The Diffusion Gap

1. Frontier abundance, mid-market scarcity

The last decade has not been short of invention. General-purpose AI, cloud computing, cheap sensing, and mature open-source tooling are, in principle, available to any firm with a network connection. The frontier is crowded. Yet the productivity dividend that should follow has not materialised across the European economy — and the reason is distributional. The tools cluster at the top of the size distribution and thin out sharply below it.

Eurostat's enterprise survey makes the gradient explicit. In 2024, 41.2% of large EU enterprises used AI, against 20.9% of medium-sized and just 11.2% of small firms.[3] A large firm is nearly four times as likely to have adopted AI as a small one. Because small and medium firms are the overwhelming majority of the economy, an adoption rate that looks merely "early" in aggregate is in fact highly concentrated: the frontier is real, but it is a frontier most of the economy has not crossed.

This is the paper's central asymmetry. Invention has a natural home in large, well-capitalised organisations; diffusion has no natural agent at all in the mid-market. Nobody's business model is to carry a €40M manufacturer across the adoption threshold at acceptable cost and risk. That absence — not a shortage of technology — is what the following sections cost out.

2. Measuring the adoption gap

Two independent measurements converge on the same conclusion.

The productivity outcome. The EU–US labour-productivity gap has widened for a quarter-century. On OECD figures, EU-27 output per hour fell from roughly 90% of the US level in 2000 to about 75% in 2024; over the same window US output per hour grew ~1.8% annually against the EU's ~1.1%.[2] The Draghi Report (European Commission, September 2024) found the EU–US income gap widened further in 2023, with the majority of that widening attributable to productivity, and traced the productivity shortfall largely to the technology sector and its diffusion — not to any single macro shock.[1] The ECB adds the recent data point: euro-area labour productivity actually fell 0.9% in 2023 while US productivity rose 1.6%.[5]

The adoption input. The same period's adoption data explains the outcome. Only 13.5% of EU enterprises (10+ employees) used AI in 2024, up from 8.0% in 2023[3] — against the EU's own Digital Decade target of 75% of enterprises using cloud, big data and/or AI by 2030.[4] The distance between a 13.5% baseline and a 75% target, six years out, is not a rounding error to be closed by exhortation; it is a structural deployment deficit.

The two panels are the same fact viewed twice. Where adoption is thin, productivity growth is thin. The policy question is therefore not "how do we invent more?" but "why does the mid-market not adopt what already exists — and what would change that?"

3. Why the mid-market under-adopts

The reflexive explanations — mid-market firms are too small, too conservative, too digitally illiterate — are mostly wrong, or at least secondary. The binding constraints are structural features of the market for deployment, not deficits in the firms themselves.

3.1 The deployment market is broken

Enterprise software is sold and implemented for enterprises. The systems-integrator model, the six-figure implementation, the multi-quarter rollout — these are priced and paced for organisations with dedicated IT functions and change-management budgets. A 150-person manufacturer has neither. The technology is accessible; the means of adopting it safely is not. There is a missing rung on the ladder between "download an app" and "engage a systems integrator," and the mid-market lives precisely on that missing rung.

3.2 No forward-deployed engineers

The firms that have crossed the AI threshold typically did so with human help embedded in their context — engineers who sit with the operators, learn the plant, and wire the tool into the real workflow. That model does not scale down: nobody can afford to forward-deploy an engineer into every 40-person casting shop. The mid-market needs the effect of a forward-deployed engineer — software that arrives already understanding operational context — without the unit economics that make human deployment impossible below a certain firm size.

3.3 The single market is fragmented in practice

The Letta Report, Much More Than a Market (April 2024), documents how Europe's single market remains fragmented across languages, tax regimes, invoicing standards and regulatory surfaces.[6] A deployment path that works must absorb this fragmentation as a feature — it must speak the customer's language, ingest the national invoicing standard (in Italy, FatturaPA/SDI), and respect each jurisdiction's data rules — rather than treating them as edge cases to be handled later.

3.4 Death by pilot

The characteristic failure mode of mid-market technology adoption is the pilot that never ends: a proof-of-concept that consumes months, produces a slide deck, and dies for lack of a decision. When time-to-value is measured in quarters, the political and financial cost of an inconclusive pilot exceeds the expected benefit, so the rational choice is not to start. Slow time-to-value doesn't just delay adoption; it prevents it, because it inverts the risk calculus before any value is ever demonstrated.

These four constraints share a root: the mid-market has no low-risk, fast, sovereign, context-aware path from "interested" to "getting value." Sections 4 and 5 specify that path and then instantiate it.

