GoodMills Senior Management Circle 2026 · Krems an der Donau · 8 September 2026

Embracing
the Future

The intelligence behind the keynote: what is coming, what it means for milling, and what leaders do.

Aragorn Meulendijks · Future-historian

Explore the findings Find your country

mills

countries

grain a year

turnover · BY 23/24

leaders in the room

GoodMills Group · BY 23/24Conference details and approximate attendance: client brief.

Motto: Leading through Change, Embracing the Future

The executive summary

Six findings: how AI and robotics remake milling by 2036

Not a balance sheet. Each finding says what is happening, what it does to grain, mills and flour within five to ten years, and what it demands of the people who lead GoodMills.

  1. Intelligence is becoming a commodity input, like electricity and wheat

    The price of a given AI capability falls about 40× a year, frontier performance reaches a consumer laptop in about eight months, and the task length an AI agent completes on its own doubles every 131 days. In a business where up to 80% of cost is grain and margin lives in fractions of a percent of yield, a near-free, compounding intelligence input is the next raw material. Denmark’s enterprises already buy it in bulk (42% AI adoption); Poland (8.4%) and Romania (5.2%) do not yet. Leadership implication: securing and deploying this input becomes as deliberate a job as grain procurement.

    epoch.ai · epoch.ai · metr.org · buhlergroup.com · ec.europa.eu

  2. Agents arrive on every layer, and the skill delta between top and bottom collapses

    40% of large organisations already scale AI agents and 32% skipped a software purchase because they built it with agents. OpenAI reported its “automated research intern” milestone on 6 September and targets an autonomous AI researcher by March 2028. AI lifts novices by 34% and experts barely at all. Within two years every mill operator, grain buyer, brand manager and plant director runs a team of agents; a 24-year-old lab technician and the Group CEO do the same kind of work, directing agents, at different scales. Leadership implication: hierarchy built on who knows more stops working; leadership built on example, ownership and judgment is what remains. That is what the five Commitments describe.

    mckinsey.com · engadget.com · nber.org

  3. The mill runs itself; the human moves from operator to designer

    Whitley Bridge has run since 2021 with one operator and more than 15,000 data points. Bühler, whose machines process 65% of the world’s milled wheat, is prototyping mills that set their own grinding gaps and moisture targets: “millers will set desired product characteristics, and the mill will automatically make the necessary adjustments.” By 2031 a mill is a self-optimising plant benchmarked hourly against the other 23; by 2036 the night shift is a humanoid crew (22,000 humanoids shipped in H1 2026, up ~300%; US food-plant robot installations +30% in 2025 while automotive fell 1%). Leadership implication: the miller’s craft moves up one level, from running the process to designing the product, the recipe and the exception.

    buhlergroup.com · world-grain.com · counterpointresearch.com · ifr.org

  4. The field becomes a data source and biology becomes programmable

    18% of EU farms already run precision farming, on 44% of the farmland; the EU publishes a satellite-driven yield forecast every month; John Deere sells autonomy as a retrofit kit for machines already in the shed. The NGT regulation entered into force on 16 July 2026 and applies from mid-2028: gene-edited, climate-resilient wheat becomes a routable EU product inside this room’s planning horizon, and it is banned from organic, so the two supply chains diverge. Leadership implication: grain procurement becomes a data-and-biology decision. The buyer who trades against satellite forecasts and knows which edited varieties are coming outruns the buyer who reads last year’s book.

    ec.europa.eu · publications.jrc.ec.europa.eu · prnewswire.com · consilium.europa.eu

  5. Physical AI compounds: better models, robots, materials, machines and crops

    AI-designed materials, AI physics simulation that evaluates 10 to 100 times more design alternatives, and robot factories that went from one unit a day to one an hour in under 120 days form one loop that feeds itself: better models design better robots, which build better machines, which run on better materials. That is why the curve is exponential rather than a series of product launches. Leadership implication: an annual budget against a four-month capability clock is reacting on a delay. Someone in every country unit watches the frontier daily and translates it into “what does this mean for our mill this quarter?”

    figure.ai · epoch.ai · press.bmwgroup.com

  6. The extreme-but-possible 2036: a flour company that is a data-and-biology company

    Fewer, larger, smarter mills (the consolidation this industry already survived, on a faster clock); grain bought by agents on field-level data; recipes designed with AI biology for customers’ ovens; humanoid crews on the shifts nobody wants; 1,500 people who each lead a hundred agents and a few dozen robots, with human work concentrated on judgment, relationships, taste and deciding what to build next. Nestlé is already moving 12,000 white-collar roles in this direction. Leadership implication: this future is not chosen by the technology. It is chosen by whichever miller’s leaders shape the change instead of managing it. The five Commitments are the operating system for that choice.

    fooddive.com · dairyreporter.com · muehlen.org

The disruption · 2031–2036

Agents on every layer. Robots in every field and mill.

