Franklin Templeton uses OpenAI to support investment research

Asset managers cannot treat AI like a consumer recommender. Katrina Dudley, SVP of Public Market Investments at Franklin Templeton, describes moving from AI that automated human-led steps to systems that surface insights for human judgment — starting with equity research workflows built partly with Codex.

Early experiments that dumped data into a database failed until agents and a semantic layer modeled how the firm's information connects. The ambition is governed, reusable intelligence spanning research, risk, compliance, and trading perspectives for portfolio managers.

Diversity of opinion and explicit risk surfacing are design requirements, not nice-to-haves.

We're taking the anti-personalization point seriously. Most consumer AI learns what you click and shows more of it. Dudley wants the opposite for investing: if a manager favors one research provider, the system should still surface dissenting notes and identified risks. "A great investment idea is one that is well reasoned, but it also has analyzed the risks," she says. That is a values choice encoded in product behavior.

The operating model matches: hackathons with Codex, Investment AI champions pushing experiments toward production with security review and data controls, practical labs soon after new tools launch, and leadership cover to innovate inside guardrails knowing not every bet will work. The older sequential analyst workflow — first-pass work, earnings calls, competitor maps — becomes continuous intelligence plus judgment, with a longer-term hope that analysts spend more time in the field talking to people the models cannot replace.

I want regulated firms to build semantic layers before they trust chat over the portfolio. The downside if "connected intelligence" ships without hard controls: leakage across Mandates, hallucinated risk flags that look authoritative, or silent homogenization of views dressed up as breadth. Useful investment AI widens the debate and names the downside. Anything else is just a faster confirmation bias engine.

OpenAI published the customer story on November 5, 2026. Franklin Templeton frames Codex as one tool among several in an ongoing build toward production-grade research workflows rather than a finished platform launch.