How J.P. Morgan Payments Orchestrates AI Agents To Turn Payment Signals Into Trusted Insights

Enterprise teams have spent years sitting on more payment data than they could ever read, and capable AI agents have finally made the backlog of questions answerable at a speed that was never practical before. Now it comes down to deciding which answers deserve to be acted on. A pattern that looks compelling on a Tuesday can turn out to be a sampling error, a privacy problem, or a coincidence that dissolves the moment it meets historical context. The approach taking shape inside large financial institutions treats that judgment as an engineering problem: a coordinated layer of narrowly scoped agents that check each other's work and decide, against a fixed set of rules, whether a signal has actually earned the word 'insight.'

Florencia Ardissone is Head of AI Pricing Transformation at J.P. Morgan Payments, where she leads the data products behind the firm's Customer Insights work for small-business, mid-market, and enterprise clients. Before moving into payments, she was a strategy consultant at McKinsey in New York and across Latin America, and she holds an MBA from the University of Chicago Booth School of Business. Her teams turn the bank's payments data into products that merchants use to make real decisions, which has given her a close read on where AI genuinely moves the work forward and where it mostly adds noise. For her, the discipline that matters most is the one that decides which answers are allowed through.

"You have to put a lot of rules into play before you're comfortable saying this is a real insight, and not a data anomaly, the wrong sampling, or something that just looks interesting," Ardissone said. The bar sits high because a wrong call at a bank costs client trust, not just a misleading chart. Those rules encode statistical relevance, privacy compliance, and historical context, and they run before a finding ever reaches a person.

• Trust the data before the insight: Every rule downstream depends on the inputs being clean and privacy-compliant in the first place, which is why her teams fence off what the agents are allowed to touch and work only from data that has already cleared a governance review. The sequence she describes starts well upstream of any model. "To trust your insights, you first have to trust your data. Then you have to trust the rules that let a signal become an insight, and then the interpretation that turns that insight into an action," she said. Skip a step and a weak link shows up later as a confident, wrong recommendation.

• The bar between a fact and a decision: "We play a lot of these what-if scenarios much faster, which gives us a lot more confidence. One agent runs the analysis while another checks that the attributes actually meet our criteria." That cross-checking is where a tempting correlation gets pressure-tested before anyone sees it: whether a behavior that reads as generational is really driven by age, for instance, or by some other demographic marker entirely. The distinction isn't academic when a cohort like Gen Z is visibly reshaping seasonal spending, since mistaking the driver sends a business after the wrong customer. It's the difference between a striking chart and something a business can act on, which is what separates a real decision system from another dashboard.

• Scale, not net-new: Ardissone is careful about what the current setup does and does not deliver. "Our agents are helping with the scale, not with the net new yet. I can see a world where AI itself gives us a greater wealth of insights, but in the way we operate today, what they give you is scale," Ardissone said. The agents replicate an established method of distilling signals from payments data across far more customers, merchants, and segments than a bounded team could ever reach by hand, the same lens the bank turns on public moments like a global sporting tournament, and for now that reach is the payoff.

What makes that scale safe rather than reckless is the shape of the system underneath it, part of a wider AI push across the bank. Ardissone is emphatic that the analyst her clients interact with is not one large model answering freely, but a set of agents with different jobs and different privileges. Each question moves through stages, and most of the engineering goes into the stages that have nothing to do with producing the answer. The point of that structure is to spend effort and money on control first and computation second.

• The analyst is a system, not a model: "When a question comes in," Ardissone said, "an agent evaluates the context, and if it's out of what we want it to answer, it stops there." From there a more capable model interprets the question that survives, while a separate model acts as a judge, confirming the working agent isn't deleting rows or straying from what the code is supposed to do. The firm assembles these controlled flows with standard agent-orchestration frameworks rather than letting a single model roam, closer to managing each deployment like a product than turning one loose.

• Spend tokens where they earn it: Cost discipline is wired into the same structure. The scope check runs on purpose, so the expensive models never engage a question the firm didn't want answered, and the routine math runs on something cheaper. "The calculation is an easier, less sophisticated model. But you need a more sophisticated model to interpret the question, and an LLM as a judge checking that, while it works, it isn't deleting rows and the code stays aligned with what you expect," Ardissone said. Measuring token use at each step is what keeps a single analysis from becoming an open-ended bill.

Even a well-built system runs into a limit that has nothing to do with engineering. Ardissone has found that the model's ability to answer has moved faster than any client's willingness to believe the answer. Closing that distance is its own body of work, and it is where she is most candid about what remains unsettled. The value of an insight only shows up when someone acts on it, which makes trust the hinge the whole exercise turns on.

• Capability has outrun trust: "There's a gap between how well the model can answer and how much the client trusts it. The only value is when you act, and if you get the right nudge but don't trust the source enough to act on it, there's still a gap." One approach her team is testing is surfacing a confidence score alongside each answer, so the person can see how much weight the system is putting behind it. The larger question of when an agent has earned the room to act on its own is one many enterprises are still working through, and it sits directly behind the customer-insights products clients actually log into.

• Draw the box: The lever she keeps returning to is how much freedom to grant the agents in the first place. A tightly drawn scope lowers the risk of a recommendation no one intended, at the cost of the serendipity a freer system might produce. "You can create a box that says this is the scope where you can generate recommendations, and if you need to move outside it, ask, or don't," she said. Open the box too wide and the burden lands on evaluation, where hallucinations, toxicity, or a signal that mimics something harmless, like a card-testing attack that reads like ordinary demand, can slip through.

• Where it already pays: The upside of the disciplined version is concrete, and it's measured in time and granularity. Work that used to consume a full quarter now lands in a fraction of it. "In the past, this would have taken three and a half months. We were able to compare a client across food segments, show how to influence the customer journey, and do it in a week and a half." Those are the gains the bank points to in its published work, where a footwear retailer credited the approach with measurable margin gains after realigning staffing and advertising spend to what the data showed.

For a CIO weighing the same move, Ardissone's counsel lands less on the technology than on the order of operations. Trust the data, then the rules that promote a signal to an insight, then the interpretation that turns it into an action, and accept that no single insight points to one obvious step. A shift in where customers concentrate might argue for a new store and a new marketing plan at the same time, and something still has to decide which comes first.

What she will not hand to the agents is how much latitude they get. "There's always a trade-off between the degree of freedom you give your AI and how prescriptive you are. That's something you need to set beforehand, and it should guide the way you let the infrastructure work for you," Ardissone said. At an institution that treats its reputation as one of its most valuable assets, that setting stays deliberately conservative.

The views and opinions expressed are those of Florencia Ardissone and do not represent the official policy or position of any organization.