Agentic AI generates code faster, but inside a payment flow it can also act on that code, and that's where the risk changes shape. The guests on our Fintech Recoded interview series shared their perspectives on handling such risks.
Most compliance frameworks were built for AI that answers a question.
They weren't built for AI that takes action inside a regulated flow on its own.
TL;DR
Provenance and audit trails matter as much as accuracy,
Cardholder data stays out of dev environments and AI tooling,
Pre-approved workflows keep compliance from bottlenecking,
Accountability stays with a person, not the agent.
What changes once AI can act inside a payment flow?
For years, AI in payments meant a model answering a question or drafting a suggestion. A person still decided what happened next.
Autonomous AI agents collapse that gap. They can chain steps together, call other systems, and execute decisions inside a live payment flow without anyone reviewing it first.
That's a different risk profile than a chatbot getting an answer wrong. A wrong agent acts on that mistake before anyone catches it.
What are the non-negotiables when AI touches a regulated payment flow?
Thredd processes payments across more than 100 fintechs and digital banks, so as AI agents in finance move from assisting to acting, every AI-touched change carries real regulatory weight for CTO Marilyn McDonald.
Her non-negotiables start with where information comes from.
"We focus on provenance: where did this information come from, and what did the model use to generate its output." — Marilyn McDonald, CTO at Thredd
Accountability then sits with a named person, regardless of who or what drafted the change. An audit trail lets a regulator walk it through end to end.
Data boundaries are set before any of that, in line with PCI compliance requirements. No cardholder data can appear in dev or test environments, and none can appear inside the AI tooling itself.
How do compliance teams keep pace without becoming a bottleneck?
Compliance moves slowly by nature, and AI tooling doesn't wait for it. Thredd's answer is what McDonald calls paved paths.
"A paved path is a pre-approved flow within established boundaries. You approve the flow once, and every implementation within those boundaries is covered." — Marilyn McDonald, CTO at Thredd
As long as an agent stays inside those boundaries, it doesn't need a fresh sign-off every time it runs. Compliance reviews the novel cases instead of re-litigating the routine ones.
TSH built a similar principle into Cleeng's Adyen payment integration. The system uses throttling, least-privilege access, and encryption at rest, designed to meet payment security standards from the architecture up, not checked after launch.
Where does human accountability become non-negotiable?
"A human has to be accountable for any change in production for a regulated flow." — Marilyn McDonald, CTO at Thredd
An agent can generate the output and gather the evidence. A person still has to say “yes” before it reaches production.
That line holds hardest during a live incident, when real money is moving and someone has to make a judgment call on incomplete information. Regulatory attestation works the same way. An agent can produce the evidence, but someone inside the firm signs off on it.
What about the vendors and models a fintech doesn't fully control?
Every agent touching a payment flow eventually calls something the fintech didn't build. That's where the risk gets harder to see.
"If we want to use a new vendor, we do a Change Risk Assessment. Doing a CRA on a third party tells you we have requirements for each vendor we want to integrate with." — Dennis Overbeeke, CTO at New10
McDonald frames the same problem from the buyer's side. Vendor models are the harder part of the equation, because fully verifying a system someone else built and trained is difficult.
Her team's workaround is to control the boundary instead. That means deciding what data goes into a vendor tool, and what it's allowed to return.
How ready is payments infrastructure for the next AI regulation?
If a new rule targeting AI in payments landed tomorrow, the harder question isn't whether engineering can adapt. It's whether a firm has visibility into its own estate.
That means knowing where AI is running, which vendor tools are involved, and what data each one can reach.
Agent-specific rules are still rare in 2026, but existing risk principles still apply to what firms build on top of them. No rulebook isn't the same as no expectations.
McDonald puts the bar in one line. A regulator should be able to walk in and ask for a firm's AI inventory.
"If it's a nightmare to produce, you're not ready." — Marilyn McDonald, CTO at Thredd
Provenance and audit trails prove a decision even when a model made the first move,
Boundaries drawn before deployment do more work than reviews done after,
Accountability for a regulated action stays with a named person, not the agent that touched it.
Agents will keep getting better at acting on their own. In payments, that's exactly why the guardrails have to be built before the agent is, not around it afterward.
Read the original conversations with Marilyn McDonald and Dennis Overbeeke to learn more about it.
The interview with Marilyn McDonald, CTO, Thredd.
The interview with Dennis Overbeeke, CTO, New10.
Authors
Adrian Senecki
Copywriter and budding fiction writer, interested in (but not limited to) the business side of software development. Likes acquiring new skills and foretelling the future.
Andrzej Wysoczański
Frontend developer with 10 years of experience. With The Software House for almost 7 years, going from a regular dev to the Head of Frontend. He loves keeping tabs on the latest frontend technologies, especially React-related. Regular of the Taby & Spacje podcast (tsh.io/taby-vs-spacje) for Polish speaking programmers.