For decades, institutional banking has operated around a simple assumption: there’s a human on the other side of the relationship. A portfolio manager calls the desk. An analyst logs into a research portal. A trader asks for pricing. The workflows, controls, compliance frameworks, and relationship models across the industry were all built around that idea.
That assumption is starting to change. Increasingly, the first interaction may not come from a person at all. It may come from software acting on the client’s behalf. Institutional banking is entering an agent-to-agent era, in which software acting for the client deals with the bank’s systems and, increasingly, with the bank’s own AI.
AI systems on the buy side are beginning to retrieve research, query pricing, access analytics, and in some cases initiate execution through programmatic interfaces. The human hasn’t disappeared from the relationship. But the access layer is becoming automated.
That isn’t a subtle shift. In certain parts of the market, client-side AI agents are already interacting directly with bank infrastructures. The broader question remains if banks are ready for that interaction as it brings ownership, governance, and infrastructural challenges to bear.
The latest AI trends, brought to you by experts
Get curated insights on the most important—and intriguing—AI news. Subscribe to our twice-weekly Think Newsletter.
A lot of the industry’s AI focus has been internal: research summarization, meeting notes, workflow automation, productivity tools for bankers. That work matters, but it’s not the biggest structural shift underway. The more important development is that institutional clients themselves are adopting AI quickly, especially hedge funds, quantitative firms, and large asset managers that already operate through highly automated infrastructure.
Prime brokerage is the clearest example. Clients already consume position data, financing data, margin analytics, and risk information through APIs and structured interfaces. Moving from traditional APIs to AI-agent-driven access is not a massive leap technically.
The gap is rarely “how do we do it” but “do we know how we did it”. What’s missing is the governance layer.
What happens when the entity requesting information is not a named analyst but an autonomous client agent? Who authorized the access? What permissions apply? What gets logged? What does supervision look like when decisions are increasingly machine-assisted?
Most banks don’t have clean answers yet.
Model Context Protocol (MCP) is an open standard that lets AI agents connect directly to external data sources and services through structured interfaces.
MCP has drawn a lot of attention recently, but the idea underneath the technical language is relatively simple. Instead of building a one-off integration for every system, an AI agent can use a common protocol to find out what a service offers and interact with it.
In practical terms, a client-side AI agent could pull research directly from a bank, check inventory or financing terms, and retrieve risk analytics. It could compare pricing across counterparties and, potentially, route execution automatically. Each step moves the agent further from reading information and closer to acting on it.
Once the interaction is machine-to-machine, the operating model looks very different from a traditional portal or client interface. It becomes a question of permissions, entitlement controls, auditability, and governance at machine speed.
Some firms are already moving in this direction. IG Group has introduced a read-only MCP layer. ThinkMarkets has supported execution-capable workflows. Aiera distributes sell-side research into buy-side AI systems with entitlement tracking built in. Huatai Securities in China has launched dedicated infrastructure aimed at institutional AI integration.
None of this is science fiction anymore. It’s early, fragmented, and unevenly adopted, but it is real.
Institutional banking is fundamentally an information business.
Successful firms tend to have:
- Better context
- Better timing
- Better integration into client workflows
- Faster access to information and liquidity
That dynamic becomes even more important when the “user” consuming your services is software. An AI agent will naturally prefer the platform with cleaner interfaces, better structured data, faster retrieval, and clearer permissions. Over time, ease of machine access may matter almost as much as the underlying relationship itself.
That’s particularly important in prime brokerage, where the product is already deeply operational and data-driven. Today, a hedge fund may have teams manually pulling financing data, inventory availability, and analytics from multiple providers. Tomorrow, much of that comparison process could happen automatically through client-controlled AI systems.
That pivot changes the entire competitive landscape because the differentiator is no longer just the strength of the sales relationship. It becomes the quality of the infrastructure layer sitting underneath it.
Across the major banks, AI investment is happening everywhere, but most of it still points inward. Banks are building research tools and internal copilots, summarizing meetings, automating workflows, and making their own knowledge easier to find. Those capabilities are useful and increasingly necessary. But they are also quickly becoming standard. The harder and less crowded problem is building the governance and infrastructure required for client-agent interaction. That requires systems most institutions were never designed to support:
- Non-human entitlement models
- Real-time permissioning
- Machine-level audit trails
- AI-specific compliance controls
- Supervision frameworks for autonomous workflows
This is one reason client-facing AI has moved much more slowly than internal deployment. The governance burden is materially higher. But that burden is also what makes the capability defensible once built. Any bank can license the same models. Very few can rapidly rebuild the underlying compliance, permissions, and audit infrastructure needed to support machine-to-machine client interaction safely at scale.
Regulators have already started paying attention. FINRA recently flagged governance around agentic AI interactions as an emerging examination priority. That matters because many existing frameworks assume a human decision-maker somewhere in the workflow.
The industry has yet to settle some basic questions. Who is accountable when an autonomous system makes a decision? How should a bank supervise activity started by a client’s AI? How do suitability obligations and best execution responsibilities apply when the counterparty is software? And how do you audit a workflow that machines largely run?
Those questions will become more urgent as execution-capable systems become more common. The challenge for banks is that governance infrastructure takes years to build, while client expectations can shift very quickly once a new access model proves useful.
The market today still talks about AI largely as a productivity story. That’s understandable because productivity gains are visible, measurable, and relatively easy to implement. Every institution can improve internal workflows with summarization, copilots, and automation.
The redesign of the institutional access layer itself is the larger shift happening. When clients begin interacting through autonomous systems rather than exclusively through people, the economics and competitive dynamics of institutional banking start to change.
The new interaction necessitates guardrails to operate at scale:
- Governed access infrastructure
- Machine-readable client interfaces
- Permissioning systems
- Auditability
- Operational trust
Those capabilities are harder to market than chatbots and copilots. They’re also much harder to replicate. Once clients begin expecting that level of integration from every provider, catching up won’t be fast.
Banks don’t need every answer today. But they do need to start asking the right questions now. Which client agents are already calling our APIs? What are they allowed to see? Could we reconstruct what they did, and why?
The banks that can answer those questions with confidence will set the standards for this market. The rest will be living by someone else’s.
Anand Pandya
Head of Modern Data Platforms - Financial Services
IBM
