Your AI Agent Has a Settlement Problem

Banks are investing heavily in agentic AI and, separately, in tokenized infrastructure. Most are treating them as parallel tracks that will eventually converge. That sequencing assumption deserves scrutiny, because the two programs are more interdependent than most technology roadmaps currently reflect.

Here is the underlying issue. Agentic AI systems are fundamentally different from the predictive models and decision-support tools that preceded them. A model surfaces an insight. An agent acts on it. That distinction is not a marketing nuance. It has direct infrastructure implications that most deployment plans have not yet accounted for.

When an agent acts, the transaction needs to settle. Not at the end of the day. Not on the next business day. At the moment of execution, because the next instruction in the workflow depends on the outcome of the current one.

Batch settlement breaks that dependency entirely. If an agent identifies a liquidity shortfall, selects the optimal collateral to move, and initiates the transfer, but settlement infrastructure cannot confirm finality until the following morning, the agent is not managing treasury in real time. It is queuing instructions into a system that will process them on a schedule designed for a world where humans were the actors. By the time those instructions settle, the market conditions that generated them may no longer hold. The agent has not failed. The rails have.

NTT DATA has described this as the “stack gap,” the chasm between what agentic AI demands and what most bank infrastructure can actually deliver. MIT research cited across multiple industry analyses found that infrastructure integration failures, not model quality, are the primary reason AI pilots in banking do not deliver measurable value at scale. The intelligence is not the limiting factor. The foundation is.

This matters particularly for treasury and payments operations, where the value of autonomous execution is most direct. An agent managing intraday collateral across counterparties, monitoring exposures continuously, and optimizing cash positions in real time requires infrastructure that can move with it. A16z’s 2026 outlook states this directly: AI agents will require payments that move at internet speed, supported by programmable settlement tools. The shift toward intent-based autonomous systems is not compatible with rails designed around human processing windows.

What autonomous financial workflows actually require is atomic settlement: the simultaneous, irrevocable exchange of value that confirms finality in real time. This is precisely what tokenized infrastructure is being built to deliver. JPMorgan’s deposit token on Base, BNY’s tokenized deposit platform for institutional clients, and the Cari Network consortium of five regional banks all represent, at their core, the construction of settlement rails that do not depend on overnight batch cycles. They are not solely a tokenization story. They are an AI infrastructure story. The institutions building programmable settlement rails today are building the prerequisite for autonomous financial operations at scale.

The sequencing implication for banks running these as separate programs is direct. At some point in the near term, the agents being deployed in treasury and payments workflows will be capable of executing decisions faster than the underlying settlement infrastructure can confirm them. When that happens, the organization faces a choice: constrain the agents to what the rails allow, accepting that autonomous execution stops at the boundary where manual handoff begins, or rebuild the rails at considerably higher cost and complexity than if the two programs had been designed as a single program from the start.

There is also a client-facing dimension worth naming. Corporate treasury teams are building their own agentic workflows. A client that constructs an AI-native treasury function will not need its bank to manage those decisions. It will need its bank’s infrastructure to support autonomous execution without reintroducing manual intervention at the settlement boundary. Banks that cannot provide that will find corporate clients gravitating toward institutions, or platforms, that can.

The practical question for every bank currently running an agentic AI program is whether the settlement infrastructure those agents will eventually depend on is being built in parallel. Not as a future consideration. As a current design decision. The two programs are not sequential. They are the same program.