In May 2026, Anthropic passed OpenAI in enterprise AI adoption for the first time. Ramp’s AI Index — which tracks paid AI subscriptions across more than 50,000 US businesses — showed Anthropic at 34.4% of businesses versus OpenAI at 32.3%. A year earlier, Anthropic sat at roughly 9% share. That kind of climb in twelve months doesn’t happen because of marketing. It happens because a large number of technical buyers made a deliberate procurement decision.
Understanding why they made it is more useful than the headline number.
xychart-beta title "Anthropic Business AI Adoption — Ramp AI Index" x-axis ["May 2025", "May 2026"] y-axis "Share of US Businesses (%)" 0 --> 40 bar [9, 34]
What the Ramp Data Measures — and What It Doesn’t
Ramp’s AI Index captures real spending decisions — which AI providers companies are actually paying for — not survey responses about what companies plan to use. It almost certainly undercounts free-tier and trial usage, which means OpenAI’s substantial advantage among individual developers and experimenters doesn’t show in this data.
What it does show is enterprise procurement behavior: where security reviews got cleared, where legal approved vendor data handling practices, where IT signed contracts that had to pass a compliance checklist. Developer-led adoption and enterprise procurement are different signals. Developer adoption reflects what’s easiest to start with and what has the best documentation. Enterprise procurement reflects what works reliably in production, at scale, with sensitive data.
Anthropic’s climb from 9% to 34% is an enterprise procurement signal. It’s worth taking seriously as such.
The Context Window Advantage That Compounds in Production
Claude’s extended context window — up to 1 million tokens in current enterprise configurations — is not a benchmark talking point. It’s a practical capability that changes what an enterprise application can actually do.
For knowledge-work applications that require reasoning over large documents — contracts, compliance filings, technical documentation, audit trails, clinical notes — fitting the full context into a single pass changes the quality of the output. Retrieval-augmented generation is a reasonable engineering workaround for shorter context windows, but it adds a pipeline: a retrieval layer, embedding maintenance, and debugging for cases where the retriever didn’t surface the right chunk. With a context window large enough to fit the whole document, that pipeline doesn’t exist.
For enterprise use cases in financial services, legal, healthcare, and insurance — where documents are long and accuracy requirements are high — this is not a marginal improvement. It’s a different category of capability. It changes what kinds of applications are feasible to build and what failure modes you’re engineering around.
Why Sensitive Data Work Shifted to Anthropic
Enterprise AI adoption has stalled in several organizations not because of cost or raw capability, but because legal and compliance teams couldn’t get a clear answer about what happens to data after it’s processed — whether it’s stored, used for model training, or accessible to the vendor in any form that creates audit or regulatory exposure.
Anthropic’s enterprise plan has been explicit about training data practices — stating that enterprise customer data is not used for model training — and has invested in compliance certifications that matter to regulated industries. This isn’t unique to Anthropic; OpenAI has moved in the same direction over time. But Anthropic was earlier to market with a documented, auditable answer to the compliance question. When the security review is the bottleneck to deployment, being first with a credible answer matters.
This dynamic mirrors something I’ve seen in platform decisions across long engagements. At H&R Block’s San Diego technology division, the engagement was narrow and focused: re-architect the SOA for TaxCut 2007 from scratch, absorbing a legacy rules engine and restructuring the application services for the next product cycle. The lesson from that kind of targeted, high-stakes build is that the core platform choice — the architecture the whole system depends on — has to hold up under real conditions: seasonal traffic spikes, consumer data at scale, regulatory constraints on data handling. Choosing wrong at the platform level is expensive to undo. The same evaluation framework applies to AI model selection in 2026.
What Enterprise Leaders Should Actually Compare
The benchmark question — which model scores highest on which leaderboard — is the wrong evaluation for most enterprise contexts. Benchmarks measure performance on test sets, not production reliability on the specific data types and query patterns your application actually runs.
The right evaluation criteria for enterprise AI platform selection are:
Does output quality hold on the specific document types and task patterns your application needs? A general benchmark won’t tell you this. A structured evaluation on your actual data will.
Does the provider have a credible, documented answer to your compliance and data governance requirements? Not a marketing claim — a legal commitment backed by certifications your compliance team recognizes.
Does the context window cover your actual use case, or does it require an engineering workaround that adds cost and failure modes?
Can you get contractual stability — model version commitments, uptime guarantees, data handling obligations — from a vendor with a realistic path to remaining operational over your application’s lifecycle?
Anthropic’s position in Ramp’s data reflects organizations that ran a version of that evaluation and came out on the Claude side. Whether your organization reaches the same conclusion depends on your actual requirements, not on the market share number.
