How do small manufacturers adopt AI, according to the research?

Slowly, and mostly at the edges. The OECD's December 2025 report finds 11.9% of firms with 10 to 49 employees using AI in 2024, against 40% of firms with 250 or more. Among small firms using generative AI, it is "mostly used for peripheral rather than core tasks". Where manufacturers do take AI into production, a 2025 study of tens of thousands of firms finds productivity falls in the short run before it rises. Those losses concentrate in older firms, not smaller ones, and about a third of them trace to management routines that lapsed.

What the usual answer says

Typical answers say small manufacturers adopt AI gradually, through pilots in quality control, maintenance and scheduling, held back by cost, skills and legacy systems. The research supports that. It adds what happens to a plant after AI reaches production.

How far adoption has got

The OECD's report on AI adoption by small and medium-sized enterprises sets out the gap. Across member countries, "40% of firms with 250 or more employees were using AI in 2024", against 20.4% of firms with 50 to 249 employees and 11.9% of firms with 10 to 49. Adoption in the ICT sector is often "more than three times higher than the manufacturing sector".

In an OECD survey, "Among SMEs using generative AI, only 29% report using it in their core activities", and 50% of SMEs said their employees lack the skills to use it. For small manufacturers, obstacles include "difficulty finding vendors of AI solutions tailored to their needs, lack of quality data and digital readiness". The report also finds that AI users' productivity advantage shrinks noticeably once their wider digital capabilities are accounted for. Much of the gain comes with the groundwork.

What happens after adoption

McElheran, Yang, Kroff and Brynjolfsson's 2025 working paper is the deepest look at manufacturers after they adopt. It combines large surveys for 2017 and 2021 with balance sheet and tax records, including a panel of about 55,000 manufacturing firms.

It reports "causal evidence of J-curve-shaped returns, where short-term performance losses precede longer-term gains". Industrial AI use "increases work-in-progress inventory, investment in industrial robots, and labor shedding, while harming productivity and profitability in the short run". Two estimates matter here. As a correlation, after controls for size, age and IT, a one standard deviation rise in the authors' AI index goes with a 1.33 percentage point drop in productivity. Their causal estimate, for the plants their method isolates, is far larger, so 1.33 points is not the size of the dip.

Growth follows for some. Earlier adopters "exhibit stronger growth over time, conditional on survival". That condition matters: plants that didn't come through the dip aren't in the growth figures.

Age, not size

The paper sets out to disentangle "size from correlated organizational characteristics such as age". It notes that "larger firms have emerged as leading adopters". The losses it finds depend on age: adjustment "is more challenging for older businesses", while young firms with growth-oriented strategies "show stronger AI-related performance improvements". Older firms "struggle to maintain vital production management practices such as monitoring key performance indicators and production targets". That fall-back, measured as a decline in structured management scores, "accounts for about one-third of their productivity loss".

Reading it through a data lens

In our view, the part of the dip most within management's control is that third. A plant can keep its daily KPI review, its targets and its shift reports running unchanged while a new system beds in, and check each week that they still happen. If throughput dips and the reviews held, the dip is the adjustment the research describes. If the reviews slipped, that is the first thing to fix. We would also put data quality first. A model trained on inconsistent tallies or downtime codes is, in our experience, soon distrusted.

The strongest objection

The J-curve concerns industrial AI in production between 2017 and 2021: machine vision, predictive systems and tools that reorganise the line. Most small manufacturers in 2026 meet AI as generative tools and features inside office and ERP software, used at the edges, as the OECD finds. No study we opened documents a production dip from that kind of use. For those firms, the barriers that matter are the familiar ones: skills, suitable vendors and data.

When it doesn't apply

Young firms, and plants built around digital systems from the start, face a shallower curve on the paper's own findings. Uses kept outside production, such as drafting quotes, carry little risk of a production dip. And the paper covers one national economy's manufacturers before generative AI.

Quarri for wood products is built around how a wood products plant runs, where the order book meets real capacity.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.