ECB Experiment Finds AI Investment Intentions Follow Domestic Peers

An AI investment plan can change in a single survey interview. A purchase order cannot. That distinction is central to a new ECB working paper on how firms respond to information about their peers. Its randomized experiment found that firms given accurate figures on peer AI investment planned to allocate more of their own investment to AI. The sharper finding is geographical: revisions to beliefs about domestic peers tracked firms' intended allocations much more strongly than revisions to beliefs about a fixed three-country benchmark.

The result is a test of what information can move inside Europe's single market. It is not a count of AI systems installed, a productivity gain, or evidence that firms disregard all foreign competition. It shows how a particular signal changed the plans of surveyed firms at the moment they received it.

One signal, two peer groups

The researchers embedded the experiment in the fourth-quarter 2025 Survey on the Access to Finance of Enterprises. Their analysis uses 3,316 firms in 12 EU countries. Firms first estimated the share of comparable domestic businesses that had invested in AI by June 2025. They also estimated the share among comparable firms in Germany, France and Italy. Comparable meant the same sector and size class, with country defining the domestic group. The same three-country benchmark applied to every respondent, so for firms based in Germany, France or Italy it also included their home country. Earlier survey data supplied the actual shares used in the experiment.

Half the respondents then received accurate figures for both groups; the other half received no new information. The researchers asked how many peers firms expected to invest in AI over the next year and what share of their own total investment they expected to allocate to it. The treatment therefore tested an information shock about past peer investment, followed by beliefs and plans for the future. It did not assign firms a subsidy or observe their later purchases.

Treated firms reported an intended AI allocation of 10.13%, compared with 8.33% among controls. The 1.8 percentage-point difference is statistically significant in the paper. It is also the cleanest causal result: the randomly provided bundle of domestic and three-country information changed reported intentions. It cannot by itself say which part of the bundle mattered most.

The paper's second step uses the randomized information gaps to estimate how changed beliefs relate to intended spending. A one-point rise in the expected share of domestic peers investing in AI is associated with a 0.570-point rise in a firm's intended AI allocation. The corresponding estimate for the three-country benchmark is minus 0.224 points and statistically insignificant. That is evidence of a stronger domestic response in this design, not a proof that the benchmark effect is exactly zero. The belief estimate gives greater weight to firms whose views moved after receiving the signal; it is not an average response to every possible change in peer behavior.

The three-country signal moved beliefs without clearly moving budgets

Respondents did not simply ignore the three-country figures. The paper reports that the information raised expectations about future peer AI investment in both the domestic group and the three-country benchmark. Respondents appear to have absorbed the signal; the split emerges when the researchers connect those revised expectations to the firm's own intended allocation.

One reading is competitive proximity. A local rival may sell to the same customers and bid for the same staff. Its AI spending can therefore look like an immediate threat. Another is learning: a firm may trust a domestic peer's experience as more relevant to its own wages, regulation and customer demand. The experiment cannot cleanly separate these channels. The authors' subgroup analysis offers clues, but several subgroup estimates have weak-instrument cautions and cannot carry a sweeping account of European competition.

The benchmark is not a roster of the cross-border competitors each respondent actually faces. For a firm exporting heavily to a neighboring country, its most consequential rival may be absent from it. The authors deliberately chose the three large economies because they expected their signals to be salient. That choice gives the test a clear target while narrowing the claim it can support.

The paper does examine export exposure. Domestic-belief responsiveness is larger among firms with lower export shares than among those with higher export shares, a pattern consistent with local competition. The lower-export subgroup also has a weak-instrument warning. A stronger next test would identify the markets where each firm sells, match the signal to its actual competitors, and follow whether planned budgets become expenditure.

Information is a lever with an unmeasured endpoint

The findings do not make financing or implementation friction disappear. The ECB's SAFE background analysis reports that firms give practical reasons for limited AI use, including doubts about usefulness, system incompatibility and missing skills. An accurate peer figure can make an investment more plausible without supplying the engineers, data access or cash to execute it. Financial autonomy appears among the controls in the working paper, but the experiment is not a direct test that peer information overcomes a funding constraint.

In his 30 October speech on European AI adoption, ECB Executive Board member Philip Lane highlighted the experiment's domestic-versus-three-country pattern. The speech places the paper in a broader economic discussion; neither it nor the working paper creates a new policy or establishes that information campaigns will produce lasting investment. The paper's authors suggest such campaigns as a possibility to study, not an intervention already proved to raise realized AI spending.

For managers, the immediate lesson is about evidence quality. A benchmarking email can change a budget answer. To know whether it changes adoption, a follow-up would need purchase records, AI deployment and perhaps productivity measures, with a credible comparison group. To know whether Europe's single market transmits competitive pressure, the comparison benchmark should reflect the markets in which each respondent actually trades.

The experiment identifies a potentially useful starting point: firms may underread the AI investment already happening around them, especially among familiar domestic peers. It leaves the expensive endpoint open. The reported allocation moved; the money has yet to be observed moving with it.