Two deep dives: how agentic AI collapses the skill delta between the top and the bottom of the organisation, and how the rise of AI powers embodied AI, materials, physics and biology along the wheat-to-flour chain. Verified evidence first, then the extreme-but-possible scenario a futurist owes the room, then what it demands of leaders.

Deep dive 1 · Agentic AI on every layer

The skill delta between top and bottom collapses

40% of large organisations now scale AI agents, up from 27% a year ago, yet only about 10% of enterprise functions run on them. The next five years of competitive separation happen in that gap. mckinsey.com · forbes.com

Inside OpenAI, 3.1 agent-workdays are spent for every human workday. The company declared its “automated research intern” milestone met on 6–7 September and targets an autonomous AI researcher by March 2028. datastudios.org · helpnetsecurity.com

The task length an agent completes on its own doubles every ~7 months long-run and every ~4 months recently: two seconds in 2019, about twelve hours in 2026. Month-long tasks land around 2027. metr.org · theaidigest.org

Measured on 5,179 workers: +34% for novices, roughly zero for experts. The tool spread the best workers’ practice to everyone and made novices learn faster. Twenty-five years of milling intuition, held privately, loses value the moment it is encoded and served to a 23-year-old through an agent. nber.org

Layers are going: Amazon removed about 30,000 corporate roles to flatten management during 21% revenue growth; Salesforce’s support workforce went from ~9,000 to ~5,000 with agents resolving ~85% of 1.5 million inquiries; Klarna’s assistant did the work of 700 agents in month one, then rehired for human escalation when cost became “too predominant”. briefs.co · aboutamazon.com · customerexperiencedive.com · openai.com · fastcompany.com

In your neighbourhood: Nestlé has 100,000+ regular AI users with agentic AI advising finance on working capital and risk; Cargill’s AI optimises grain flows, plant yields and port logistics; Bimbo Bakeries cut forecast error by up to 30% across 57,000 routes. Your grain counterparties already trade with better eyes than yours; your bakery customers order better than you predict them. foodnavigator.com · cargill.com · supplychainbrain.com

Organisational factors drive 67% of AI impact, twice the weight of individual mindset. The bottleneck is the operating model. microsoft.com

Deep dive 2 · Embodied AI, from field to flour

Better models, robots, materials, machines and crops, one loop

Humanoids crossed from demo to deployment this year: more than 22,000 shipped in H1 2026 (+~300%), over 50,000 projected for the year, five Chinese firms holding 86%. Schaeffler deploys up to 2,000 humanoids in two German plants from late 2026; Figure 02 ran 1,250+ hours at BMW. counterpointresearch.com · dcvelocity.com · press.bmwgroup.com

Food is the lagging sector for one solvable reason: no humanoid platform has published IP69K wash-down certification yet; 2026–27 is the stated horizon. Quadrupeds already inspect food plants at AB InBev and Cargill. ifactoryapp.com

The field: John Deere’s autonomy kit launches fully in 2026 and retrofits older machines, so the installed base converts without a replacement cycle. Carbon Robotics’ LaserWeeder kills 200,000 weeds an hour, 10 billion since 2022. In Europe 18% of farms run precision farming on 44% of the land; Luxembourg, Finland and Estonia above 75% of farmland, Greece and Romania at 10–15%. A pan-European miller takes data-rich and data-dark grain into the same intake. deere.com · carbonrobotics.com · ec.europa.eu

The mill: Bühler’s SmartMill is a self-optimising, self-learning programme with prototypes deciding grinding gaps and moisture; reported digital-service results 48% less downtime, 49% fewer errors, 18% less energy. Whitley Bridge: one operator, 15,000+ data points, since 2021. buhlergroup.com · world-grain.com · crown.co.za · buhlergroup.com

Materials and physics: DeepMind’s GNoME predicted 2.2 million new crystals, 736 already synthesised by outside labs (wear-resistant rolls, lighter robot limbs). NVIDIA’s Apollo physics models search airflow, dust, thermal and structural designs at tens-to-thousands× speed: a mill becomes a searchable design space, not a one-shot drawing. deepmind.google · simtool.com

Biology: Rothamsted’s CRISPR wheat cuts free asparagine 59% (up to 93%) with no yield loss and acrylamide in bread up to 45% lower. The EU NGT regulation is in force since 16 July 2026 and applies from 17 July 2028. AI enzyme design is producing de novo industrial enzymes; amylases and xylanases for baking are exactly that category. From July 2028 a European miller can legally source wheat engineered for what happens in the customer’s oven. rothamsted.ac.uk · food.ec.europa.eu · sciencedirect.com

