The FBI counted $893 million in AI-linked fraud losses in its first year of tracking them

The call in the FBI's file starts with a grandson's voice, shaken, saying there has been an accident and he needs money before anyone finds out. The voice is software that learned him from a clip posted online, working down a list of phone numbers ^[28]. Divide roughly $893 million by 22,000 complaints and the average reported loss is about $40,600, and the real average is lower, because the case count is "more than 22,000" ^[23].

Investment fraud carries $632 million of the total, about 71 percent, which leaves roughly $261 million spread across every other AI-linked complaint type ^[21][20]. The panic call is the vivid case, but the money is mostly in investment flows. The account breaks the $893 million down two ways, by that category and by victims over 60. It does not separate household money from business money ^[27].

Cloning a voice now takes a few seconds of audio and cheap consumer tools ^[7]. In one study, listeners identified an AI-generated voice only about 60 percent of the time, so they missed roughly four in ten ^[8][25]. The cons are built around fear and urgency, and stress pushes people toward fast judgments at the moment they need slow ones ^[15].

The other version of this never speaks to the victim. Stolen personal data sells for a few dollars on dark web markets, and criminals feed it to automated AI agents that probe bank and fintech systems around the clock ^[13]. Last fall Anthropic disrupted an espionage campaign in which an AI agent performed 80 percent to 90 percent of the intrusion work against roughly 30 targets, including financial institutions ^[12].

In 2024 a finance employee at the architecture and design firm Arup wired about $25 million to fraudsters after a video meeting staffed by deepfakes of the chief financial officer and several colleagues ^[9]. That single transfer equals about 2.8 percent of everything Americans reported to the FBI as AI-linked loss the following year ^[26]. Deloitte projects AI will help push overall US fraud losses to $40 billion by 2027, up from $12.3 billion in 2023, about 3.25 times the earlier figure ^[6][24].

The article's author, a finance professor who studies household finance and how people use AI to make money decisions ^[16], wrote that "Banks defend their own wire rooms with procedures, not vigilance." ^[17] The procedure offered to households is a callback: hang up, then dial a number you already know, such as your bank's fraud hotline ^[18].

Two axes sort the flows a product team owns. Whether one user can finish the flow alone under time pressure, and whether it can be undone within an hour. In the quadrant that is solo and irreversible, the callback becomes a build item: the app supplies the number and the transfer pauses until the callback clears. Whether it worked shows up in how often a paused transfer is never resumed.