What Happens When Al Agents Can Spend Your Money?

AI agents that act for users could change who controls online shopping, advertising and routine financial decisions. The speaker argues that agents may make price comparison and benefit optimization common even among people who would not spend time doing those tasks themselves. That could weaken business models built on users overlooking coupons, premium benefits or cheaper alternatives, and force businesses to reconsider offers that stop being profitable when customers claim them consistently.

Town’s initial product used large language models to turn small-business financial records and documents such as leases into tax returns. The business generated just under $1 million in its first year, and its founders projected $2.5 million to $3 million in its second, but concluded that annual purchasing cycles, customer acquisition costs and incumbents’ distribution limited the opportunity for rapid growth. Town then shifted toward AI assistance with email, calendars and meetings, where an agent can prepare users, follow up and handle routine replies.

The product is designed to make useful work visible while leaving users in control of whether to accept the result. The speaker called it a “token allocation machine”: it identifies tasks where AI might help, and users decide whether to proceed. Access to company data and APIs makes the assistant more capable, so Town starts with email and calendars and seeks permissions for tools such as CRM as users see value. Falling token costs make previously impractical tasks, such as researching meeting attendees and preparing notes for every meeting, affordable; the speaker estimated this could cost about a penny.

The same logic could apply directly to spending. A restaurant owner might receive wholesaler prices in spreadsheets or PDFs, and an agent could compare them, flag a supplier switch or suggest removing a menu item. The speaker estimated that checking prices could cost about $1 a month and, if it identifies a problem, save hundreds or thousands of dollars. Another example was paying $5 annually for an agent to find $500 in savings. These figures illustrate potential value, not guaranteed results.

The speaker sees agents as a challenge to the ad-supported internet, where free services such as Yahoo’s stock quotes were supported by users’ attention to ads. Businesses could treat agent access like a new form of SEO and welcome agent traffic rather than block it, although reports of agent traffic exceeding human traffic largely refer to agents in general, not specifically assistant agents. The speaker favors personal agents paid for by users, arguing that recommendations could be distorted if an agent provider also sells advertising. Ads can still help people discover products, however: someone searching Amazon for Lego might welcome ads for unfamiliar construction toys before asking an agent to evaluate them.

Amazon could lose a significant part of its advertising business if it resists agents, the speaker argued, and simply aggregating supply may become less compelling. One possible response would be a better personal shopping agent, perhaps bundled with Prime, but the speaker was unsure Amazon could build one. Walmart might benefit from Amazon’s resistance because it does not have the same advertising business and may have different incentives. The speaker suggested Amazon could preserve discovery by offering a separate, worthwhile ad-supported experience for users who want it, rather than blocking agents across its services.

Other models could connect agents to commerce without conventional ads. The speaker suggested Muse might remain free and ad-free if its data improved ad targeting on Instagram, with a wallet and product search potentially linking the services. CamelCamelCamel, described as having 20 million American users after 18 years, alerts shoppers when tracked Amazon products reach target prices, but users must still complete the purchase themselves. Agents could automate more of that work. For enterprise customers, adoption also depends on controls: Town is building audit logs, protections for privileged information and visibility into spending and token use, while prioritizing self-serve small and mid-market users and developing enterprise security and compliance. The founder said large-company adoption also requires a consumer-quality experience.