Extreme but possible · 2031–2036 · Aragorn’s frame

The grain buyer with twelve agents

By 2033 every GoodMills grain buyer runs a standing team of eight to twelve agents: one on satellite biomass across the Danube basin hour by hour, one on freight and energy curves, one modelling protein and falling number per silo, one negotiating against the counterparty’s agent inside a pre-approved envelope, one writing the memo the buyer used to spend Thursday on. The buyer sets the risk envelope, says no, and owns the consequence. A two-year hire and a thirty-year veteran both direct competent agent teams; what separates them is judgment under ambiguity, the only remaining scarce good. The pyramid becomes a mesh: fifty people, four hundred agents, one shared model of the business. Two management layers in a 24-mill group disappear, not by redundancy programme but because they existed to relay information that now moves at machine speed. The CEO and the newest plant engineer open the same interface and are measured on the same thing: the quality of the decisions they authorise.

What it demands of leaders

  1. Stop managing information flow; start managing decision rights. Aggregation is now free. Redesign every role around the decision it owns and the risk envelope it runs in, or the layer is removed for you.
  2. Encode the masters before they retire. The +34% only happens when best practice is captured. Every senior miller, buyer and blender is a private model; turning that into a shared, agent-served asset is the highest-return leadership act of the next three years, done with them, framed as legacy.
  3. Judge leaders on how well they direct agents, not on headcount. If promotion tracks span-of-humans, your best people optimise for hiring instead of leverage.

Bridges: Setting an example · Taking responsibility · Promoting development

Extreme but possible · 2031–2036 · Aragorn’s frame

A data-and-biology company that owns 24 grinding assets

By 2034 the wheat at a GoodMills intake carries a data payload heavier than its paperwork: every hectare planted, weeded, scouted and harvested by autonomous machines, every lot tagged with cultivar, soil, nitrogen curve, heat-stress hours, mycotoxin risk and asparagine precursor level, and priced on those attributes rather than a protein band. Procurement stops buying “milling wheat” and buys a functional specification, hedged by agents the moment the satellite biomass signal moves. The night shift has no humans in it: wash-down-certified humanoids handle sack changes, sampling, cleaning validation and the small non-standard interventions that blocked full automation for thirty years, while the process self-optimises against a live model of every customer’s bakery. The mill itself was searched, not drawn, in alloys that were computed rather than found. And the flour is programmable: gene-edited low-acrylamide, climate-resilient wheat matched with AI-designed enzymes lets GoodMills sell a baking outcome with a health claim instead of a commodity with a protein number. The assets are the least defensible part of that company.

What it demands of leaders

  1. Buy attributes, not tonnes, and build the data spine now. Pricing grain on measured functional attributes is three years of plumbing away and cannot be bought in the year you need it.
  2. Treat 17 July 2028 as a product-roadmap deadline. Low-acrylamide, climate-resilient wheat becomes routine in the EU on that date; decide this year whether GoodMills leads that conversation with bakeries and consumers or has it forced on it by a competitor’s marketing.
  3. Design the next mill for a workforce that is part human, part machine. A mill commissioned in 2028 lives to 2050; specify robot access, charging, wash-down-robot readiness and full instrumentation in the capex, or it is obsolete before it is depreciated.

Bridges: Creating focus & clarity · Shaping change

The keynote, in six threads

The future has a history

From a Roman millstone to intelligent mills: the threads shared with GoodMills on 2 September, connected to the evidence.

Thread 01

Almost everything we call technology is new

20 A.D.

A millstone dated to 20 A.D. was unearthed under the Castellmühle site. On that same site, GoodMills built a mill with “Milling on Demand”. The long history belongs to you. So does the next chapter.

millermagazine.com

Thread 02

Progress was never linear

<5% → 95.0%

In May 2025, leading models remained below 5% on ARC-AGI-2. By September 2026, Astra reached 95.0% on the semi-private set. Plan for capability to move while your project is still being built.

arcprize.org · arcprize.org

Thread 03

AI was predicted, not hype

1–3 September

The release window spans Fable 5.1 on 1 September, Gemini 3.8 Flash and Muse Spark 1.3 on 2 September, and GPT-6 Astra on 3 September. Four models across three days. Release dates follow the cited research trackers; the changing capabilities are the point.

llmgateway.io · artificialanalysis.ai · theneuron.ai

Thread 04

Frontier → mainstream in two years or less

~8 months

Epoch measures about eight months for frontier performance to reach consumer hardware, and a roughly 40× annual price decline for GPT-4-level performance on PhD science questions. These are measured trends, not a universal adoption deadline. In your supply chain, the NGT regulation entered into force on 16 July 2026; most provisions are expected to apply from mid-2028.

