Navy Federal Uses AI Bots to Fight Scams, Keeps $125 Million Safe

In January 2026, Navy Federal Credit Union deployed a new line of defense against scams targeting its members: a platform of bots that talk to scammers in place of the victims. The move breaks with the usual playbook of wire transfer filters and generic warnings. The result, as reported by Banking Dive on September 11, 2026^[1], is roughly $125 million kept out of wire transfer scams over the past 19 months [1].

That figure deserves a careful reading: it covers a period longer than the Cube AI partnership, and it comes from the credit union itself. It is still a before-and-after comparison with a clear denominator, and that makes it worth studying.

The original insight: fraud is a mature problem, scams are not

Carrie Foran Sepulveda is vice president of fraud and physical security at Navy Federal, the largest credit union in the United States, headquartered in Vienna, Virginia, with $204 billion in assets [1]. Her position is clear-cut: on fraud, the industry has built solid defenses. Fraud attempts against members dropped by roughly 25% from 2024 to 2025, according to what the credit union told Banking Dive [1].

The crime has moved elsewhere. Foran Sepulveda describes a "crime balloon": squeeze it on one side and the volume swells on the other, toward scams, which are harder to see [1].

That distinction is the heart of the case. In fraud, a stranger moves the customer's money, and the signals are well known: a new device, an unusual time of day, behavior that breaks the pattern. In a scam, the customer moves the money themselves, persuaded by an impostor posing as the bank or the police [1].

"For the most part it's an entirely different set of signals, data and tools," the executive said [1]. To the consumer the two experiences look alike; to those fighting them, they are two different trades.

The mechanism: a bot that pretends to fall for the scam

This is where Cube AI comes in. The platform uses conversational bots that reply to scammers the way a gullible victim would, and keep the dialogue going until the decisive moment: when the scammer names the account the money should be sent to [1].

That account is the data point that matters. Once obtained, it is reported to the financial institutions involved, which can block or monitor incoming transfers. Navy Federal can then warn a member who is about to wire money to an account already identified as a scam destination, backed by concrete proof rather than a vague suspicion [1].

The operational difference is significant. A warning grounded in hard evidence persuades the member to stop, while a vague alert is ignored by someone already under pressure.

The credit union also works with the Global Anti-Scam Alliance, other financial institutions, social media platforms and law enforcement [1]. The bot produces the information; the network makes it useful.

The independent measure: what the research says

The scambaiting mechanism built on language models has also been evaluated outside a commercial setting. Siadati, Jafarian and Jafarikhah published the study "Send to which account? Evaluation of an LLM-based Scambaiting System"^[2] on arXiv on September 10, 2025 [3].

The numbers from that test, over five months of operation [3]:

• more than 2,600 conversations with real scammers;
• over 18,700 messages exchanged;
• a sensitive information disclosure rate of about 32%;
• a Human Acceptance Rate of about 70%.

The 32% measures the share of conversations in which the scammer revealed a useful detail, such as an account or a payment channel. The 70% shows how often human reviewers approved the model-generated replies before they were sent. Together they show that the method works at scale and that a human stays in the loop [3].

To be clear: the study evaluates an academic system, and its numbers describe that system. The lesson for the Navy Federal case is that this class of tools holds up under scrutiny, and that separates the story from a mere vendor announcement.

The friction point: fewer than half reply

Every verified story contains a correction or a limit, and here the limit lies in the opening move. In the study, 48.7% of scammers replied to the bot's first message [3]. More than one in two ignored the bait.

That figure cuts the image of the infallible bot down to size. The yield depends on volume: many conversations must be started to produce a useful number of flagged accounts. That is why Navy Federal's scale, with millions of members and a network of partners, is part of both the design and the advantage.

There is also a friction in the timeline. The Cube AI partnership began in January 2026, while the $125 million figure covers 19 months [1]. Readers must keep the two horizons apart: the first is a recent investment, the second the outcome of a broader program.

How the institution talks to its members

Technology alone leaves the weak spot exposed: the person on the phone. Navy Federal has published a dedicated page, "How to Spot and Avoid AI Scams"^[3], explaining to members the scams built with AI, from cloned voices to messages impersonating a relative or the credit union itself [2].

The tone is direct. The page describes the warning signs, such as urgency, demands for secrecy and unusual payment channels, and lays out the steps to take: hang up, call back on an official number, check with the real person [2].

It is the half of the system that is often overlooked. The bot intercepts the scammer's account; the informed member recognizes the cloned voice before typing in the amount. The two defenses add up.

What to take away from this case

For a founder or an SME CEO, the lesson is in the method: before buying a tool, ask which crime you are really fighting. Navy Federal separated fraud from scams and chose different signals for each [1].

For a CTO or a head of product, the replicable pattern is the role reversal: using AI to produce a hard data point, the destination account, instead of a probabilistic risk score. The hard data point activates the partner network; the score stays locked inside the company.

For a board, the bar has moved. A not-for-profit cooperative institution reports $125 million protected in 19 months [1], and an independent study confirms that the mechanism holds up [3]. The benchmark for anyone handling retail payments has changed.

For a manager, the idea to take home is simple: the human review loop, with the 70% acceptance rate measured in the study [3], is what makes the automation sustainable.

The open question

The case leaves a question for every organization that handles money or personal data. Which of your defenses protects the customer from a stranger, and which protects them when they themselves have been talked into it?

Navy Federal answered by building a bot that listens to scammers. Your answer may be different, as long as it starts from the same distinction.