Claude edges enterprise APIs

Menlo’s data says Claude is edging out OpenAI on enterprise APIs. Why? Vertical, compliance-first agents. Excel sidecar with cell-level provenance, MCP into LSEG/Moody’s, and baked-in DCF/coverage skills. Audit wins. Vibes don’t.

I’ve yet to meet a CFO who cares about personality. They want a receipt.

Cell lineage, input hashes, model/version pins, and a rerunnable plan. If Finance can hit rerun next quarter and get the same DCF, that’s trust.

Architecture that’s winning looks boring in a good way.

Claude with MCP connectors into LSEG, Moody’s, Snowflake, and internal services. Scoped tokens, JIT permissions, and an Excel sidecar that annotates every cell with source, prompt, and policy.

Reproducibility isn’t magic, it’s plumbing.

Versioned prompts in git, frozen tool schemas, seed control, and outputs logged as artifacts. Tie each cell to a source doc and a tool call with arguments. Now you have a story an auditor can follow.

Why Claude here? Anthropic leaned into compliance-native agents.

Prebuilt skills for coverage notes and DCFs, step-by-step rationales you can inspect, and a bias toward least-privilege access. That beats a 2% win on a benchmark you can’t reproduce anyway.

The moat isn’t model magic. It’s permissions, provenance, and predictability.

Design for the auditor first and the operator will thank you. Try the reverse and, well, Legal will book a meeting.

As a founder, I ship for the spreadsheet and the security review at the same time.

Where can you add provenance by default? What is your least-privilege story? My bet: reproducible outputs will beat flash in every enterprise cycle this year.