Hey, fintech fam 💜

August is moving!! We’re officially in that pre-fall sprint where somehow everything is happening at once.

My team and I are deep in FTW: San Francisco planning, having some very fun conversations with partners about what value we are bringing to the stage, event coverage and content campaigns we’ll execute, and starting to map out 2027 (yes, already 😅).

We’re opening up conversations around some bigger collaborations across the newsletter, Humans of Fintech, FTW, live events, and original content — and I’m especially excited about the brands thinking beyond one-off sponsorships and getting creative with us throughout the year.

Our fall calendar is filling up quickly, and 2027 conversations are starting now. So if Fintech Is Femme has been on your partnership wishlist, consider this your sign to reply and say hi. 👀

As for today: I’ve been doing a lot of reporting on AI lately, and one question keeps coming up that I think every financial services leader needs to be asking.

Let’s get into it. ✨

There’s a question I’ve been thinking about after several recent conversations with financial services leaders:

Do we actually know what makes an AI model good?

For the past few years, the AI race has been relatively easy to follow from the outside. Which model is smartest? Which one scores highest on benchmarks? Which can reason better? Which is fastest? And increasingly, which brings the greatest ROI?

Those questions matter.

But as financial services moves from experimenting with AI to actually putting AI agents to work, I’m increasingly convinced they’re not the most important questions anymore.

Because financial services doesn't need the world's most impressive LLM.

It needs AI that can survive contact with the real financial system.

And that requires a very different definition of "good."

This distinction came up in a recent conversation I had with Kasey Roh, U.S. CEO of Upstage, an enterprise AI company founded in South Korea.

Upstage is launching Solar Pro 4, its new flagship commercial large language model, and Roh made me think differently about how financial institutions should evaluate the models sitting underneath their AI applications.

An AI agent isn't simply answering a question like you or I might ask ChatGPT.

An agent could be asked to complete a series of steps, retrieve information from different systems, call external tools or APIs, follow an institution's policies, return information in a specific format, and determine when something needs to be escalated to a human.

Then it has to do that again.

And again.

Potentially thousands or millions of times.

That's where Roh argues behavioral reliability becomes just as important as intelligence.

In plain English: Does the model consistently do what you told it to do?

Think about a model that gets a task right on its second or third attempt.

For you and me using an AI chatbot, that's mildly annoying. We prompt it again.

Inside an enterprise agent, every retry can mean more tokens consumed, another model call, additional latency, and potentially another opportunity for something to go wrong.

Suddenly, the model with the lowest advertised price isn't necessarily the least expensive model to operate.

And the model capable of producing the most impressive answer isn't necessarily the one you want executing a financial workflow.

That's the distinction Upstage is betting on with Solar Pro 4. Rather than optimizing only around raw intelligence or completing a task at all costs, the company says it has focused on instruction following, maintaining tool-call structure, and producing outputs that stay within the schema an enterprise has established.

That might sound like technical plumbing.

In financial services, it's the product.

If an AI system is summarizing an internal document, a malformed response may be inconvenient.

If it's helping investigate a suspicious transaction, interacting with a customer, supporting a lending workflow, or eventually initiating a financial action, behaving predictably becomes much more consequential.

The question for executives becomes less about which model wins the latest benchmark and more about what happens once that model is actually inside your business.

Of course, none of this changes the pressure financial institutions are under to move quickly.

Banking offers perhaps the clearest example.

McKinsey recently argued that agentic AI could fundamentally change the economics of banking because agents won't behave like traditional consumers.

Consider deposits.

Net interest income accounts for roughly 60% of retail-bank revenue globally, according to McKinsey. Historically, many customers have prioritized the convenience of keeping their financial lives in one place rather than constantly moving deposits to capture the highest possible yield.

An AI agent doesn't have that inertia.

It can continuously monitor balances, compare rates across institutions, move idle cash into higher-yielding accounts, and move it back before bills are due.

Now apply that same logic to credit cards, loyalty programs, lending, wealth management, insurance, and payments.

Agents can compare.

Agents can switch.

Agents can negotiate.

Eventually, agents can transact.

And consumers are adopting the underlying technology remarkably quickly. McKinsey reports that generative AI reached 45% of the U.S. working-age population in roughly two years, compared with approximately 15 years for digital banking. By 2025, usage had reached 55%.

That's an extraordinary compression of the adoption curve.

