Food manufacturing bought AI for language, not for production
The share of Swedish food producers using at least one AI technology went from 4.71 to 21.39 percent in two years. The share using AI for production processes fell over the same period, and that decides what the sector figure is worth to you.
Key insights
- The share of Swedish food producers with ten or more persons employed using at least one AI technology was 21.39 percent in 2025, against 4.71 percent in 2023, in Eurostat's survey of ICT usage in enterprises.
- Seven technologies are measured. The three that grew all handle language: text analysis to 16.13 percent, language generation to 10.10 and speech recognition to 5.62. The other four moved under six tenths of a point.
- The share using AI for production processes fell from 2.03 percent in 2023 to 0.70 percent in 2025. Across Swedish manufacturing as a whole the same item rose to 5.04 percent.
- Danish food producers stood at 5.48 percent on that item in 2025, Norway at 2.63 and the EU27 at 2.22. Sweden is lowest of the four.
- The sector figure does not measure where in the flow AI is used, and for a producer whose margin varies with the raw material it describes work that never touches that margin.
Eurostat's survey of ICT usage in enterprises measures AI by economic activity, and for Swedish food, beverages and tobacco manufacturing the development looks like a breakthrough. The share of enterprises with ten or more persons employed using at least one AI technology went from 4.71 percent in 2023 to 17.24 percent in 2024 and 21.39 percent in 2025.
That is the figure that ends up in a board presentation. The same survey breaks the use down by technology and by purpose, and the breakdown shows that everything growing is language technology, while one of the seven purposes measured across the whole period has fallen at every measurement since 2023.
Everything that grows is language technology
The survey separates which technologies enterprises use from what they use them for. Seven technologies are measured, and three of them grow. Analysis of written language went from 2.62 percent in 2023 to 16.13 percent in 2025. Language generation and speech synthesis went from 0.50 to 10.10 percent. Speech recognition went from 1.13 to 5.62 percent. The figures are shares of all enterprises in the population and not of each other: 21.39 percent report at least one technology, 16.13 percent report text analysis specifically.
The other four stand still or fall. Image recognition went from 1.04 to 1.41 percent, machine learning for data analysis from 1.04 to 1.31, process automation from 3.20 to 2.91 and autonomous machines from 1.66 to 1.10. None of the four has moved more than six tenths of a percentage point in two years, while the smallest of the three that grew has moved four and a half.
The technologies that grow do not sort themselves by how advanced they are, but by whether they handle words.
That makes the headline figure something other than what it looks like. A sector where four of seven measured technologies barely move, and where the three that do move all handle language, has not broadened its AI use. It has adopted one kind of AI. What that kind is used for sits in the other half of the survey.
One purpose falls, and it is the one that touches the product
The purpose side measures what the AI is used for, as items of its own. Seven of them have data across the whole comparison, and six of the seven stand higher in 2025 than in 2023. The largest are marketing or sales, from 0.52 to 5.63 percent, and business administration or management, from 1.80 to 5.71 percent. Accounting and finance management, ICT security, research and development, and logistics move less.
The seventh item is AI for production processes, and it is the only one that falls, at every measurement: 2.03 percent in 2023, 1.63 percent in 2024 and 0.70 percent in 2025.
Share of enterprises with ten or more persons employed in NACE C10-C12. All purposes with published values for 2025. The item for human resources and recruitment is not published. An enterprise can appear in several items.
Source: Eurostat, isoc_eb_ain2, Sweden, NACE C10-C12, retrieved 2 August 2026
The two that grow most have something in common. Administration, marketing and sales are office functions, and they are also the functions where the work already consists of text. The technology that grows and the purposes that grow most describe the same thing from two directions. The item closest to what the company actually manufactures is the only one going the other way.
The rest of manufacturing moves in the opposite direction
The same survey measures manufacturing as a whole, NACE C, which the food companies are themselves part of. There, AI for production processes rises: 2.59 percent in 2023, 3.90 percent in 2024 and 5.04 percent in 2025.
That is as close to a control condition as a survey gets. Same country, same survey, same year, same question wording, and a population the food companies are a subset of. What differs is the economic activity.
Share of enterprises with ten or more persons employed. Food is NACE C10-C12, manufacturing is all of C. Both series carry the break-in-series flag on the 2023 observation.
Source: Eurostat, isoc_eb_ain2, Sweden, retrieved 2 August 2026
The distance is not primarily in the headline figure. Manufacturing as a whole stands at 32.86 percent for at least one AI technology in 2025 against food manufacturing's 21.39, which is a difference but not a gulf. On the production item it is 5.04 against 0.70 percent. That data already sitting there has rarely had a question put to it is a pattern that recurs across manufacturing, and machine logs are the clearest example.
Danish food producers do the opposite
The same breakdown exists for other countries, and it tests whether the pattern belongs to the sector or to Sweden. Danish food, beverages and tobacco manufacturing stood at 5.48 percent for AI in production processes in 2025, against 3.94 percent in 2024 and 4.40 percent in 2023. Norway stood at 2.63 percent and the EU27 at 2.22 percent.
Share of enterprises with ten or more persons employed in NACE C10-C12. Only the Swedish series carries the break-in-series flag, set on the 2023 observation.
Source: Eurostat, isoc_eb_ain2, NACE C10-C12, retrieved 2 August 2026
On the headline figure Sweden is not behind. 21.39 percent against Denmark's 26.48 and Norway's 21.32 percent. It is on the production item that Sweden comes last of the four, and that is the observation that decides how the rest should be read. A Danish producer in the same sector, under the same EU law, runs AI in production at almost eight times the share of enterprises. The question is therefore not whether it can be done in food production.
