01 · THESIS
Evidence first, workflow second, architecture third, product fourth.
The frontier has moved from gen-AI pilots to production agentic systems. The reference deployments field teams must know cold: JPMorgan Chase (LLM Suite, ~200,000 users onboarded in 8 months, 450+ use cases, roadmap to 1,000 by end-2026), Morgan Stanley (“To date, 98% of Financial Advisor teams have adopted the Assistant”), BBVA (ChatGPT Enterprise to all 120,000 employees), Wells Fargo (Fargo on Google Cloud, 245.4M interactions in 2024 with zero PII sent to the LLM) and HSBC (“two to four times more financial crime… 60% fewer false positives”). ✓ VERIFIED
Google Cloud’s winning wedge in FSI is the full stack under strict data control— and the commercial reality is multi-vendor: the same bank runs Google, OpenAI/Microsoft and Anthropic simultaneously. The seller’s job is to anchor the data + agent + security layers to Google even when models are mixed. Positioning discipline beats product claims. ◆ ANALYSIS
The four-stage maturity model
| Stage | Typical implementation | Financial-services reality | Production label |
|---|---|---|---|
| assistive | Search, summarisation, drafting, coding and agent assist | Broadest production adoption and clearest productivity evidence | Production — assistive |
| bounded-agent | Agents complete multi-step tasks through approved tools, with human checkpoints | Entering production in onboarding, servicing, research and internal operations | Production — bounded agent |
| multi-agent | Specialised agents plan, retrieve, analyse, validate and prepare actions | Production evidence exists, but disclosure is limited and controls are evolving | Limited production — multi-agent |
| autonomous | AI executes material decisions or transactions without routine human approval | Rare in regulated, customer-impacting workflows; generally inappropriate where explanations, suitability or legal accountability are required | Experimental — autonomous |
Reading this site
Every claim carries one of three labels — ✓ VERIFIED fact traceable to a named source, ◆ ANALYSIS the author’s positioning, ◇ COMPANY-STATED self-reported and rarely independently audited. Every case carries an evidence grade from VERIFIED PRODUCTION down to UNVERIFIED. Missing fields say Not disclosed — nothing is inferred. Where the three source files genuinely disagree, the case shows a Conflicting account note instead of a silently chosen number.