epoch.ai · epoch.ai · europarl.europa.eu · consilium.europa.eu

Thread 05

After AI comes embodied AI

22,000+

Counterpoint reports more than 22,000 humanoids shipped in H1 2026, up nearly 300% year on year. BMW says Figure 02 supported more than 30,000 X3 vehicles in around 1,250 operating hours. Schaeffler’s announced rollout begins in December 2026, with a four-digit deployment planned by 2032. These are shipments, a factory case and a commitment: three different kinds of evidence.

counterpointresearch.com · press.bmwgroup.com · roboticstomorrow.com

Thread 06

Leadership reinvented

32%

McKinsey finds 32% of respondents’ organisations skipped a software purchase because they could build it with agentic coding tools. Only about 6% are AI high performers; nearly three-quarters of that group redesigned workflows. Amazon’s Andy Jassy asked for at least a 15% increase in the individual-contributor-to-manager ratio by Q1 2025. Leadership increasingly means designing the work, not just supervising it.

mckinsey.com · aboutamazon.com

Frontier update · 7 September 2026

What moved this week

The latest model releases, the benchmarks behind the headlines, and what has actually been demonstrated.

Source check · 7 September 2026

The frontier moved while this keynote was being prepared.

Fable 5.1, Gemini 3.8 Flash, Muse Spark 1.3 and GPT-6 Astra arrived across the release window of 1–3 September. The practical opportunity is to learn what the newest tools can do for real work, with the same care used to commission any new equipment.

anthropic.com · artificialanalysis.ai · llmgateway.io · openai.com

The model landscape

Release dates and published results in the research snapshot. Scores are attributed to the named test; they are not interchangeable.

Frontier models, release dates and source-linked capability evidence
Model / labReleasedPublished evidence
Claude Fable 5.1Anthropic1 Sep 202665.7% on the AA Intelligence Index; 81.2% on SWE-bench Pro.anthropic.com · benchlm.ai · codingfleet.com
Gemini 3.8 FlashGoogle2 Sep 2026AA Index 59 at high reasoning; $0.75 / $3.75 per million input / output tokens through December.artificialanalysis.ai
Muse Spark 1.3Meta2 Sep 202662.1% on the AA Intelligence Index, as listed by BenchLM.llmgateway.io · benchlm.ai
GPT-6 AstraOpenAI3 Sep 202695.0% on ARC-AGI-2 at max reasoning; cybersecurity capability classified Critical by OpenAI.openai.com · arcprize.org · openai.com
Claude Opus 5Anthropic2026, before Fable 5.196% on SWE-bench Verified; 63.0% on the AA Intelligence Index.benchlm.ai · benchlm.ai
Grok 4.6xAI12 Aug 202660.9% on the AA Intelligence Index, as listed by BenchLM.llmgateway.io · benchlm.ai
Kimi K3Moonshot AI16 Jul 202659.7% on the AA Intelligence Index, as listed by BenchLM.llmgateway.io · benchlm.ai
GLM-5.3Z.AI2026; exact date unverified59.5% on the AA Intelligence Index, as listed by BenchLM.benchlm.ai
GPT-5.6 SolOpenAI9 Jul 202658.9% on the AA Intelligence Index, as listed by BenchLM.llmgateway.io · benchlm.ai
Qwen3.8-Max / Qwen3.8-27BAlibaba3 Aug / 2 Sep 2026SWE-bench Pro: 67.7% for Max and 61.7% for 27B. Vendor-reported results.llmgateway.io · codingfleet.com

Read the benchmark. Read the conditions.

Verified benchmark snapshot and the limits of each comparison
BenchmarkPublished resultWhat it measures / conditions
SimpleBenchClaude Fable 5 · 81.9%Below the 83.7% human baseline. The human reference is a small sample of nine participants; this is a specific common-sense test.simple-bench.com · lmcouncil.ai
Humanity’s Last ExamClaude Fable 5.1 · 65%Tool-assisted result. Expert-written questions; no single human baseline. Scores without tools are a different comparison.benchlm.ai
ARC-AGI-1GPT-6 Astra · 98.5%Semi-private evaluation; high / xhigh reasoning.arcprize.org
ARC-AGI-2GPT-6 Astra · 95.0%Semi-private evaluation; max reasoning. Above the 85% Grand Prize target; the target is not an average-human score.arcprize.org · arcprize.org
ARC-AGI-3GPT-6 Astra · 62.71%Standard harness, max reasoning. Provider Adapter results use different conditions and are not interchangeable.arcprize.org
SWE-bench VerifiedClaude Opus 5 · 96%Published coding leaderboard snapshot. Solving benchmark issues is different from operating a production system.benchlm.ai
SWE-bench ProClaude Fable 5.1 · 81.2%Vendor-reported scores unless otherwise noted by the leaderboard; treat small gaps cautiously.codingfleet.com
AA Intelligence IndexClaude Fable 5.1 · 65.7%Composite index, reported by BenchLM. It does not measure the same tasks as the individual tests above.benchlm.ai
ExploitBenchGPT-6 Astra · 100%OpenAI evaluation without production safeguards. This is a research test result, not unrestricted public access.csoonline.com · openai.com

Five developments to carry into the conversation

3 September

The open-model ecosystem attracts a major infrastructure bet.

