Giving an artificial intelligence agent a wallet is easy to understand. Giving it permission to use that wallet safely is where the payments industry gets complicated.
“Large language models are probabilistic, and financial execution can’t be,” Chris Trainor, head of platform strategy, innovation and AI at Paymentus, told PYMNTS for the September edition of the “What’s Next in Payments Series: The Fall Draft.”
While the first wave of generative AI rewarded systems that could understand and produce language, the next wave is being asked to execute. These execution requests are moving right to the center of commerce, with AI agents now technically capable of changing an autopay setting, resolving billing disputes, establishing a payment arrangement and moving money without requiring a human to navigate every step.
Identity Is Becoming AI’s Control Plane
Financial execution introduces something an AI model cannot solve simply by becoming smarter. The system needs to know who is asking, what that person controls, what they are authorized to do, what the service provider permits and which deterministic infrastructure should ultimately execute the transaction.
Trainor said his first requirement is what he called “trusted customer identity,” a persistent identity layer connecting an authenticated customer with accounts, policies, payment credentials, permissions and workflows.
“Authentication remains foundational in an AI landscape,” Trainor said.
Without it, AI can converse. With it, AI can potentially act. Take, for example, the case of a customer asking why an insurance bill increased. An authenticated system with account-level context can potentially identify what changed, determine which remedies are available and guide the customer toward an authorized action.
In that world, identity becomes less like the doorway into a financial application and more like the operating context governing everything an AI agent can do once inside.
The AI Bottleneck Today Is Integration, Not Intelligence
There is another reason better models alone will not solve the problem. An agent may perfectly understand a customer asking to change a payment date. But if customer information lives in one system, billing rules in another, payments in a third and communications preferences somewhere else, intelligence stops at the boundaries of enterprise architecture.
“There’s tremendous attention right now on what AI can say, but there’s a lot less attention to what AI can actually reach,” Trainor said.
AI-powered integration is emerging as an overlooked part of the equation, Trainor said. Future-ready firms are now using AI to interpret schemas, map information and connect workflows across disparate systems without forcing companies to replace every underlying system of record.
“Many organizations have accumulated separate technologies for billing, payments, wallets, communication, fraud, disbursements and customer service,” Trainor said, adding that while these point solutions once represented specialization, in an agentic environment, they can become an integration tax.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Every additional system expands the number of connections, permissions and business rules an AI layer must understand before it can reliably complete an action. The paradox is that companies may discover their AI strategy depends as much on simplifying yesterday’s technology stack as buying tomorrow’s models.
That helps explain the skepticism Trainor said he has toward unconstrained agentic commerce.
“I would characterize unconstrained agentic commerce as probably having too much buzz,” he said.
Agentic Commerce Is Already Hitting a Determinism Wall
Autonomous purchasing makes intuitive sense when the stakes are low. Let an agent reorder household goods or find a cheaper product, for example. Service commerce introduces a different risk profile. Changing autopay, establishing a payment plan or paying an insurance premium requires more than a payment credential.
The agent needs authenticated identity, account context, eligibility rules, provider policies and evidence that the customer authorized the action. That means the future architecture may deliberately separate decision intelligence from transaction execution.
“The mistake that many have is expecting that model itself to become the system of record or a transaction processing engine,” Trainor said. “The future is not unconstrained artificial intelligence. It is governed agency.”
The wallet, in other words, is not the interesting part. The permission structure around it is.
“That’s far more impactful than adding a chatbot to a payment page,” Trainor said. “It’s a new operating model for how consumers and businesses manage ongoing financial service relationships.”
Watch the full PYMNTS TV interview with Paymentus’ Chris Trainor to hear more about:
- Why the smartest AI agent is useless if it doesn’t know who it’s dealing with. Trusted customer identity could become the foundation of agentic payments, connecting customers to their accounts, payment credentials, permissions and business rules, so AI can move from generic answers to authorized action, Trainor said.
- Why AI’s biggest payments bottleneck may be what it can reach, not what it can say. Fragmented billing, payments, fraud and customer service systems could become an “integration tax” on agentic AI, making connectivity across the existing enterprise stack valuable, Trainor said.
- Why giving AI a wallet may be solving the wrong problem. “The future is not unconstrained artificial intelligence,” Trainor said. “It is governed agency,” with AI interpreting intent while permission layers and deterministic systems control consequential financial actions.
For all PYMNTS AI coverage, subscribe to the daily AI newsletter.
Add as Preferred Source
