Bühler Group

Bühler's AI is inside the machine, not the office. The models that matter most ship inside capital equipment its customers already buy. The evidence trail stretches back to 2018, and none of the published results pass a before-and-after test.

Section one

What problem does Bühler's AI actually solve?

Raw materials account for up to 80 per cent of the cost of grain milling. A model that keeps more good product from being thrown away with the bad goes directly to the processor's margin. That is the economic logic behind every AI sorter Bühler has built.

The food safety problem is narrower and more urgent. Aflatoxin clusters in hotspots, so a laboratory sample misses it. Mycotoxin contamination of as few as two kernels in 10,000 can make a batch unfit for purpose, and climate change is increasing the frequency of infected harvests. Allergens carry the same mathematics: a single barley kernel in a certified gluten-free oat lot creates a product liability event.

These are problems where a statistical sample is structurally insufficient, every unit must be inspected, and inspection must happen at line speed. That is what a camera and a trained model can do that a person cannot.

Inside Bühler, the problems the AI addresses are more prosaic. At Uzwil, 600 products each had their own paper drawings, bills of material and work instructions. Across 140 countries, content had to exist in multiple languages. The Squirro partnership, announced in November 2025, targets knowledge access across the business. These are digitisation and retrieval problems, not transformation ones.

Section two

Which systems are running, and how do we know?

Of 19 AI or AI-adjacent systems in the public record, five are running at a named site, eight are offered commercially with no published installation counts, and six are piloted, announced, demonstrated or gone quiet. An announcement is not a deployment, and this register treats them separately.

Scroll the table sideways to see every column.

Source: GAIF evidence register, built from Bühler media releases, partner case studies and trade press. Period: April 2018 to September 2026. "Named site" means the record names a customer location or a Bühler facility. "Offered commercially" means the product page is live but no installation count is published.

Section three

Where does Bühler's AI work best, and why?

The pattern in the successful deployments is consistent. Every mature system targets a problem where the right answer can be confirmed cheaply: a laboratory can verify whether aflatoxin fluoresces, a person can label a barley kernel in an oat stream, a sensor can measure rice whiteness. Bühler has deployed AI first where ground truth is cheap and an error is costly.

The second pattern is commoditisation by one category at a time. SORTEX AI700 launched in May 2025 for oats only. Merlin Ai launched in 2021 for wheat and rye only. LumoVision launched in April 2018 for maize only. Each model was trained on a single, manageable labelling task before the team extended it. The roadmap is visible; the pace is deliberate.

The third pattern is infrastructure before product. Bühler Insights launched in September 2018. The deep-learning flagship arrived in May 2025, seven years later. The connectivity layer, the data science teams and the labelled image libraries were built first. Organisations now surprised that their generative AI pilots do not scale usually skipped the equivalent of Bühler Insights.

Section four

What did each outcome figure actually claim, and does any reach evidence?

Nine outcome figures appear in the public record across eight years. All come from Bühler or its partners. None is independently verified. None compares a production line before and after on measured output.

Tiers: T1 Controlled, T2 Before and after, T3 Estimated, T4 Asserted. A vendor's own figure is T4 until something independent confirms it. Source: GAIF evidence register, September 2026.

Section five

Who governs the AI, and who is conspicuously absent?

Bühler has no Chief AI Officer. No Head of AI or Head of Generative AI appears in any release, annual report or executive listing reviewed for this study.

AI authority is held by three people. Ian Roberts, Chief Technology Officer since 2011, spoke for Bühler at the Microsoft partnership launch in 2018 and at the Squirro signing in November 2025. Vidor Kapy, Chief Information Officer and in Bühler IT since 2000, owns the SAP estate, the digital workplace and the Squirro deployment. Mark Macus, Chief Financial Officer since September 2019, was present at the Squirro signing, which signals budget sponsorship of the enterprise AI layer.

The product models are governed by the SORTEX business in London: Stuart Bashford, who held the Chief Digital Officer title in 2019 Microsoft material and later became Head of R&D at Bühler SORTEX; Melvyn Penna, product manager for both Merlin Ai and the AI700; and David Kinsella, Head of R&D Digital Technologies at the time of Merlin Ai's launch in 2021.

Samuel Schär became Chief Executive Officer on 1 January 2026. His first public statement on taking office does not mention AI. Mike Häfeli, who leads Grains and Food, the segment where the product AI sits, has made no public AI statement. No HR leader, general counsel or data protection officer appears in any AI announcement.

