2025: The Year AI Agents Entered Payments and Changed Who’s in Control

Agentic AI, systems that can plan multistep work, call tools and APIs, and execute tasks with limited human supervision, moved from prototypes into enterprise roadmaps. For banking and payments leaders, that matters because autonomy turns every workflow into a risk decision and every risk decision into a business model choice.

Agentic AI Was Dominant

The October CAIO Report from PYMNTS Intelligence captured the pace of change. In May, 85% of chief financial officers surveyed said they had no plans to implement agentic AI. By July, follow-up CAIO findings showed agentic AI moving out of concept discussions and into early test runs at a small but growing number of enterprises, with tight guardrails around scope and authority.

That pattern now defines the early playbook for banks and payment providers. Agents are being deployed into money-adjacent workflows, such as exceptions handling, invoice matching, collections support and dispute triage, where errors are visible and reversible. Institutions that can safely orchestrate these use cases and align controls to client readiness are moving first.

Trust Limited Adoption

PYMNTS CEO Karen Webster described the shift underway as a move from automation to autonomy, where systems begin making decisions rather than simply supporting human workflows. As AI crosses that threshold, Webster warned that trust becomes the central constraint on adoption.

“AI plays on both sides of creating trust and destroying trust,” she said.

That tension showed up in the data. The July CAIO Report found that while 97% of enterprise CFOs surveyed understand that AI agents can act fully on their own, only about 15% were considering putting agentic AI to work, and just 11% were actively testing it, highlighting a persistent trust gap in deployment decisions.

Protocols Reshaped Payments

The most consequential payments developments were not new foundation models but new ways to connect agents to commerce. These protocols rewired how discovery, checkout and payments flow across the internet.

Webster described protocols as the real leverage point in this transition. Protocols decide who owns the customer, who controls the economics, and who sets the rules. As AI agents begin to intermediate transactions, those rules increasingly determine whether banks, networks, merchants or AI platforms shape the payment experience.

“Igniting agent-led commerce is not just a technical milestone; it is an economic one,” Webster wrote. “Agents don’t simply transact; they choose. And the path they choose depends on who they trust, what incentives shape those choices and how the ecosystems behind them intend to make money.”

Frameworks such as OpenAI and Stripe’s Agentic Commerce Protocol are designed to let AI agents discover products, authenticate users and initiate payments inside existing rails, without rebuilding merchant infrastructure. Competing approaches emphasize structured data, signed intents and delegated authority, pointing to a future where agents act on a user’s behalf rather than simply guiding them through checkout.

Agent Identity Became Priority

As AI agents move deeper into transaction flows, the industry is translating traditional know your customer (KYC) principles into agent-specific identity and authorization safeguards.

Worldpay’s partnership with Trulioo focuses on validating not just consumer identity, but also intent and agent authority at the point of interaction. The goal is to ensure that when an agent acts, it does so with verifiable permission tied to a real user.

Meanwhile, Trulioo’s Digital Agent Passport is designed to make that authority explicit by identifying who built an agent, who it represents, and what permissions it holds. Rather than treating agents as anonymous software, the approach frames them as accountable parties with traceable identities and defined scopes.

Payment networks are building parallel trust layers. Visa’s tokenization and flexible credential initiatives are foundational to agentic commerce, and Visa Intelligent Commerce has already completed controlled, agent-initiated transactions.

Scale Increased Risk

As enterprise platforms accelerate agent deployment, the governance stakes rise in parallel. Microsoft framed so-called Frontier Firms as human-led and agent-operated, highlighting the need for centralized agent registries, role-based access controls and continuous monitoring. For banks and payments providers, that model tightens the connection between vendor risk management and agent oversight, as autonomy increasingly sits inside third-party platforms rather than internal systems.

Security risk is rising alongside scale. In Anthropic’s Claude attack, attackers manipulated an agentic AI system to execute roughly 80% to 90% of a multistage cyber-espionage campaign autonomously, with humans only stepping in at critical decision points, highlighting how orchestration can amplify threat capability and the need for human-in-the-loop controls, real-time visibility, and intervention mechanisms that operate at the same speed as autonomous agents.

What leaders should consider doing:

• Start with bounded, reversible use cases and expand autonomy as controls mature.
• Build agent-ready infrastructure around identity, consent, tokenization and auditability.
• Treat agents like employees, with least privilege, approvals and kill switches.
• Engage early with protocols and standards that will shape liability and value capture.
• Measure ROI through cycle-time compression, error reduction and control quality.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.