What argues against this reading
The evidence has three weaknesses that belong in plain sight. The Swedish series that were checked carry Eurostat's break-in-series flag on the 2023 observation, which makes 2021 unusable as a comparison and makes 2023 the first point under the new definition. The flag has also not been checked for every series: for language generation, speech recognition, process automation, machine learning and several of the purpose items it is unknown.
Eurostat publishes no uncertainty figure for this breakdown. The metadata file states that the national statistical institutes supply standard errors for a selection of indicators and breakdowns, and this breakdown is not one of them. The response from the API carries status flags but no confidence interval.
And the movement is small in absolute terms. The fall from 2.03 to 0.70 percent is 1.33 percentage points, in a table where five series peak in 2024 and fall back in 2025: machine learning, image recognition, process automation, accounting, and research and development. Individual cells in this breakdown move sharply between years.
A figure with no published uncertainty should never be left to carry an argument on its own.
What still carries the reading is that it does not rest on one cell. Seven technologies are measured, and the three that grow are the three that handle language, while the four covering data, images, workflows and machines stand still. Seven purposes are measured, and the only one that falls is the one closest to production. The control group moves in the opposite direction on the same item, the same year, in the same survey. And three other populations sit above the Swedish one on that item. Noise explains a cell. It does not explain a sort.
The question that decides whether this applies to you
The sector figure does not describe every food producer equally well, and the difference can be put as a question: where does the variation in your margin come from?
If it sits in the raw material, that is in every batch differing and in yield, waste and deviations varying between runs, then the growing items describe work that never touches it. Administration gets cheaper and marketing gets faster, and the margin stays where it stood. For that producer, the sector's 21.39 percent is no evidence that the work has begun.
If the variation sits instead in the order book, that is in demand swinging and the assortment changing, then the growing items are the ones closest to the money, and the sector figure describes the business fairly well.
Most producers have both, and then it comes down to which of the two costs more right now. Demand forecasting, production planning and quality documentation all sit in the seam between them, and all three rest on history the company already owns: sales, orders, deviations, complaints and batches. That is also why the question can be settled without replacing any system.
What the sector figure cannot say is where in your flow a model would make a difference. It says how many in the sector answered yes to a question. What points out the processes is a bounded analysis on your own data, and that is what an AI readiness assessment does for a single unit.
Common questions
21.39 percent of Swedish enterprises in food, beverages and tobacco with ten or more persons employed reported in 2025 that they use at least one AI technology. The corresponding share was 17.24 percent in 2024 and 4.71 percent in 2023. The figures come from Eurostat's survey of ICT usage in enterprises, which Statistics Sweden runs nationally, and cover NACE C10-C12. Smaller producers with fewer than ten persons employed are outside the population.
Analysis of written language, which Eurostat measures as its own item. The share of Swedish food producers reporting it went from 2.62 percent in 2023 to 7.10 percent in 2024 and 16.13 percent in 2025. Two other language technologies also grew: language generation and speech synthesis to 10.10 percent, and speech recognition to 5.62 percent. The four technologies measured beyond those, namely image recognition, machine learning for data analysis, process automation and autonomous machines, were all below three percent in 2025. The item measures analysis of written language and is not the same thing as generative AI or chatbots.
The share of Swedish food producers reporting that they use AI for production processes has fallen at every measurement since 2023: 2.03 percent in 2023, 1.63 percent in 2024 and 0.70 percent in 2025. It is the only one of the seven purposes with data across the whole period that falls. Eurostat publishes no uncertainty figure for this breakdown, so a single cell should be read with care. That the same item rises across Swedish manufacturing as a whole and in Danish food manufacturing still argues against the movement being nothing but sampling variation.
On the headline figure Sweden sits close to Norway and below Denmark: 21.39 percent against Norway's 21.32 and Denmark's 26.48 percent in 2025. On AI for production processes Sweden is lowest of the measured populations at 0.70 percent, against Denmark's 5.48, Norway's 2.63 and the EU27's 2.22 percent. The Danish and Norwegian series carry no status flag, while the Swedish ones carry a break in the time series at 2023.
Every Swedish series that was checked carries Eurostat's break-in-series flag on the 2023 observation. A break means the definition or the method changed, so values before the break are not comparable with values after it. The comparison therefore runs 2023, 2024 and 2025. The Danish, Norwegian and European series carry no such flag and can be read further back.
It is one of the purposes the survey measures, alongside marketing or sales, organisation of business administration or management, accounting and finance management, ICT security, research and development, and logistics. The purposes are measured separately from the technologies, that is separately from the items for text analysis, language generation, speech recognition, machine learning, image recognition, process automation and autonomous machines. An enterprise can therefore be counted in several items at once, and the purpose item says what the AI is used for rather than which technology sits behind it.
Not without knowing where in the flow the use sits. A sector share says how many enterprises answered yes to a question, not whether the AI is used where the margin is decided. For a producer whose variation lies in the raw material, that is in yield, waste and deviations between batches, the growing items describe work in administration and marketing that does not touch that variation.
In the history the company already owns: sales, orders, deviations, complaints and batches. Demand forecasting, production planning and quality documentation all rest on those records and need no system replacement to be analysed. The work starts by naming which processes actually carry the cost, which is a bounded analysis on the company's own data.
If this lands on your desk, we should talk.
Ampliro Insights
New analysis, roughly weekly.
We write when the rules change and when something turns out to work in practice. One piece at a time, no sequences, and you can leave from any issue.
We store your address to send Ampliro Insights, and for nothing else. More in the privacy policy.