NVIDIA announced an agreement to acquire Hugging Face for $12.93 billion. An announced agreement is distinct from a completed acquisition.

blogs.nvidia.com

For a miller: keep the ability to choose tools and suppliers as the technology market changes.

6 September

AI starts helping to build the next AI.

OpenAI says it has reached its automated research intern milestone: well-defined tasks under human direction, including work that takes a skilled researcher days. Its next target is an automated researcher by March 2028.

openai.com

This is the company’s reported milestone and future target; people still set priorities and judge results.

3–4 September

Greater capability brings a greater leadership responsibility.

OpenAI classified Astra as its first Critical cybersecurity model: 100% on ExploitBench without production safeguards, and two previously unknown zero-days found during testing. Enterprise access is off by default at launch.

csoonline.com · openai.com · openai.com

For a miller: agree who may connect an AI system to business data and operational tools, and who remains accountable.

This week’s benchmark check

Keep the breakthrough and the counterweight together.

ARC Prize reports 95.0% for Astra on ARC-AGI-2. On SimpleBench, the best published score is Claude Fable 5 at 81.9%, below the 83.7% human baseline.

arcprize.org · simple-bench.com · lmcouncil.ai

Test the actual job you want done. A benchmark is evidence about its own task, not a universal licence to delegate.

In this week’s business debate

The build-or-buy decision is changing.

In McKinsey’s 2026 survey, 32% of respondents said their organisations skipped at least one software product or feature because they could build it internally with agentic coding tools.

mckinsey.com

For a miller: a small, useful workflow can start with the people closest to the work. Include maintenance and ownership in the decision.

The evidence, at a glance

SimpleBench: Claude Fable 5 scores 81.9 percent, below the 83.7 percent human baseline.
Common sense still asks for care

SimpleBench: the best published AI score remains below the human reference. Scores shown follow the verified research snapshot.

simple-bench.com · lmcouncil.ai
Humanity’s Last Exam tool-assisted comparison: Fable 5.1 65 percent, Opus 5 64.7, Mythos 5 64.5, and Muse Spark 1.1 62.1.
Expert questions, with tools

Humanity’s Last Exam: tool-assisted results only. There is no single human baseline, and these results must not be mixed with scores without tools.

benchlm.ai
ARC benchmark comparison: Astra 98.5 percent on ARC-AGI-1 and 95 percent on ARC-AGI-2; Fable 5.1 90 percent on ARC-AGI-2.
Abstract reasoning moves forward

ARC Prize’s semi-private results. Benchmark versions differ; no invented average-human reference is shown.

arcprize.org · openai.com
Selected benchmark leaders: SimpleBench 81.9 percent, tool-assisted HLE 65, ARC-AGI-2 95, and SWE-bench Verified 96. These are distinct tests.
Different tests, different strengths

Selected published leaders across distinct tests. This is a comparison of benchmark results, not a single scale of intelligence.

simple-bench.com · benchlm.ai · arcprize.org · benchlm.ai

Your world

A familiar landscape, in motion

The pressures are real. So is the expertise this room has built by working through them.

Demand

The bread basket is changing

Gollegger named “stagnant European demand, declining margins and lack of efficiency” as consolidation drivers in his historical Krefeld-construction interview. millermagazine.com

German households bought nearly 10% less bread in 2024 than in 2019. Bakery purchasing reached about 45% of households in May 2024–April 2025, versus 52% five years earlier. This is household reach, not a bakery market-share measure. lebensmittelpraxis.de

Margins

The grain is the cost base

Bühler puts raw materials at up to 80% of milling costs. Small gains in yield therefore matter directly to the economics of the mill. buhlergroup.com

The annual global wheat benchmark fell from about $360/t in 2022 to $173/t in 2025 — a 52% decline. These are nominal annual averages, not local procurement prices. fred.stlouisfed.org

Energy

Efficiency remains a European advantage to earn

For energy-intensive industry in 2025, EU electricity prices were roughly double US levels and more than 50% above China, according to the IEA. iea.org

Leadership implication: make energy performance part of everyday operating decisions, alongside quality and yield.

