The Failover Economy
I wanted to switch AI agents without starting over.
I fixed that at my desk, then found the same pattern in 30 months of corporate card data.
The numbers
- In May 2026, Anthropic passed OpenAI on Ramp’s index for the first time: 41.01% to 39.54%.
- By June, the gap had widened: Anthropic was at 42.39% and OpenAI at 39.48%. In total, 54.95% of US businesses on Ramp paid for AI.
- Those vendor shares overlap. At least 26.9% of all Ramp businesses paid both, or roughly half of the businesses that paid for AI.
The full monthly series is embedded in this page as JSON and mirrored at ramp-ai-index.json. The scroll story below walks the same numbers.
Why I needed a backup
The handoff problem
When Claude Code’s usage meter approaches 100%, I can keep working. What I lose is the agent that already understands the project.
Adam Schoenfeld had a better answer: switch to another agent and keep going.
I wanted the second agent to inherit the work, not just open the same files. So I made Codex a real backup, then looked at whether companies were doing the same thing at a larger scale.
June 2024
In June 2024, OpenAI appeared on 30.7% of US businesses on Ramp.
Anthropic appeared on 5.7%.
OpenAI looked uncatchable. Anthropic barely registered.
May 2026
Twenty-three months later, Anthropic had passed OpenAI: 41.0% to 39.5%.
The crossover got the attention. The overlap matters more.
The story
The two vendor shares add to 81.9%, but only 54.9% of Ramp businesses paid for any AI.
Those numbers can coexist only if at least 27% of all businesses paid both vendors. In June 2024, that minimum was 3%.
Businesses were no longer choosing one. They were paying for a backup.
What I did about it
That is what I built at my desk. Claude Code and Codex now read the same rulebook, use the same memory files, and leave one another a handoff when I switch mid-task. A symlink points both agents to the rulebook (ln -s CLAUDE.md AGENTS.md). The rest was writing down what the second agent would otherwise have to guess.
Installing Codex took minutes. Making the handoff trustworthy took most of a day. I tested it by starting Codex cold and asking it to tell me the rules before it touched the repo.
I no longer watch Claude’s usage meter like a deadline.
The full walkthrough, including how to keep the shared rulebook current, is Make Your AI Agent Replaceable.
The takeaway
Roughly half of AI-adopting US businesses on Ramp now pay both OpenAI and Anthropic. My desk is one tiny version of the same choice: one project, two agents.
Method
Ramp publishes the share of its roughly 50,000 US business customers with a paid subscription to each AI vendor, plus the share paying for any AI. Vendor shares overlap. The lower bound uses inclusion-exclusion: OpenAI% + Anthropic% can exceed total AI adoption only when some businesses pay both. The true overlap is therefore at least the difference, and may be higher. Ramp’s customer base skews toward companies that put more spending on corporate cards. The data is embedded in this page as JSON: view source, take it, and check the math. I checked all 30 months against Ramp again. They matched, and June 2026 was still the latest month Ramp had released.
Companion essay: Make Your AI Agent Replaceable, the setup that puts this chart into practice at desk level.