The story of artificial intelligence in business has followed a familiar script for the past few years. To help employees work faster, companies deployed AI to summarize documents, draft emails, write code, and answer questions. The role was clear: AI advised, humans decided, and humans acted.
That division of labor is starting to blur. A growing number of businesses, led by JPMorgan, are moving beyond AI as a productivity assistant and exploring something with considerably different implications, namely, AI systems like those made by Anthropic that do not wait for human approval but act on their own authority.
These systems execute transactions, manage workflows, purchase services, and interact with financial infrastructure directly. The shift is early, but the direction is becoming hard to ignore.
From assistant to operator: what agentic AI actually means
Most AI in use today is still advisory. It produces an output and waits. A human reads the recommendation, approves or adjusts it, then takes the action. Gartner predicts that 40% of enterprise applications will have task-specific AI agents by the end of 2026, up from under 5% in 2025.
Agentic AI inverts the traditional sequence. Instead of producing a recommendation and waiting, an AI system is given a goal and the authority to pursue it. For example, AI might call APIs, initiate payments, or deploy software, all at a speed and scale very different from humans’.
More AI:
- Goldman Sachs has blunt message for AI stock investors
- Microsoft CEO sends a blunt warning on AI and the tech ecosystem
- The next AI infrastructure race has nothing to do with chips
Michael Heinrich, CEO of 0G Labs, has been building infrastructure for exactly this kind of AI activity. He told TheStreet the sectors most likely to change first share a common characteristic: “The first domains to flip are the ones that are already digital and rule-bound: payments and treasury, software development, and digital asset operations.”
If a process already runs through software, an AI agent can be inserted into it without rebuilding everything from scratch. Payments that go through APIs, treasury functions that operate on rules, and software pipelines that follow defined steps are all structurally ready for autonomous participation in ways a physical retail operation isn’t.
Why finance may feel this shift before any other industry
Financial systems have been moving toward programmability for years. Payments can be initiated via API, assets can be represented digitally, and transactions can settle through automated contracts.
The IMF noted in April 2026 that a new generation of agentic AI systems has begun emerging across payments and financial markets, capable of initiating and managing financial operations under delegated authority. The infrastructure that once required human operators at every step has become increasingly machine-readable.
The competitive pressure to move in this direction is already visible. Anthropic unveiled 10 AI agents designed specifically for financial services in May 2026, targeting tasks from investment banking pitch decks to compliance review, Bloomberg reported. That creates an environment where AI agents can do more than observe.
Eric Swartz, general counsel of Panther Hollow Ventures, told TheStreet that the change runs deeper than most people are framing it. “The biggest shift may not be automation, but participation,” he said.
An AI that automates a task is still a tool, but an AI agent that participates in a market is something else.
Related: Sam Altman makes stunning admission about AI
A picture is already emerging regarding what participation looks like in practice.
An AI agent authorized to manage treasury operations could monitor liquidity in real time, rebalance allocations, and initiate payments when conditions are met, all without a human approving each step. The human sets the parameters, and the agent operates inside them.
Casper Network CTO Michael Steuer frames the underlying shift in terms of what AI systems are now capable of that they were not before. “The most profound transformation will happen where digital systems move from passive analysis to active execution,” he told TheStreet.
Casper recently launched its AI Toolkit, designed to support agent-driven interactions and micropayments between machines, Benzinga noted.
The trust problem that no amount of AI intelligence can solve on its own
The case for agentic AI in business is not hard to make. Many enterprise workflows are not intellectually demanding; they are operationally fragmented. Employees spend enormous amounts of time moving information between systems, waiting for approvals, and executing tasks that follow predictable rules.
AI agents could compress all of that. The question is not whether they are capable, but whether anyone is willing to let them.
Trust is where the conversation gets complicated. Financial systems were built around human identity and human accountability. When something goes wrong, there is a person responsible, and when a transaction is disputed, there is a legal framework for resolving it. Autonomous AI agents do not fit cleanly into either of those structures.
Swartz believes the problem comes down to a single word. “The key challenge is trust. Financial systems are built around human identity, liability, and accountability. If AI agents are going to transact, borrow, lend, or hold assets, markets will need new frameworks for identity, permissions, collateral, governance, and risk management.”
“The technology may arrive before the market structure,” he added, flagging the timing risk. Building those frameworks will require collaboration among technologists, financial institutions, and regulators who are still working out how to think about AI liability in contexts far simpler than autonomous financial agents.
Cravetiger/Getty Images
What AI guardrails actually look like in practice
One of the more persistent misconceptions about agentic AI is that autonomy and oversight are in tension. The people building this infrastructure generally disagree. JPMorgan told CNBC it plans to deploy autonomous AI agents in 2026, with guardrails built at the account level rather than left to the model.
Steuer is unambiguous on the principle. “The value will not come from giving an AI unconstrained custody of capital,” he said.
The more serious implementations share a common design: a bounded operating environment where an AI agent is given a defined role, permitted actions, spending limits, and an audit trail. Humans design the envelope. The interesting question is not how much freedom to give the agent but how precisely those boundaries can be defined and verified.
There is a related problem that gets less attention. The question of what an agent is allowed to do is separate from whether you can actually verify what it did.
“Privacy, transparency, verifiability, and safety are not nice-to-haves for autonomous systems, they are the precondition for trusting one with a transaction,” Heinrich explained. For AI agents to move beyond pilots and into real financial activity, the audit infrastructure has to be as good as the agents themselves.
The bigger picture: what agentic AI means for the economy
Pull back from the individual use cases, and the implications are harder to ignore. If AI agents can independently purchase services, manage resources, coordinate workflows, and execute transactions, they are not just tools augmenting human workers. They are becoming economic participants in their own right.
This is an early transition. Large parts of the necessary infrastructure, including legal frameworks, identity systems, and settlement rails designed for machine-to-machine commerce, are still being built. The distance between what AI agents can technically do and what financial and legal systems are prepared to accommodate could slow adoption even as capability advances.
McKinsey’s 2025 global survey on AI found that 62% of organizations are at least experimenting with AI agents, but only 23% are actively scaling them, and most of those only within one or two functions, McKinsey reported. Those numbers are a useful reminder that the technology is further along than the deployment.
The more immediate opportunity sits in workflows that do not require new legal frameworks. These include internal treasury operations, software pipelines, and data management tasks that already run through software and rule-based systems.
Whether AI agents eventually become genuine economic actors at scale depends on what gets built around them over the next several years, both in technology and in the institutional structures that determine who is responsible when something goes wrong.
Related: Apple’s iOS 27 surprise could change the AI narrative