The raw-material cycle · annual averages

A price shock, and its unwinding

Global price of wheat · nominal US dollars per metric tonne

Source: IMF, Global price of Wheat, via FRED; annual series retrieved through Alpha Vantage on 7 September 2026. Values rounded to cents. fred.stlouisfed.org · alphavantage.co
View the price data
Global wheat price, annual average
YearUS dollars / tonne
2019163.26
2020185.51
2021265.79
2022360.17
2023272.33
2024201.19
2025173.09

Consolidation

You have already lived a transformation

Germany: almost 19,000 mills in 1950 → 2,500 in 1980 → 500 today, in the milling association’s account. muehlen.org

Europe: around 15,000 mills in 1960 and 3,800 companies in the association’s 2020–21 manual, which reported average capacity use of about 65%. The current facts page says over 3,000 mills. These are sector figures, not GoodMills utilisation. iaom-eurasia.info · flourmillers.eu

Austria had 92 mills in 2020; the largest five produced 63% of its flour. landschafftleben.at

The UK cleared ABF–Hovis on 16 June 2026. In the US, nearly 50% of flour-milling assets changed hands in 21 days in May 2014. world-grain.com · bakingbusiness.com

Digital adoption

Different starting points, shared opportunity

In 2025, 19.95% of EU enterprises used AI, up 6.47 percentage points. Eurostat covers enterprises with at least ten people employed in its surveyed sectors. ec.europa.eu

Two separate Bitkom surveys
PopulationYearUsing AI
German food companies20198%
German business overall202641%

Both cover companies with at least 20 employees. Different years and populations: this is not a food-industry trend line. bitkom.org · bitkom.org

Talent upstream

Experience worth passing on

In the EU’s 2023 farm-manager data, 61.0% were aged 55 or over, 5.6% were under 35, and 71.4% had practical experience only. These describe farm managers, not mill employees. ec.europa.eu

Austria had the EU’s highest share of young farmers: 24.1%. Here, “young” means under 40. ec.europa.eu

Leadership implication: capture experienced people’s judgment and make it easier for the next generation to learn.

In your own words

“Not a single one of the 100 or so grain industry companies taking part was able to say that they’d had a successful year.”

Scott Wellcome · GoodMills · 22 December 2025

In the same account, Wellcome said AI had improved quality controls for incoming grain and enhanced production efficiency. goodmills.com

“The future of milling will revolve around people who can combine modern production, new technologies and the human element.”

Susanne Richter · GoodMills · 22 December 2025

GoodMills announced its Grain IQ academy for 2026, starting in the Czech Republic with UK Flour Milling: learning in an app, practice in the mill and support from a mentor. goodmills.com

From field to flour to food

Classical industries get remade too

The frontier already reached the mill. Now the opportunity is to connect what happens upstream, in production and at the customer's door.

Upstream · the field

Intelligence joins the growing season

John Deere’s second-generation autonomy system uses 16 cameras and can be retrofitted to certain existing machines, as announced at CES 2025. prnewswire.com

About 18% of EU farms used precision-farming technology in 2023 — but managed roughly 44% of agricultural land. Adoption by farm count and by land area tell different stories. ec.europa.eu

The JRC already combines Earth observation and crop models in its monthly MARS yield forecasts. publications.jrc.ec.europa.eu

GoodMills supports the Sustainable Wheat Initiative’s goal of 30% lower wheat and flour emissions by 2030. Around 80% of wheat’s footprint arises on the farm, according to Milling and Grain. goodmills.com · millingandgrain.com

At the mill · the process

The frontier is already grinding flour

Bühler introduced Mill E3 in 2019. Its first operational plant opened at Whitley Bridge in 2021, with more than 15,000 data points and, according to Bühler’s case study, one operator for the whole mill. buhlergroup.com · buhlergroup.com · buhlergroup.com

SmartMill has prototype solutions making autonomous decisions on grinding gaps and flour moisture. A fully self-learning mill remains the programme’s ambition. world-grain.com

Launched in May 2025, the SORTEX AI700 uses convolutional neural networks to inspect colour, shape and texture, initially separating barley, wheat and rye from oats. buhlergroup.com

Downstream · the customer

Food plants become the next robotics story

US food-sector robot installations grew 30% in 2025, while automotive installations fell 1%, according to the International Federation of Robotics. ifr.org

Nestlé announced 16,000 role reductions in October 2025 over the following two years, alongside a focus on automation and faster decisions, as reported by Food Dive. fooddive.com

Bimbo Bakeries’ demand forecasting incorporates weather, local events and information from routes and retailers. automationworld.com

TechCrunch names Unilever and Land O’Lakes among Microsoft Frontier Company’s early partners. techcrunch.com

Making more of what is already there

The surplus nobody could see

DairyReporter reports that Nestlé’s Zest pilot identified 4.8 tonnes of edible surplus on a single production line. Across the project, 201.9 tonnes were redistributed to people. The single-line finding and the project-wide redistribution are different measures. dairyreporter.com

Interpack 2026 · adjacent food processing

What equipment makers now put in writing

Manufacturer claims for specific equipment and comparison conditions; not independently audited savings at GoodMills.

