Issue 01 · Operating Intelligence
Why Germany and Switzerland are running a different race than the United States — and what a mid-sized firm should do about it, 2026–29.
In January 2025, the Swiss building-technology group Meier Tobler went live on a single SAP core after consolidating multiple ERP systems and straightening its end-to-end processes. A year later the company reported that the expected efficiency gains had not yet arrived. Commissioning costs ran ahead of benefits. Meier Tobler’s project was an ERP consolidation. Operational AI runs through those same core processes and records. Chat tools run fine on top of messy systems. Changing a measured workflow requires a clearly defined process, usable records and someone who owns the result. That work takes as long as it takes, whatever the model labs ship.
Both economies innovate under different pressures. In the United States, headlines centre on frontier models and hyperscale infrastructure. The German federal government’s own framing is sober; the High-Tech Agenda describes Europe as being in a catch-up position and identifies gaps in SME know-how, staff, finance and dependence on non-European components. For German and Swiss operators, the near-term challenge is mainly absorption: connecting capable models to installed systems and governed workflows.
AI capability improves faster than mid-sized firms can absorb it into their productive systems. Many run on legacy software and fragmented data, with thin management bandwidth and real compliance constraints. Six forces are at work, each on its own clock.
Integration, workflow redesign, governance and ROI remain firm-level constraints.
Watched · external clockThe AI Act’s revised timetable is fixed. Inventory, accountability, logging, vendor diligence.
Watched · legislative clockQuoting, engineering documentation, maintenance, QA, finance, procurement, service. Shorter cycles, fewer errors, lower burden — outcomes to test, and benefits only once a baseline proves them.
Driven · firm-ledValue sits in ERP, MES, CAD, CRM, QMS and service logs — fragmented, undocumented, expert-dependent. Connecting AI to these systems is the hard, decisive step.
Driven · firm-ledWhere confidentiality and traceability are non-negotiable, AI needs demonstrable controls. Swiss data centres can support that objective; storage location alone does not resolve access, subprocessors, contracts or cross-border transfers.
Varies by sectorAgeing workforces make knowledge continuity an important use case: preserve expert review, help apprentices learn from governed records. The deadline is succession — KfW counted 231,000 owners planning to close rather than hand over by end-2025, and 532,000 transfers intended by 2028.
Horizon-wideQuelle Regulation (EU) 2026/1744, Official Journal, 24 July 2026. Legal applicability remains system- and role-specific. Runway measured from this issue, August 2026.
Regulation and compute programmes run on legislative and vendor timelines a firm does not control. Data fragmentation and workflow productivity move only when someone inside the firm decides to fix them. Labour redesign runs across the whole horizon; whether its design objectives show up as employment outcomes is a question for a firm’s own measurements.
German SMEs are dealing with skilled-labour constraints, weak demand, energy costs and international competitive pressure all at once. KfW reports that AI use is more common among larger, internationally active and digitally mature SMEs. The pattern is consistent with constraints beyond model access. Managers still have to decide who owns the work, how people are rewarded and whether the firm’s systems can support the integration.
Rammer, Fernández and Czarnitzki, working from Community Innovation Survey data, report that 5.8% of German firms were actively using AI in operations or products in 2019 — a self-reported survey measure of use. In the same data, AI use is associated with roughly €16bn in annual world-first product-innovation sales, about 18% of that total, and AI-supported process innovation with about 6% of total annual cost savings in the German business sector. These are associations from one country during early diffusion; the study does not identify causal effects.
The ifo Business Survey does not specify what counts as AI, so it records what firms themselves consider AI use: on that measure 13.3% of German firms were using AI in June 2023, and 27% in June 2024. Neither can be put in a series with the 5.8% above; that survey counted something else.
The federal cabinet adopted the High-Tech Agenda in July 2025. Its AI chapter sets a policy target that 10% of German economic output be AI-based by 2030. Its measures remain subject to available budgets.
The EU’s InvestAI initiative aims to mobilise €200 billion. On 30 July 2026 the Commission opened the call for up to seven AI gigafactories, backed by up to €10 billion in EU and national funding and expected to draw at least €20 billion in private investment.
A gigafactory adds compute capacity inside European jurisdiction, in time perhaps lower prices, and a wider set of vendors that a buyer under data constraints can consider. The programmes fund capacity. Connecting a model to the firm’s own systems and records, redesigning one workflow and counting the result is the firm’s own work. The Agenda’s target and the measured base do not share a scale: one is denominated in economic output, the other counts firms.
Where can AI produce a measurable change in revenue, engineering throughput, production reliability, quality documentation or administration? Each candidate in the table below stays a hypothesis until a baseline, an owner and a result exist.
| Result area | Candidate workflows |
|---|---|
| Revenue | RFQ parsing and draft quotes |
| Engineering throughput | Documentation and variant analysis |
| Production reliability | Anomaly triage and maintenance support |
| Quality documentation | Audit preparation and CAPA drafting |
| Administration | Invoices, cash forecasts, supplier-risk checks, multilingual service, onboarding |
The table is a menu of places to look. No savings figure belongs in a proposal until a firm’s workflow and software stack have been scoped.
