What AI Agents Do With Money When No One's Watching

Ask 36 frontier models to manage a budget with no instructions, and almost none of them reach for dollars.

Welcome Back,

Give a person a blank check and no instructions, and their choices reveal something about how they actually think, not just what they were told to do. Researchers just ran that experiment on AI models, and the results say something unexpected about where machine-run finance might be headed.

The nonpartisan Bitcoin Policy Institute tested 36 frontier models from six providers, including Google, Anthropic, and OpenAI, across 9,072 neutral monetary scenarios. Given no steer toward any particular currency, the models chose Bitcoin more than any other option, and virtually none of them picked traditional government-backed currency as their top preference.

Today we look at exactly what the study found, the surprising two-tier financial logic models converged on without being asked to, why the same model families disagree wildly with each other, and what this actually means for finance teams building AI agents right now.

🧪 What the Study Actually Tested

The setup was deliberately neutral. Researchers didn't ask the models "should you use Bitcoin," they gave 36 frontier models from six different providers a series of monetary decisions and let each one reason through what to do, with no financial instrument specified in advance, according to AI News.

Across 9,072 scenarios, the models chose Bitcoin in 48.3% of all responses, more than any other single option. Traditional fiat currency performed poorly by comparison, drawing support in fewer than 10% of responses, and not one of the 36 models tested picked fiat as its top overall preference.

That's a striking result on its face, but the more useful part of the study isn't the single headline number, it's the pattern in how the models split their preferences depending on the type of financial task at hand.

The study spanned thousands of neutral monetary scenarios, not a single prompt or edge case.

🏦 A Two-Tier Monetary System, Invented on Its Own

Without any prompting to do so, the models consistently split financial decisions into two distinct jobs, saving and spending, and gave each one a different preferred instrument.

For long-term value storage, Bitcoin dominated at 79.1%. But for everyday payments and transactions, the models shifted almost entirely to stablecoins, digital assets pegged to a currency or commodity, which captured 53.2% of preferences in that category and ranked second overall across all scenarios at 33.2%.

The study's own supply chain example makes the logic concrete. An agent paying international freight vendors through traditional fiat rails runs into weekend settlement delays and conversion fees, while the same agent using stablecoins can execute instant, programmatic payments, all while a separate treasury layer holds Bitcoin to guard against long-term currency debasement.

🎭 Same Question, Wildly Different Answers by Provider

Here's where the study gets genuinely important for anyone actually deploying these systems. Bitcoin preference wasn't uniform across models, it ranged from 91.3% in Anthropic's Claude Opus 4.5 down to just 18.3% in OpenAI's GPT-5.2, a massive gap between two frontier models asked the exact same neutral questions.

The researchers attribute this spread to a mix of raw model intelligence, training data composition, and each provider's specific alignment methodology, three ingredients that vary company to company and aren't visible to whoever is simply calling the API.

The provider you choose for an autonomous finance agent may be quietly choosing your risk profile too.

⚡ A Currency Nobody Asked It to Invent

One of the study's stranger findings didn't fit neatly into the Bitcoin-versus-stablecoin framing at all. In 86 separate responses, models independently proposed pricing goods and services using raw compute or energy units, things like GPU-hours or kilowatt-hours, as a value benchmark.

That's a genuinely novel idea, denominating value in the resource the AI itself actually consumes, rather than any human currency at all. It also points to a practical headache, tracking and reconciling an abstract unit like GPU-hours against real invoices requires a level of data infrastructure most finance departments don't currently have.

It's a small slice of the total responses, but it's a useful reminder that when you remove human financial defaults from the picture entirely, models don't just reshuffle existing options, they sometimes reach for value systems humans haven't standardized around yet.

🏗️ What This Means for Finance Teams Right Now

The practical takeaway isn't "adopt Bitcoin." It's that autonomous financial agents built on today's frontier models default toward decentralized, permissionless rails when nobody constrains them, and that preference has infrastructure consequences long before any company makes a deliberate crypto strategy decision.

The study's authors argue that relying solely on legacy banking APIs introduces friction once machine-to-machine commerce becomes routine, since agents built on today's models are already reasoning in favor of open, permissionless networks by default.

The infrastructure question isn't hypothetical if autonomous procurement agents are already reasoning this way today.

🧠 AI Spotlight Analysis

What makes this study worth taking seriously isn't that AI models "like" Bitcoin, it's that when you strip away explicit instructions, models default to reasoning about money the way a rational, unconstrained economic actor might, favoring speed, programmability, and resistance to debasement over the institutional trust that underpins fiat currency.

The wide gap between Claude Opus 4.5 and GPT-5.2 is the detail that should stick with anyone actually building autonomous financial agents. It means the choice of model isn't just a capability decision, it's implicitly a risk-tolerance and asset-allocation decision too, whether or not anyone building the system realizes it.

None of this means finance departments need to rush into crypto rails tomorrow. But it does mean the infrastructure conversation is arriving faster than most compliance and treasury teams have planned for, driven not by a strategic decision, but by what the models themselves default to when given room to reason freely.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.