Balyasny Cut a 3-5 Day Merger-Arb Analysis to 30 Minutes With Claude Fable 5

A merger-arbitrage analysis that used to take Balyasny Asset Management’s analysts three to five days now takes about 30 minutes. The $38 billion asset manager built BAMAgent, a proprietary platform that runs Claude Fable 5 across financial-analysis and trading workflows, letting the model plan its own steps, select tools, evaluate evidence, recover from its own errors, and produce a finished deliverable without an analyst in every loop. In Balyasny’s internal evaluation against thousands of real financial tasks, Fable 5 hit an 89.4% success rate versus 86.1% for the prior production model — and solved internal economics problems the firm hadn’t been able to crack before.

What makes the case study worth reading past the benchmark numbers is Chief AI Officer Charlie Flanagan’s framing of safety: not a model-selection question but “a product and operating-model question.” Concretely, that means approved data boundaries, least-privilege access, tool-level permissions on what each agent can touch, logging on every action, human-review requirements on defined categories of work, and escalation paths back to a person — controls built into how the system runs, not a compliance layer bolted on after deployment. Balyasny is now scaling to more than 300 agents doing continuous analysis, and Flanagan describes the shift as one from treating AI as tools to treating it as teammates that improve with repeated use.

That’s one bet on where financial-services differentiation comes from. Forrester has been tracking a different one: banks building proprietary “transaction foundation models” trained on their own behavioral data — Revolut’s PRAGMA, trained on 24 billion events across 111 countries; Nubank’s nuFormer; Visa’s TransactionGPT. Forrester’s read is that the model itself isn’t the moat, the accumulated transaction and outcome data feeding it is. Balyasny’s approach skips that build entirely: adopt the frontier general model, and put the firm’s real investment into the governance wrapper around it. Two regulated firms, two different theories of where AI advantage actually accrues — one in the data, one in the operating model.