It means financial institutions are simultaneously facing pressure to deploy AI internally while preparing for customers who increasingly have AI working on the other side of the transaction.

The question isn't whether financial services should participate.

It's how institutions move quickly without moving recklessly.

This is where another recent interview has stayed with me.

Kim Olson, Chief Risk Officer at Green Dot, looks at AI from the other side of the equation: risk.

One of the most important realities of deploying AI in financial services is that automation doesn't magically remove the institution's underlying obligations.

If anything, scale can make mistakes more consequential.

A bad decision made manually might affect one customer. A flawed automated process can replicate that decision across thousands of customers before someone realizes what's happening.

That makes governance, monitoring, escalation, explainability, and human oversight part of the AI product itself.

And it reinforces why the conversation Kasey and I had about behavioral reliability matters.

We're moving toward a financial system in which AI isn't simply helping humans find information. It's increasingly helping determine what happens next.

Financial institutions therefore need to understand not only what their models can do, but how those models behave when instructions conflict, data changes, a workflow breaks, or an agent encounters something its designers didn't anticipate.

That's a risk conversation.

It's also a product conversation.

And increasingly, it's a C-suite conversation.

There's one more layer to this that technology alone can't solve.

Jyoti Menon, VP of Product at Bread Financial, said something to me on the Humans of Fintech podcast that I haven't stopped thinking about:

"I hope we don't lose the ability to think critically."

Menon's concern isn't that financial services shouldn't adopt AI. Quite the opposite.

She's concerned about what happens when people begin accepting AI outputs without interrogating whether those answers actually make sense for their company, customer, regulatory environment, or problem.

And there's an interesting connection between her warning and what I'm hearing from the technical side.

We need models that follow instructions reliably.

But we also need humans capable of determining whether those were the right instructions in the first place.

Menon described prompting to me almost like writing a paper. What context did you provide? What story are you telling? What information does the model need to understand the problem?

That's critical thinking.

And as AI gets better at executing work, I think that human capability becomes more valuable, not less.

Because the financial institutions that win the AI era won't simply be the ones with access to the best models.

Everyone will have models.

The advantage will come from understanding where to deploy them, what to trust them with, how to govern them, and when a human still needs to make the call.

Maybe that's the larger shift happening right now.

The first phase of generative AI was about capability.

Look what this thing can do.

The next phase is about production.

Can we actually trust it to do the work?

For financial services leaders, that means the AI conversation has to become much more sophisticated.

Don't just ask how intelligent a model is.

Ask how reliably it follows instructions.

Ask what happens when it fails.

Ask how often it retries.

Ask where your data goes.

Ask how the system is monitored.

Ask when a human steps in.

Ask what the economics look like at scale.

And perhaps most importantly, ask whether your organization understands the technology well enough to know which questions it should be asking in the first place.

Because financial services isn't experimenting with AI in a vacuum.

We're putting it inside an industry that moves money, extends credit, protects assets, manages risk, and serves real people under real regulatory scrutiny.

The opportunity is enormous. So is the responsibility to get it right.

I'm going deeper into what financial services leaders need to understand about the model layer, behavioral reliability, risk, and the economics of agentic AI in my next Forbes column. Stay tuned. 👀

This conversation is much bigger than one article.

It’s also exactly why we’re building FTW: San Francisco around the thesis that fintech owns the agentic AI narrative.

Financial services isn't sitting on the sidelines of the AI revolution. We’re building the infrastructure, risk frameworks, payment rails, identity systems, and financial products that will determine what happens when AI agents start moving from answering questions to taking action.

And I’m inviting the people actually building that future in the room.

FTW: San Francisco is September 29–October 1. If you're a fintech or financial services leader thinking seriously about where AI is taking our industry, come join us.

📰 Share this newsletter with a friend and start growing your network.

🔗 Connect with me on LinkedIn for daily insights on leadership.

🤝 Grow your business through content & community by partnering with me.

📣 Promote yourself to 50,000 subscribers by sponsoring this newsletter or the Humans of Fintech podcast.

🎤 Host an epic event by booking me as a speaker, moderator, or emcee.

📚 Increase your expertise by ordering your copy of my book, Fintech Feminists: Increasing Inclusion, Redefining Innovation, and Changing the Future for Women Around the World.

Thanks for spending time with me today!

Love,

Nicole 💜