Anthropic leads — and doesn't: two AI-adoption reports, two truths, one lesson for the Mittelstand

What happened

The June edition of the Ramp AI Index — run by the finance-platform vendor Ramp on actual card-payment and bill-pay data from more than 70,000 US businesses — puts Anthropic up 2.5 percentage points at 41 percent of firms with a paid AI subscription. That makes Claude, for the first time, the most-procured enterprise AI vendor in the US, ahead of OpenAI; the crossover had already become visible in the May edition. Almost simultaneously, The Register (11 June) reports on an IDC survey (FERS survey, fieldwork March 2026, more than 1,000 organisations): only 19 percent use Claude “extensively”, 25 percent are evaluating it — while OpenAI (around 42 percent) and Google (around 38 percent) lead on extensive use. VentureBeat frames the Ramp finding with three risks that could cost the lead again: compute costs, capacity limits and usage-based pricing.

Reading it

The two reports do not contradict each other; they measure different axes. Ramp measures breadth: does a company pay for the vendor? That is a binary procurement signal from real transaction data — good at making broad uptake visible, but blind to how deeply the tool sits in production. IDC measures depth: how intensively is a model used inside the organisation? That is a self-report that reflects the incumbency advantages of earlier-established vendors. Read together, a consistent picture emerges: Anthropic wins the breadth question — driven above all by Claude Code — but still trails on depth of entrenchment. The structural point for any platform decision: a “market leader” label says nothing about whether a tool is load-bearing. Understanding the metric matters more than the ranking.

What it means for the Mittelstand

The first consequence is defensive: do not choose an AI vendor by a leaderboard. “We have a subscription” and “it carries our production” are two different statements, and an adoption headline blurs exactly that distinction. What matters for your house is not the US market share but the depth of use you can measure yourself.

The second finding is commercial. Anthropic switched enterprise customers to usage-based billing in April; for heavy users that can double or triple costs. A lead built on bottom-up adoption can be clawed back through price. Anyone rolling out Claude Code broadly should model a cost-runaway scenario before usage scales — not after.

The third finding is the data-protection reflex, and it does not hang on the ranking. Whether Anthropic, OpenAI or Google leads — all three are US vendors: the third-country question, including CLOUD Act exposure, persists in every case. For every vendor that becomes load-bearing in your house you need the same contractual base: a legal basis, a data-processing agreement under Art. 28 GDPR with the actual processor, a viable transfer mechanism. Which data path actually applies is a question for your data protection officer — before scaling, not afterwards.

What it means for technical development

Architecturally, adoption figures concern the commercial consumption layer of the agent stack, not the interoperability beneath it. Which vendor currently leads does not change the interfaces — and that is precisely the point for your own architecture: keep the model layer swappable. If Claude Code is your growth driver, the stack should speak standardised protocols (MCP, A2A), so that changing vendor stays a configuration question rather than a rebuild.

The second consequence is about measurement. Reading adoption only from external headlines means steering blind. Capture depth of use internally via telemetry — which workflows run how often and how reliably over which model. That figure, not the Ramp index, tells you where a vendor switch would be expensive.

One concrete recommendation

In this order. First, separate breadth from depth in your own house: list which AI subscriptions you pay for and which of them actually carry production load — those are two different lists. Second, for the load-bearing vendors check the pricing model (seat-based versus usage-based) and model a cost-runaway scenario before usage scales. Third, with your data protection officer clarify the third-country and DPA basis for every load-bearing vendor — the ranking changes nothing here. Fourth, measure your architecture against the question of whether the model endpoint stays swappable (MCP/A2A), so that a pricing or compliance change is a configuration decision, not a rebuild. This article reflects our technical and strategic assessment. It is not legal advice and not a data protection impact assessment.

Sources

• Ramp AI Index — June edition “How much does it cost to be AI-pilled?” (Anthropic 41%, most-procured enterprise AI vendor) (10 June 2026)
• The Register: “Claude is ready for its corporate close-up” — IDC FERS survey, 19% extensive Claude use (11 June 2026)
• VentureBeat: “Anthropic finally beat OpenAI in business AI adoption — but 3 big threats could erase its lead” (May/June 2026)
• TechWire Asia: “Anthropic builds out Claude as OpenAI and Google stay ahead” — IDC context (12 June 2026)