.
Three scenarios frame the path. The base case is gradual adoption with gains remaining concentrated enough that the disinflationary supply effect dominates, but with enough income-expectation lift to keep services inflation sticky and delay the ECB's easing cycle. The upside inflation case is rapid, broad-based adoption that lifts expected incomes across the workforce before productivity arrives - the exact condition Panetta flags. The downside case is automation-dominated adoption that compresses employment and consumption, bringing disinflation sooner and forcing earlier rate cuts.
For investors and policymakers, the practical implication is that AI is not a single-factor trade. The productivity estimates - anywhere from Acemoglu's 0.66% cumulative TFP gain to McKinsey's 3.4 percentage points of annual productivity growth - span a range wide enough to accommodate any outcome. What will decide the inflation path is not the size of the pie but who eats it. Panetta's judgment compresses to a single line: the market is pricing AI as a supply shock, but the inflation outcome will be decided by the distribution of its gains. Central banks that watch only productivity will miss the demand channel that actually sets prices.
The policy lesson is asymmetrical. If Panetta is wrong and AI is a pure supply shock, central banks that hesitate to ease will err on the side of price stability - a forgivable mistake for institutions with a 2% mandate. If he is right and demand outruns supply, central banks that dismiss the distribution question will be behind the curve on inflation. For a central bank that has just raised rates into a war-driven inflation spike, the cost of the second error is the higher one.
Explore more exclusive insights at nextfin.ai.
