The challenge isn't the models. It's the workflows they were never designed to support.
The conversations are usually pretty candid. The models have improved faster than most people predicted. The data is better governed than it was a few years ago. Pilots have gone well. Individual users are genuinely more productive. But operational impact, the kind that changes what a team can actually produce, at what scale, with what consistency, keeps proving elusive.
My view is that this isn't a technology problem. It's a workflow problem. As more institutions adopt AI for financial research, many are discovering that their existing processes were never designed to support institutional AI workflows at scale. And the industry is only now starting to name it clearly enough to do something about it.
The research process in most institutions follows a logic that makes sense: data sourced, transformed, validated, analyzed, communicated. That sequence exists for good reasons. The standards embedded in it took years to develop.
What nobody had to state explicitly was that humans were doing the coordination; navigating between systems that don't talk to each other, reconciling data that doesn't align by default, translating outputs into formats the next person in the chain can actually use. That coordination was invisible in the sense that it was just... how things worked. People knew how to do it. They did it.
AI systems don't navigate seams the way people do. They operate across steps, not within hand-offs. So, when you introduce AI into a workflow built around human coordination, you get productivity gains at the individual task level and friction everywhere in between. Which is exactly what most organizations are experiencing.
The dominant deployment pattern right now is adding a conversational layer, a chat interface, a co-pilot, on top of the existing stack. These tools are useful. I do not want to dismiss that.
But a chat interface helps you think faster inside a single step. It doesn't change the structure of the workflow those steps sit inside.
The data still needs pulling from wherever it lives. The transformations and vintage reconciliations still need to be applied. The outputs still need to be moved into formats that decision makers can use and routed to the right people at the right moment. All of that still runs on the old rails, with human coordination doing the work it has always done.
The bottleneck in financial research has never been analysis in isolation. It has always been the coordination around the analysis, the hand-offs, the translations, the implicit institutional knowledge that determines how things flow. That is where productivity lives or dies. And it is the part that most AI deployments, as currently structured, do not reach.
Not a new workflow. That is not what I am arguing for. The logic of serious research, the sequencing, the validation, the domain conventions about what a trustworthy output looks like, that logic is genuinely valuable. Throwing it out to build something AI-native would be a mistake.
What needs to evolve is how the workflow executes. Not what it does, but how the steps connect to each other. Whether the coordination between them lives in people's heads and informal processes, or whether it is embedded in the research infrastructure itself, in structured data flows, explicit context, connected systems that do not require human navigation at every seam.
That is a different kind of investment than most organizations are making right now. And I think it is the one that separates incremental gains from something that compounds.
The organizations that figure this out will not look like they have built something radically new. They will just be able to do things at a scale and consistency that others cannot quite match, and it will not always be obvious why.
The ones that do not will keep seeing the same ceiling: meaningful AI adoption at the individual level, genuine enthusiasm from users, and persistent difficulty translating that into operational impact at the team or firm level. Not because the technology failed, but because the workflow infrastructure was not ready to support it.
At Macrobond, this is the problem we have been working on. Not AI as a feature layer added to existing products.
We believe the future of institutional AI workflows depends on trusted macroeconomic intelligence with the infrastructure required to support consistent execution.
Connecting workflow infrastructure, data quality, domain intelligence, and execution in a way that reflects how research gets done, and that lets AI operate inside the process rather than adjacent to it.
The research workflows that get us through the next decade will not look completely different from the ones we have now. But they will be built to run without depending on people to hold them together at every seam.
