Rachael Hadaway, VP AI Product Management, FICO

Today we're delighted to speak with Rachael Hadaway, who leads AI Product Management at FICO. With nearly 70 years of predictive analytics under its belt, FICO is now bringing generative and agentic AI into regulated financial services - without sacrificing the explainability and trust that's always defined the company. Rachael shares her vision for how algorithmic, generative, and agentic AI stack together, why banks can't afford to stay in the slow lane, and what needs to be true before agent-to-agent banking becomes safe and scalable. FICO last appeared here in April, when Scott Zoldi shared his advice to banks on AI.

My questions are in bold - over to you Rachael:

Tell us about FICO and your role.

At the core, FICO helps banks and other regulated companies make better decisions, whether that's who to lend to, how to catch fraud, or how to collect on a late payment, and we've been doing that with AI and analytics for decades, long before "AI" was a buzzword. That track record is backed by a portfolio of 150+ patented AI technologies. My job, leading AI Product Management, is basically figuring out how we bring, generative and agentic AI, into that world without losing the trust and rigour that's always been our reputation.

How does FICO think about AI, and how do algorithmic, generative, and agentic AI fit together?

We've been doing AI for a long time, mostly the algorithmic kind, and we've been doing predictive analytics models for even longer - nearly 70 years now, and mostly in retail banking. That history matters because it means we're wired to always ask, "Can you actually explain this decision?" Generative AI is the newer layer on top, helping with things like letting an analyst just ask a question in plain English instead of writing a query. Agentic AI is the next step up from that, where instead of one model spitting out a score, you've got something that can actually go do a multi-step task. We don't see these as competing they're layers stacked on each other, all running through the same governed core: an auditable, enterprise-scale execution layer, more than 100 production-ready AI models, and an agentic AI suite which puts that intelligence into the actual flow of work. The trusted decision-making core is still doing the heavy lifting underneath all of it.

Where would you prioritise agentic AI in retail banking?

The obvious answer is the high-volume, grind work onboarding, collections outreach, document chasing and yes, AI is already proving itself there; our own FICO Assistant is doing exactly that today, helping bankers investigate fraud cases and draft decision strategies faster. But I'd push back on the idea that agentic AI should stay confined to that lane. We're already seeing it move into territory that used to require actual human judgement, such as flagging which fraud cases deserve a closer look, or deciding how to structure a workout plan for someone in distress. I don't think the answer is to keep AI away from anything that looks like judgement, as that's basically where the biggest value is. The real question is whether you've built the explainability and oversight to back it up.

How does a bank compete with tech companies, given younger folks trust generative AI platforms too?

Younger customers have already shown they don't have much loyalty to a logo. They'll use whichever AI solves their problem faster, whether it's their bank, a fintech app, or ChatGPT with a plugin. So, banks can't coast on being the trusted incumbent that advantage is shrinking fast, and I don't think it's coming back. Banks do have one edge that's real: they see a customer's full financial picture, held inside a governed, auditable decisioning environment, in a way a generic AI platform doesn't. But that only matters if they build something that uses it well. If they don't, they're going to lose that relationship to whoever builds the better agent, full stop.

How can AI help banks balance risk, experience, and growth?

These used to feel like a seesaw, tighten risk and you slow everyone down, loosen it and you're taking on more danger. Good AI-driven decisioning lets you get smarter instead of just picking a side, like approving someone on different terms instead of a flat yes or no. It also means decisions happen in real time instead of someone waiting three days to hear back, which is its own kind of win for experience. Really, it's about giving the risk, product, and marketing folks the same picture so they're not all pulling in different directions.

How should banks prep for customers showing up with their own AI agents?

This one's coming faster than people think. Banks need to build for a world where it's not always a person tapping through an app sometimes it's their AI agent negotiating or handling something on their behalf. That means solid APIs, a real way to verify an agent is who it claims to be, and a plan for spotting a fake agent trying to impersonate a real customer. That's a lot of what we're building into our platform's native agent layer right now identity, access, and policy enforcement so you've got role-based controls on what an agent can touch, plus observability and audit trails across every agent run so you're not flying blind. Whoever gets that plumbing right early and reduces the friction for their customer to operate this way, is going to have a real edge later.

How can lenders make sure their AI meets rules that aren't even written yet?

The safe move is to build to the toughest standard you already know applies, like fair lending and existing model risk rules, and just assume that bar carries over to the new stuff too. Keep good records of what the AI did and why, keep a human able to step in and override things, and test constantly for bias or weird outcomes. Also, don't just wait around for regulators get in front of them and show your work, that tends to go a lot better than playing catch-up later.