Bühler’s stated energy performance
EquipmentManufacturer claim
Lucent cocoa roasterAt least 20% less energy per batch than the preceding Tornado roaster
CompactMixUp to 30% less energy than conventional models
TurbuBake Smart ovenUp to 15% less energy consumption
OptiBake wafer ovenUp to 50% less energy than conventional gas-heated systems

Bühler release · 7 May 2026. buhlergroup.com

A possible destination

What a winning European miller looks like in 2030

Keynote interpretation drawn from the evidence below. This is a strategic scenario, not a measured outcome or a promise that every element will be deployed by then.

The operating model, grounded in what already exists
LayerThe opportunityEvidence today
The millSet the desired flour characteristics; let connected systems adjust and benchmark performance across sites.SmartMill prototypes already adjust grinding gaps and moisture. Bühler’s Yield Management System compares performance and recipes across plants. world-grain.com · buhlergroup.com
People at the millMore output per operator, with people focused on product quality, oversight and continuous improvement.Whitley Bridge demonstrates a mill operated by a single control-room operator. buhlergroup.com
Grain intakeDeep-learning inspection becomes an ordinary part of protecting quality and gluten-free integrity.The CNN-based SORTEX AI700 is a launched product, with an initial application separating other grains from oats. buhlergroup.com
ProcurementCombine commercial judgment with weather, satellite observations and crop-model forecasts.The JRC MARS Bulletin already supplies monthly EU crop-yield forecasts. publications.jrc.ec.europa.eu
Supply baseBring farm-level agronomy and carbon data into procurement and the emissions ledger.GoodMills has joined the Sustainable Wheat Initiative; most of the wheat footprint arises on farms. goodmills.com · millingandgrain.com
Raw materialPlan separate routes for eligible new genomic varieties and organic grain, with clear traceability.The NGT framework is expected to apply from mid-2028; NGT plants are excluded from organic production. europarl.europa.eu · consilium.europa.eu
Product developmentBuild margins through functional, plant-based and clean-label ingredients alongside commodity flour.GoodMills Innovation already develops functional ingredients and plant-based solutions. goodmills.com
Customer interfaceConnect milling decisions to the demand signals of bakeries, stores and delivery routes.Bimbo Bakeries uses forecasting inputs including weather, local events and route-driver observations. automationworld.com
LeadershipLeaders learn the tools themselves and make workflow improvement a continuing part of work.Bühler’s SmartMill account identifies satisfaction with the status quo as the central adoption obstacle. world-grain.com

Seven countries, one GoodMills

Local knowledge. Shared possibility.

Different harvests, brands and starting points. A common capacity to learn from each other.

Local expertise · shared ambition

Austria

3 mills · Schwechat, Rannersdorf, Graz-Raaba

Footprint in the FY21/22 company presentation, including the organic mill. The Schwechat project will consolidate organic production currently at Rannersdorf.

assets.new.siemens.com

Brands in everyday life

Fini’s Feinstes · Farina · Küchenperle · Eselmehl

Regional brands are supported by the secondary LLI profile.

goodmills.com · de.wikipedia.org

Investing in the next mill

Around €25 million is being invested in the new organic mill at Schwechat, with completion planned for 2027 on the existing site.

ots.at

Scale is already the story

Austria had 92 mills in 2020. The 5 largest produced 63% of its flour.

landschafftleben.at

A strong next generation

24.1% of farm managers were young farmers, under 40, in 2023: the highest share in the EU.

ec.europa.eu

The line for this table · keynote interpretation

Austria can bring its next generation of farmers and its next generation of mills into the same learning story, then share it across ONE GoodMills.

Leadership in practice

Four bridges. Five commitments.

The keynote meets the leadership commitments you already share. The next step is to make them tangible in everyday work.

Preparing “people and organisations for a world we do not yet know.”

Progress is not linear; AI is not hype

A 131-day post-2023 capability doubling calls for leaders who keep revisiting what is possible. Make change part of the operating rhythm.

Shaping change
metr.org

Frontier becomes mainstream

With an approximately 8-month frontier-to-consumer-hardware lag, focus the team on a useful outcome and revisit the route as the tools improve.

Creating focus & clarity
epoch.ai

Always learning, with digital hygiene

WEF expects 39% of existing skills to change by 2030. Create space for practice, safe data use, thoughtful review and learning from mistakes.

Promoting development
weforum.org

Leaders use the tools themselves

Only about 6% of McKinsey respondents are AI high performers. Hands-on leaders can redesign workflows and retain clear ownership of the outcome.

Setting an exampleTaking responsibility
mckinsey.com

Leadership Commitments: exact wording supplied by GoodMills in the client brief; internal, not a public framework.

