Memo work is the first honest win
Anthropic’s May 2026 financial-services agents target pitchbooks, credit memos, KYC, and statement audits — the documents a committee already knows how to review. Goldman’s work with Anthropic on onboarding and transaction accounting sits in the same band: high document volume, a human still signing.
JPMorgan’s Legal Agentic Workflow is the best published accuracy claim in this set: >95% on contract termination dates versus <3% for a raw GPT-3.5-Turbo baseline. That is a retrieval-and-workflow result, not a general legal intelligence result.
Portfolio agents are still mostly research
JPMorgan’s eight regime agents (Goldilocks / reflation / stagflation / risk-off) beat 60/40 by up to 0.7pp annualized in a two-decade backtest, with lower volatility. That is a research architecture with a published number. It is not a live book. Writing “AI hedge fund” on top of a backtest is how this topic gets junked.
InvestmentNews (July 2026) also notes banks moving agents from research aids into “digital coworkers” in trading, treasury, and wealth — vetting clients, watching annuity maturities. Coworker means a queue and a supervisor, not an unsupervised wire.
The bar that does not move
EU AI Act and DORA want traceability. US model-risk guidance still wants a challenger, a monitoring plan, and an owner. An agent that writes a credit memo is a model. An agent that files a SAR is a model plus a legal person. Neither is exempt because the vendor called it a coworker.