For Swiss firms, trusted deployment is shaped by confidentiality, traceability, professional liability and speed. A firm that can demonstrate all four has something to sell — and the platforms are building the same capabilities for themselves. Microsoft’s EU Data Boundary commitment now covers its Copilot enterprise stack, with carve-outs a careful buyer still has to read (web search queries and, currently, Anthropic models sit outside it). Mistral’s regional endpoints, generally available since August 2026, pin inference and the associated processing to Europe; the company’s own announcement notes limited, safeguarded transfers to sub-processors that may sit outside the region.
Microsoft says it is investing $400 million to expand cloud and AI infrastructure near Zurich and Geneva. It also reported more than 500,000 people reached through skilling initiatives by April 2026. A treasurer at a Swiss SME can read every one of those announcements and still not know what came back to the firm. The company’s telemetry-adjusted model estimated Swiss generative-AI user share at 37.8% of the working-age population in Q1 2026, versus 17.8% globally — a modelled population estimate, several steps removed from any single firm’s workflow ROI.
of the working-age population estimated to use generative AI, Q1 2026.
Microsoft AI Diffusion ReportThe same modelled measure, worldwide.
Microsoft AI Diffusion ReportSwitzerland above the global estimate on this measure.
Derived from the two figuresQuelle Microsoft AI Diffusion Report 2026 Q1 — first-party company research. A modelled, telemetry-adjusted estimate of population-level use.
One Swiss-owned project sits at the far end of the planning horizon. At the Star of Laufenburg, the 1958 origin node of the European interconnected grid, the privately financed FlexBase group is building a redox-flow battery beside a data centre it markets as its Sovereign AI Factory. FlexBase calls the battery the world’s largest, with a planned scale above 1.2 GW and 2.1 GWh in its January 2026 figures. That month the company reported that Swissgrid had approved an 800 MW first-phase grid connection; FlexBase targets operation in 2029.
Swiss FDPIC guidance leaves the controller responsible for processor, subprocessor, security and cross-border-transfer diligence. Swiss storage can be one control, but it does not by itself settle access, jurisdiction, contractual rights or lawful processing. The implications vary across finance, insurance, medtech, pharma suppliers, precision manufacturing, legal work and the public sector.
In higher-trust settings, that shifts the deployment questions toward data access, traceability and the ability to justify a decision.
Supports individuals with email, summaries, translation and notes.
Changes a named workflow and makes before-and-after measurement possible.
Connects governed workflows to core systems; can improve control of institutional learning.
Two risks remain: a firm can fall behind while competitors improve real workflows, or adopt AI without the integration, governance or bargaining power needed to retain value. PwC estimates that 74% of AI’s economic value is currently captured by 20% of organisations.
The first risk: pilot purgatory. AI is announced but never wired to the production function. Shadow tools spread, ROI stays unclear and sensitive-data questions go unanswered.
The second risk is harder to observe. A firm can succeed in adopting AI and still discover that part of the resulting advantage sits outside the firm. That depends on where the workflow runs, who controls the interface, which data and corrections remain portable, and how costly it becomes to switch providers. The question is where the learning and bargaining power accumulate as AI becomes embedded in operations.
OpenAI’s own product page for advertising in ChatGPT describes ads shown to logged-in adults on the Free and Go tiers, states that Plus, Pro, Business, Enterprise and Education are ad-free, and prices opting out of ads in fewer daily free messages. The page names no EU market, and whether an EU rollout would carry additional duties under the AI Act or the DSA is unresolved.
Sam Altman says OpenAI has no wish to “eat every startup”, only to provide the platform. In the same interview he says that intelligence itself is trending towards a commodity, with durable advantage sitting in compute fleets, workflows, integrations and brand.
Satya Nadella calls one possible transmission channel the reverse information paradox. Firms, on this hypothesis, may disclose context, traces, evals and corrections to make purchased intelligence useful. Major providers say business and API data are not used for training by default. Product-specific diligence covers retention, access, feedback, fine-tuning and contractual rights.
Applied Intuition’s Dana launch supplies a physical-AI example. The company says Dana is used by Isuzu and Komatsu and describes its position as a horizontal intelligence supplier. In a promotional a16z interview, a founder described design wins as long relationships that are hard to unwind. That description suggests a switching-cost risk, but it comes from someone selling the product. It leaves open whether Applied owns customer data, whether every manufacturer cedes its learning loop, and how long the lock-in lasts.
Taken together, the cases suggest four questions worth asking before choosing a workflow to build.
Name the workflow result and the person accountable for it.
Record the starting measure and confirm that the workflow recurs often enough to test again.
List the systems and data the workflow needs. Set the terms for access, retention, permitted learning, correction ownership, export and replay. Keep the model layer replaceable where practical, so switching models does not mean rebuilding the workflow.
Make each output traceable to its sources, the process step and human approval. The responsible person must be able to defend the decision. Relax the gate only when the evidence supports it.
Choose one recurring workflow, give it an owner and record the starting measure. Connect only the systems it needs and keep consequential outputs behind human approval until the evidence supports doing otherwise. Meier Tobler’s first year is a reminder that the comparison has to include implementation costs as well as realised gains.
The firm still has to redesign the workflow while retaining its evidence, bargaining power and operating knowledge.
One finding would change our mind: firms that simply bought tools and posted the same measured gains as firms that redesigned workflows. If that pattern shows up, the premise of this publication fails, and we will say so.
Primary sources · verified 27 July–10 September 2026
Macro facts carry sources; structural arguments are patterns and labelled as such. Firm-specific savings require a scoped audit.