The leadership gap, made visible

Different studies, populations and measures — kept distinct
EvidencePublished resultWhat it means here
Safety to suggest new ways of working with AI81% vs 67%
Leaders vs employees		The room may feel less safe to your team than to you. microsoft.com
Managers make space for AI experimentation78% vs 59%
Leaders vs employees		Make permission visible in everyday work. microsoft.com
Sharing tips, learnings and mistakes61% vs 36%
Frontier Professionals vs other AI users		This comparison concerns AI user groups, not seniority. assets-c4akfrf5b4d3f4b7.z01.azurefd.net
Productivity in a customer-support rollout14% average · 34% novice / low-skilled · minimal for expertsBrynjolfsson, Li & Raymond’s study measures support agents, not milling productivity. Use it as a reason to test how expertise can spread. nber.org
Skills and jobs by 203039% of skills transformed or outdated
170M jobs created · 92M displaced · net +78M		WEF employer projections cover multiple economic and technological trends. This is not an AI-only jobs forecast. weforum.org
Exposure in advanced economies~60% of jobs may be affected by AIIMF estimates exposure, which includes augmentation. It is not a forecast of jobs lost. imf.org
Shadow AI and breaches20% of studied organisations
As much as USD 670K additional breach cost		IBM reports breaches linked to unsanctioned AI. Give people useful approved tools, clear data boundaries and review. ibm.com
Psychological safety in manufacturing51 work teamsEdmondson found that learning behaviour mediated the relationship between psychological safety and team performance. Help people surface mistakes while they can still become lessons. web.mit.edu

The future-historian's lens

Five theses, mapped to GoodMills

Aragorn's forecasts and interpretations, offered as a hopeful frame for action. Their timelines are his own, distinct from the measured evidence above.

Speaker-supplied forecasts and interpretations. The linked sources provide present-day context; they do not validate or endorse Aragorn’s forecasts. These are possibilities to prepare for, not GoodMills targets or established outcomes.

Aragorn’s forecast

The Exponential Blindspot

Capability will compound faster than familiar planning cycles allow for.

METR measured a post-2023 doubling time of 131 days for its 50%-reliability task horizon. This measures agent capability on a task suite, rather than business results.

metr.org

A hopeful future for milling

A mill that keeps learning can turn each improvement into better grain use, more consistent flour and more room for invention. Build a learning rhythm that updates as the tools improve.

Aragorn’s forecast

The Robotics Revolution

Around 150 million humanoid robots by 2030; around 600 million by 2035.

For scale, Counterpoint reports more than 22,000 humanoids shipped globally in H1 2026, up nearly 300% year on year. Aragorn’s long-range forecast is a separate, much more expansive scenario.

counterpointresearch.com

A hopeful future for milling

Imagine more of the lifting, carrying and routine inspection handled by machines, with people developing safer processes, better products and stronger customer relationships. Start with a useful, measurable job inside the mill.

Aragorn’s forecast

The Great Liberation

30–70% job displacement by 2035: the end of forced labour, not the end of purpose.

This is Aragorn’s forecast and interpretation. GoodMills already names a purpose that reaches beyond any job description: “Nourishment for a better, plant-based future.”

goodmills.com

A hopeful future for milling

A future with less compulsory routine can make more room for learning, care and creation. For GoodMills, the hopeful destination is nourishing people with less waste and helping colleagues share in that progress.

Aragorn’s forecast

Homo Nexus

Human–AI convergence: leaders running teams of hundreds of agents within 2 years.

Aragorn sees intelligence becoming something we work with continuously. The practical seed is already visible in research: an AI assistant improved novice support-worker productivity by 34%, with minimal impact on experienced experts in that study.

nber.org

A hopeful future for milling

Picture a leader exploring harvest scenarios, mill performance and customer needs with a team of agents, then setting direction with the people closest to the work. Make judgment, learning and accountability part of that expanded team.

Aragorn’s forecast

The Prediction Establishment is Broken

Aragorn argues that institutional forecasts can miss exponential change by orders of magnitude; resilience will matter more than a single consensus prediction.

The grain-market analogy is familiar: GoodMills’ own Scott Wellcome described deep uncertainty about growers’ next decisions and the industry’s outlook. Treat uncertainty as a condition to prepare for.

goodmills.com

A hopeful future for milling

A resilient miller keeps options open: adaptable operations, trusted supplier relationships and people who learn together. Build a business that can flourish across several plausible futures, and revise the forecast when the evidence moves.

Follow the evidence

Sources & verification

Sources from the research files, fetched on 7 September 2026, with publication dates and measurement limits kept alongside the claims.

Research fetched and links checked 7 September 2026. Original reference years and source qualifications retained. Some automated rechecks were unavailable.

GoodMills & LLI
Milling & agri
Capability & benchmarks
Adoption & leadership
Robotics
Regulation
Critiques

Capability progress does not guarantee economic returns. These perspectives examine productivity, implementation and reliability.