A daily automated research routine has been running quietly since September 1st, scanning for new institution–vendor relationships and opening a pull request each morning rather than touching the live site directly — deliberately, so a human reviews every finding before it reaches Supabase. Today was the first time those PRs were actually worked through: two days of automated findings, spot-checked and merged, taking the pilot from 93 to 121 institutions and from 117 to 160 documented vendor relationships.
What "review" actually meant here
Not a rubber stamp. Per the project's own verification checklist — written after the Citigroup incident — a sample of the automation's highest-leverage claims was independently re-fetched before anything was inserted: the Google Cloud press release backing three institutions at once (CME Group, BNY, Macquarie Bank), the Anthropic "finance agents" page backing four (Citadel, Carlyle Group, Walleye Capital, BNY again for a second family), the BPCE interview claiming a four-vendor broker architecture, and a Japanese brokerage's MCP-server launch that a direct fetch couldn't reach (403), corroborated instead through four independent Japanese tech-press outlets. Every one held up. The automation itself had already done the harder work: two of its rejected leads were an institution's investment stake in a model lab and a data-center financing deal, both correctly excluded under the same investment-isn't-usage rule from the Citigroup writeup.
A new record for simultaneous model families
BNY (Bank of New York Mellon) now has three confirmed families — Google, OpenAI, and Anthropic — the first institution in the sample to reach three via three separate official sources rather than one multi-vendor announcement. Wells Fargo, HSBC, and Commonwealth Bank of Australia all independently reached three confirmed families in this same batch. None of this is a "record" claim we're making without checking — see the Citigroup post for exactly why that check matters now.
The more interesting pattern: models are showing up inside connectors, not just chat windows
Several of today's additions aren't "employees use ChatGPT" stories — they're infrastructure. Verisk and Monex Securities both shipped official Model Context Protocol (MCP) connectors putting their own data (insurance underwriting signals, brokerage account balances) directly inside Claude or ChatGPT's own interface, rather than building a separate AI product. Groupe BPCE described, on the record, an internal "LLM broker" that routes requests across OpenAI, Google, Anthropic, and a self-hosted Mistral instance depending on the task — the multi-vendor strategy the family-not-vendor-name methodology was built to measure, now something an institution describes as a deliberate architecture rather than something we infer from separate press releases.
What didn't make it in
Two Nikkei-sourced relationships (Japan Post Bank, Sumitomo Life) date to 2023–2024 — old, but included because the institution or vendor pair was otherwise completely absent from the sample; a stale-but-real relationship is still more accurate than a gap. One finding, Nippon Life's Microsoft 365 Copilot pilot, is marked Probable rather than Confirmed: a single trade-press source, not yet corroborated. And a longer list — FactSet's Claude Code adoption chief among them — was left at Unverified or excluded outright for category fit or thin sourcing, detailed in the research PR itself rather than repeated here.
See the full data table for every row and its source, and the portfolio simulator if you want to see what this batch does to the concentration numbers for a specific slice of the sample.