What does agent-to-agent interaction look like in banking, and what must be true for it to be safe?

Picture a customer's personal finance agent negotiating a payment plan directly with the bank's collections agent, no humans involved unless something goes sideways. For that to be safe, everyone needs a reliable way to know who they're talking to, there need to be hard limits on what an agent's allowed to agree to on its own, and there needs to be a clean, auditable record of the whole exchange logs, traces, the works so it can be checked later. Honestly, the industry probably needs something like a shared rulebook for this, like how payment networks agreed on common standards years ago.

Who governs the interactions between the bank's agents, the customer's agents, and third-party agents?

Nobody, yet it's still being figured out. The banks on the hook for its own agents no matter who they're talking to, contracts will need to spell out who's liable when things cross between companies, and eventually there'll probably be some industry-wide standard. In the meantime, every bank needs its own internal watchdog real quality, evaluation, and monitoring built in, with thresholds and rollout gates, so something is tracking what its agents are doing in real time and can step in if things go off script.

As a product leader, how do you decide what's ready to build now versus what needs more guardrails?

I basically ask myself a few things. Can this be undone easily if it goes wrong? Is the decision itself clear-cut, or does it need real judgement? Can I explain after the fact what the agent did and why? And how much regulatory heat is on this if it goes sideways? If something scores well on all of those, I'll push it forward that's basically what a good rollout gate looks like. If it's shaky on any of them, especially explainability, it stays in a smaller pilot with a human closely involved until we've earned the right to loosen the leash.

Why do so many AI pilots stall out, and what does it take to scale one?

Usually, it's because the pilot was built to prove the model works, but nobody thought about how it plugs into the messy old systems and workflows already in place. A lot of the time there's also no clear owner it's someone's pet project without the budget or authority to scale it. And people underestimate how long the compliance and documentation side takes. A model can work great in testing and still need months of validation before it's allowed near real customers. Getting to production means baking all that in from day one instead of bolting it on at the end.

How should a bank's AI strategy shift from single-purpose tools to full agentic workflows?

Terrific question. A fraud model just needs to be accurate at one job. An agentic workflow is judged on the whole chain of decisions it makes, so you must think less like "tune this one model" and more like "manage a whole system" versioning it, deploying it, promoting it through environments the way you would any production software. That means way more attention on monitoring the entire process, not just the result, and building in spots where the agent knows to stop and hand off to a person. It also means risk and compliance need to be at the table from the start, not brought in at the end to sign off.

What role does explainability play once AI agents are involved in decisions?

It matters more than ever, because you lose that natural "let me walk you through my thinking" a human underwriter could give you. Every decision an agent makes needs a trail you could hand to a regulator or a customer and have it make sense. That's part of why, when we build focused language models for financial services, we don't stop at the model. We layer in Trust Scores and knowledge anchoring on top, so outputs get checked against verified, high-confidence sources and hallucination risk get pushed down before anything reaches a decision.

What will "AI-native" banking look like in a couple years, and what must change internally?

Most routine stuff, servicing, small underwriting tweaks, initial fraud digging gets handled by agents, with people focused on the harder judgement calls and actual relationships. Decisions happen in real time instead of batch cycles, and your banking "assistant" can do things, not just generic chat. Some of that's already on our roadmap at the platform level letting both people and automations rapidly assemble, test, and tune the business performance of features, rules, and models, and eventually using agentic and ML techniques to find the best operating point across a whole line of business, not just one decision at a time. To get there, banks need clean data that agents can use, a compliance function set up to keep watching things continuously instead of just approving once, and teams that genuinely work together instead of tossing projects over the wall to each other.

What's a lesson from outside banking that's shaping how you think about agentic AI now?

Back in retail, I spent a lot of time thinking about the actual moment someone's standing at the shelf, deciding what to buy. Not the org chart behind it, not which department owned pricing versus merchandising versus loyalty, just that one moment where a real person is trying to decide and everything either helps them or gets in their way. That's exactly the lesson I keep coming back to with agentic AI. The whole point of an agent is that it can move across a problem end to end, pull a transaction record, check a policy, take an action, follow up, without stopping at every department boundary the way a person or a legacy system would. But most banks are still organised around underwriting, servicing, collections, and fraud as separate kingdoms with separate data, separate approvals, and separate owners, and that structure actively fights against what makes agentic AI valuable in the first place. So, when I think about agentic AI now, I'm less worried about the model and more worried about whether we're willing to build around the customer's actual problem instead of our internal boxes, because the silos, not the technology, are usually what slows this down.

Many thanks to Rachael for taking the time to share her insights with FinTech Profile. You can learn more about FICO on their website.