Point a language model at a database and you do not get an analyst. You get a fluent reader who has never been told which numbers belong to which argument.
Franklin Templeton already ran a version of that experiment. OpenAI’s customer story on the firm says the earlier setup put data in a database and connected models to it. The answers fell short of what the teams needed, because the models did not understand how the data fit together. The line about numbers and arguments is my shorthand for that gap, not a line the firm published.
The account describes an attempt to rebuild investment research. Portfolio managers are supposed to hold competing views and the risks under an idea in one place. Research, risk, compliance, and trading show up as related inputs, not as four separate piles of files.
Automating the old handoff leaves the blindness intact
Katrina Dudley, senior vice president of public market investments at Franklin Templeton, draws the line that matters. In the OpenAI customer story she says:
The biggest unlock from using OpenAI’s tools is the ability to merge continuous intelligence with human judgment. It has really transitioned because, beforehand, AI was automating our human-led process. Now AI is actually automating the ability to surface intelligent insights, and that intelligence layer plus human judgment is just a fantastic unlock.
This is a customer endorsement, not an independent audit. OpenAI reports no accuracy, time saved, error rate, or return for the work. The distinction is still worth taking seriously. Continuous intelligence, as I read her point, means the work stays live: research, risk, compliance, and trading views kept in relation to one another, instead of being stitched back together after the fact. She is not describing a dashboard that refreshes.
She also says, “We have moved from augmenting our research process to reimagining the investment research workflow.” A lot of what gets called AI automation, in my experience, stops short of that turn. It makes the old handoff faster and leaves the blindness intact.
Agents are useless without a map of meaning
The fix OpenAI describes is two pieces, not a product name. One is agents. The other is a semantic layer, which OpenAI defines as a structured model of what the firm’s data means and how it is connected. Think of it as the legend on the map. Without it, an agent can pull a filing, a risk note, and a trading comment and still not know they are three arguments about the same position.
“We’ve been able to build what we think of as a connected intelligence layer, ” Dudley says. In her account, teams are bringing data into the tools “to surface insights to the people who are making decisions, which are our portfolio managers.” Codex is one of the tools named for developing that layer, alongside others. The product name is the least informative fact on the page.
The verbs are the status report. They do not all point the same way.
The database-and-model stage is the one that failed. OpenAI says agents plus the semantic layer changed the results. Dudley speaks as if the connected intelligence layer is already in use, surfacing insights for portfolio managers.
OpenAI also says the firm wants this kind of help and envisions a governed foundation. Investment AI champions are still working to turn hackathon experiments into production-ready workflows. Practical labs follow soon after new tools launch. Security review, expert support, data controls, and data agreements sit on that path. A layer in the building is not the same thing as every desk running on it each morning.
What the software is forbidden to optimize
The sharpest ideas here are not about models. They are about what the software is not allowed to treat as success.
If a tool notices that someone favors one research provider, the easy next step is to feed them more of that provider and call it personalization. Franklin Templeton says it wants the opposite. Dudley: “We want to make sure that they’re accessing that diversity of research and diversity of opinion.” A system that narrows the inputs will look helpful right up to the day the missing view was the one that mattered.
The second choice is about risk language. Ranking systems often treat caution as friction. Downrank it and the memo feels cleaner. Dudley argues the reverse:
We need AI to understand that a great investment idea is one that is well reasoned, but it also has analyzed the risks and makes our portfolio managers aware of the risk.
Identifying the risks in a note should make the decision more informed, not less visible. That is a design requirement, and it travels past asset management. Anywhere a score can punish the awkward paragraph, the tool will learn to bury it unless you forbid that move.
The part that can leak
An agent that can see research, risk, compliance, and trading can also leak across those walls if the data agreements are sloppy. That is the unglamorous reason the path into production matters more than the demo.
Dudley passes along the leadership line:
Our message from leadership is consistent: we encourage innovation and experimentation with AI within clearly defined risk guardrails. Not every initiative will succeed, but each should generate insights that strengthen our capabilities, inform future decisions, and drive continuous improvement.
Read that as permission to kill projects. The intended end state is a governed, reusable foundation: trusted connectors, portfolio and workflow context, and agent tools that surface insights for analysts and portfolio managers. A clever workflow that lives on one team’s laptop is a demo. A layer other groups can query, under the same controls, is infrastructure.
In the same account, the longer aim is organizational. Make the layer available from analysts up to investment-group leaders, and give analysts more time in the field with outside experts, including executives at private companies. Human reporting stays essential. Software carries synthesis. People go get what is not in the files yet.
What another leadership team should actually copy
The pattern worth stealing is narrower than buying the same tools, and harder.
First, stop automating a sequence the model does not understand. If your data has no shared definition of meaning and connection, more agents will only retrieve faster confusion. The semantic layer is the asset. The model is rented.
Second, write the anti-goals down. Diversity of sources is a requirement, not a preference. Risk language is a feature. Personalization that learns a user’s favorite input and hides the rest is a defect in a decision system.
Third, separate the hackathon from the desk. Pair the people who can build with the people who live with the output. Then put security review, expert support, data controls, and data agreements between “it worked on Thursday” and “it runs the book.” Not every experiment should survive. The ones that die should still teach the next guardrail.
Fourth, measure something the customer story does not. Time to a comparable brief. Risks caught before a committee. Cases where the tool surfaced a dissenting source a human would have missed. Cases where it invented a relationship the semantic layer did not support. Without those counts, connected intelligence is a phrase a vendor and a client can both enjoy.
The firms that get AI agents into real work will look boring from the outside. They will spend quarters on meaning, permissions, and dissent, then let people leave the building to learn what the files still do not know. That is the version of AI automation a business can use on Monday.
Questions worth asking before you copy it
-
Why did the first AI setup fall short?
Data sat in a database with models connected to it, and the models did not understand how the data fit together. OpenAI’s customer story is explicit on that. The missing piece was how the records related. Whether a model could reach them was not the gap.
-
What is the semantic layer, in plain terms?
OpenAI calls it a structured model of what the firm’s data means and how it is connected. That map lets agents relate research, risk, compliance, and trading. Without it, they answer each pile of files on its own.
-
What is actually in production?
A connected intelligence layer the firm says it has built, and is using to surface insights for portfolio managers. Role-specific workflows are still moving from hackathons and practical labs toward production, under security review, expert support, data controls, and data agreements. OpenAI reports no accuracy, time saved, or return figure. What is live is a design. It is not a scored result.
-
What should a non-investment business take from it?
Treat it as a design to test, not a method Franklin Templeton has proved. Build a shared model of what your data means before you staff it with agents. Forbid the tool from narrowing people onto one source. Treat risk and dissent as signal. Keep a governed path from experiment to production, and spend any time the synthesis saves on the human reporting the files cannot do.