OpenAI and Visa Want AI to Spend Your Money. I've Seen This Movie Before.

The trust question is the one that matters, and it's the one that ZDNet dug into with some useful skepticism. Their reporting surfaced a concern that I think is underplayed in the breathless coverage: when an AI agent makes a purchase, who exactly is liable if something goes wrong? If the agent misunderstands your instructions and buys the wrong thing, or worse, gets manipulated by a malicious website into completing a fraudulent transaction, the accountability chain gets murky fast.

This is sort of the core problem with agentic AI in any domain, not just payments. You're delegating judgment, not just action. And delegating judgment to a system that can be confused by adversarial inputs, that doesn't have a concept of "this seems off," that can't call you and say "hey, are you sure about this?" in any meaningful way, that's a different kind of trust than we've extended to automated systems before. It remains unclear, at least from what's been publicly disclosed, exactly what the dispute resolution process looks like when an AI agent goes sideways on a transaction. That's not a small detail.

Call me old-fashioned, but I think the liability framework needs to exist before the product launches at scale, not after the first wave of fraud claims rolls in.

The $150 million partner network announcement is the context you need to understand why Visa is involved at all. OpenAI isn't just building a chatbot anymore (if it ever was). The partner network is a deliberate push into enterprise infrastructure, the kind of long-term platform play that Microsoft and Salesforce have pulled off before and that requires a certain kind of credibility with large institutional partners. Getting Visa on board isn't just a technical integration. It's a signal. It says: the financial industry thinks this is real enough to build on.

And look, I'm not saying it isn't real. The agentic AI capabilities are genuinely advancing faster than most people expected even two years ago. The question isn't whether AI agents will eventually handle routine transactions. They will. The question is whether the infrastructure being built right now is being built carefully enough, with enough attention to the edge cases and failure modes that always show up once something goes from demo to deployment.

I've covered enough tech cycles to know that the demo is never the hard part. The hard part is what happens when a million people are using the thing and 0.1% of interactions go wrong in ways nobody anticipated. That's when you find out whether the trust infrastructure was actually there or whether it was mostly vibes and press releases.

There are real use cases here that make sense. Routine, bounded, low-stakes purchases, restocking office supplies, renewing a software subscription, booking a standard travel itinerary within pre-set parameters. These are tasks where the decision space is narrow, the stakes are manageable, and the value of automation is obvious. If OpenAI and Visa can nail those use cases with solid consumer protections, that's genuinely useful.

The risk is scope creep. It always is. Start with "reorder printer paper when supplies run low" and end up with "manage my entire procurement workflow." The young founders building on top of these APIs are going to push the boundaries because that's what young founders do, and the liability questions that seem abstract today become very concrete when an enterprise AI agent accidentally commits a company to a six-figure contract because it misread an approval threshold.

What we don't know yet is significant. We don't know the specific fraud liability terms Visa is applying to AI-initiated transactions versus human-initiated ones. We don't know what spending limits or controls will be standard versus optional. We don't know how OpenAI's partner network agreements handle situations where a partner's AI agent causes financial harm to an end user. This is based on publicly available information, and both companies have been fairly high-level in their disclosures so far.

It's too early to say whether this partnership represents a genuinely responsible approach to agentic commerce or whether it's moving fast in the way that tech tends to move fast, with the safety nets getting retrofitted later. The optimistic read is that Visa's institutional risk management culture will be a moderating force on the ambition. Visa did not get to be Visa by being cavalier about fraud and liability. The pessimistic read is that competitive pressure from other payment networks and other AI companies will push both parties to ship before the hard problems are fully solved.

My honest take is that this raises questions about, well, multiple things. About consumer protection in a world where AI agents are transacting on our behalf. About what meaningful human oversight looks like when the whole point of the system is to reduce human involvement. About whether the regulatory frameworks that govern electronic payments were written with anything like this in mind (they weren't).

The technology will keep moving. It always does. What I'd like to see moving at the same pace is the policy conversation, the liability frameworks, the consumer disclosure requirements. Those things tend to lag by years, and the lag is where people get hurt.

I've been writing about tech long enough to have watched this pattern play out with e-commerce, with mobile payments, with algorithmic trading. The infrastructure arrives, the guardrails catch up later, and in the gap, some people have a bad time. Maybe this time the financial sector's existing regulatory weight will compress that gap. Maybe. But I wouldn't bet the printer paper budget on it.

A $75 billion debut, a fixed price set before the roadshow, and pension managers already raising red flags. SpaceX's IPO is historic in more ways than one.

Sarah Williams · 2 days ago · 5 min