4. What a working deployment path must look like

Stated as neutral requirements — a specification any vendor could be measured against, not a pitch. A deployment path that actually closes the diffusion gap for the mid-market must satisfy five conditions:

1. Fast, unconditional time-to-value. It must produce a concrete, credible result on the customer's own data within days, before any deep integration — defeating death-by-pilot by making the first proof cheap and unambiguous. Value must not be contingent on a long project completing.
2. No system-of-record replacement. It must sit on top of the existing ERP/gestionale and read from it, never demand a rip-and-replace. The incumbent system of record is load-bearing and politically untouchable; a path that threatens it will be rejected regardless of merit.
3. Sovereign by construction. For European industrial firms — especially in food, pharma, defense and energy — data residency and control are not premium features but preconditions. The path must be able to run fully on the customer's own infrastructure, so that "where does my data go?" has the answer "nowhere," by default rather than by upgrade.
4. Human-governed. Adoption stalls on the trust question: can I let software act? The path must keep a human in the loop by construction — proposing, not executing, until trust is explicitly and incrementally granted — so that governance is the product's mechanism, not a compliance bolt-on. This is also how it satisfies the EU AI Act's human-oversight obligations natively.
5. Modular and scale-down. It must start at the size of the problem the customer actually has — one wedge, one module — and grow only as the customer switches more on. A monolith that must be adopted whole cannot scale down to a 20-person firm; a manifest-driven module system can.

Note what this specification is not: it is not "cheaper enterprise software," and it is not "a chatbot." It is a different shape of product — one whose first job is to collect and connect the firm's data well, and whose adoption curve is designed around the mid-market's real risk tolerance.

5. Dimbo as a diffusion vehicle

Dimbo was built to the specification in §4 — not by coincidence, but because the diffusion gap is the problem it addresses. Three design choices map directly onto the five requirements.

The wedge defeats death-by-pilot (Req. 1). Dimbo lands with a 48-hour Deadline Audit: a deterministic scan of a firm's open commitments, overdue invoices and expiring contracts into per-client 30/60/90-day risk buckets with hard cash figures — on a data slice, with no installs. It needs no cross-pillar magic and no integration project; it produces a number on the customer's own data in the first week. The audit converts "interested" into "getting value" before any pilot risk accrues — the single most important move against the mid-market's inverted risk calculus.

Zero rip-and-replace meets the incumbent where it is (Req. 2 & 3). Dimbo sits above the ERP and reads it; in Italy its erp_connectors module ingests FatturaPA/SDI invoice XML directly into its finance records from day one, with adapters for the national platforms (TeamSystem, Zucchetti). Nothing is ripped out. And because Dimbo can run fully on-prem on a benchmarked local model — a gemma-class local LLM at reference-parity on a single workstation-class GPU, with local vision and local transcription — sovereignty is the default, not a premium tier. When an external model is ever used, PII is anonymised at the gateway first. This is the honest, real property set: GDPR-by-design, on-prem, PII anonymization, full audit trail — no certification claimed that is not held.

The autonomy ladder makes trust incremental and governance native (Req. 4). Every Dimbo capability begins as a proposal a human approves in an Action Center. Autonomy is earned per process, only by measured track record, and promotion is always the customer's decision; at the top of the ladder the system acts with an undo window and instant downward demotion the moment a human disagrees. This is human-in-the-loop by construction — and therefore EU AI Act human-oversight obligations are met by the product's core mechanism rather than a compliance appendix. It is also the adoption engine: trust grows one process at a time, which is exactly the pace the mid-market can absorb.

The module system scales down and compounds (Req. 5). Dimbo is a manifest-driven plug-in platform, not a monolith. A firm lights up only what it needs, and because every module writes into one shared graph and one shared knowledge store, the second module roughly triples value rather than doubling it — the join across domains is where the non-obvious value lives. The same platform therefore fits a 20-person service firm and a 500-person factory, differing only in which modules are switched on.

Europe closes its productivity gap not by inventing more, but by deploying what already exists — one governed decision at a time. — The diffusion thesis

A representative scenario. Molitoria San Prospero, a fictional €40M flour and semolina producer, runs the Deadline Audit on a slice of its order and invoice data. In week one, the audit surfaces roughly €180,000 of avoidable retailer late-delivery penalty exposure (the penali GDO that were previously invisible until the penalty invoice arrived), bucketed by which contracts are 30, 60 and 90 days from breach. No system was replaced; no data left the building; a human approved each proposed intervention. That is diffusion made concrete: a frontier capability delivering a governed, sovereign result inside a mid-market firm — in days, not quarters.

6. Conclusion — deploy what exists

Europe does not primarily have an invention problem; it has a diffusion problem. The frontier tools that raise productivity exist and are largely accessible, but the mid-market backbone — the 99.8% of enterprises and two-thirds of employment — cannot reach them, because the market for deployment is broken: too slow, too risky, too sovereign-hostile, too monolithic. The EU–US productivity gap and the 13.5%-versus-75% adoption gap are two views of that single failure.

Closing it does not require another moonshot. It requires a deployment path engineered to the mid-market's real constraints: fast unconditional time-to-value, no system-of-record replacement, sovereignty by default, human governance by construction, and modular scale-down. Dimbo is one working instance of that specification — landing with a wedge, growing with a graph, and acting only as far up the autonomy ladder as the customer has decided to trust it. The productivity dividend Europe has been waiting to invent is, for most of its economy, already available to deploy.