Section six

Where does Bühler depend on something it does not govern?

The service layer runs on Microsoft. Bühler Insights, the platform running all digital services, is built on Azure IoT Hub. Bühler writes the algorithms. Microsoft governs the cloud, the IoT services and their lifecycle. The public record contains a worked example of what that lifecycle risk means.

In September 2018 Bühler put grain traceability on Azure Blockchain Workbench with Whitworth Brothers, a UK miller. Microsoft retired the related Azure Blockchain Service on 10 September 2021. No later Bühler report on the grain ledger was found. The ledger went quiet.

The enterprise layer runs on SAP's release cadence. SAP Digital Manufacturing is a public cloud product. SAP's own December 2025 case study states that Bühler can be confident it is always on the latest technology because SAP keeps releasing innovations, especially in AI. That sentence also describes who decides what changes and when. The natural-language shop floor Bühler expects will arrive when SAP ships it.

The new generative layer runs on a model nobody has named. Squirro describes its platform as model-agnostic and documents connectors for Azure OpenAI and for open-weight models. Which model Bühler runs, where it is hosted and who retrains it are not public.

The mirror image applies to Bühler's own customers. A food processor running an AI700 runs a sorting model trained by Bühler on Bühler's labelled images, with Bühler deciding when the model is extended to new commodities. The processor is in the same position relative to Bühler as Bühler is relative to Microsoft and SAP.

Section seven

What went quiet, and what does the absence tell us?

Three disclosed programmes have no published follow-up. The grain traceability ledger (2018) went quiet when the underlying Azure service was retired. TotalSense, a rice quality tool launched alongside Bühler Insights in September 2018, has no later record. Mill E3, a data programme at a UK site framed in September 2021 as the precursor to a self-adjusting mill, is still described on Bühler's SmartMill page as a journey.

One conspicuous absence has no announced programme: die casting. Bühler has taken orders for 50 Carat megacasting solutions and describes stabilising those cells through standardised platforms with real-time closed-loop shot control under a single automation layer, with process engineers deployed on site after commissioning. In its highest-value process, the stated answer is engineering and people. No AI claim for die casting appears in the public record.

1. April 2018LumoVision launched with Microsoft at Hannover Messe

   Aflatoxin sorting for maize, using UV cameras and cloud risk models. First AI product in the current series.

2. September 2018Bühler Insights platform launched at Microsoft Ignite

   Azure IoT platform for all digital services. Bühler stated it could connect more than 85 per cent of its solutions. Also launched: MoisturePro, GrainiGo, TotalSense, grain traceability ledger.

3. March 2021Connected rice mill announced; Riseria Taverne pilot

   Machine-learning grading of shine, smoothness, whiteness and broken rice. Described by Bühler as a world first in connected milling.

4. September 2021Mill E3 data programme announced; Azure Blockchain Service retired

   Mill E3 framed as the data foundation for a self-adjusting mill. In the same month, Microsoft retired the service underlying the 2018 grain ledger. No later record of the ledger was found.

5. May 2025SORTEX AI700 launched: deep learning, oats first

   Convolutional neural networks trained on millions of labelled images. First deep-learning sorter in the product line. Launched for oats only, with other commodities on the roadmap.

6. November 2025Squirro Enterprise AI Platform partnership announced

   Graph intelligence and agentic workflows across Bühler's businesses. First co-developed use cases described as pending pilots. The model running under the platform was not named.

7. January 2026Samuel Schär becomes Chief Executive Officer

   Schär's first public statement on taking office does not mention AI. The first AI decision of his tenure is not yet in the public record.

Section eight

What should a CXO do with this?

• Buy AI inside an asset the organisation already owns. The budget line exists, the owner is named, and the ROI argument is attached to something physical.
• Before signing any AI contract, request before-and-after evidence from a named customer site, not from the vendor's own case study.
• Write retraining disclosure, model versioning and exit terms into every platform contract. Bühler's grain ledger went quiet when the underlying platform was retired. The contract that prevents that is not the AI contract: it is the platform contract underneath it.
• Name one owner for enterprise AI. Three signatures at a signing is not ownership. When the model changes and something goes wrong, there is no one on record as responsible.
• Map which AI systems in the organisation sit in the exposure gap before the next contract renewal: high operational dependency, low governance.

Global AI Forum · GAIF-DP-001 · v1.0 Research current as of September 2026 gaiforum.com