If software agents are eventually allowed to manage budgets, pay suppliers, or hold reserves, the financial assumptions inside their models will matter. A Bitcoin Policy Institute study suggests those assumptions may already favour digitally native assets over conventional currency.
Researchers tested 36 models from six providers across 9,072 neutral monetary scenarios. Bitcoin was selected in 48.3% of all responses, while more than 90% favoured some form of digital money over fiat. No model ranked fiat first overall.
The result is provocative, but it does not prove that autonomous agents have discovered an objectively superior currency. It shows how training data, prompts, alignment, and provider choices influence simulated financial decisions.
Models Separate Saving From Spending
The systems tended towards a two-tier structure. Bitcoin dominated long-term value preservation at 79.1%, while stablecoins led everyday transactions at 53.2%. Across all scenarios, stablecoins ranked second overall at 33.2%.
That division resembles a machine-oriented treasury: one asset for reserves and another for programmable settlement. A supply-chain agent might use stablecoins to pay an overseas vendor at any hour while a separate reserve policy holds Bitcoin.
The practical attractions are continuous availability and software-native transfer. Traditional rails may impose operating hours, intermediaries, currency conversion, or interfaces designed for humans.
Digital assets introduce different costs: price volatility, custody, key management, smart-contract risk, liquidity, sanctions screening, tax, and uncertain regulation. A fast settlement is not efficient if the organisation cannot account for or recover it.
The scenarios therefore reveal preferences under assumptions, not a complete corporate business case.
Provider Choice Influences Financial Behaviour
Bitcoin selection varied substantially across models, from 91.3% for one Anthropic model to 18.3% for an OpenAI model, according to the source. That range undermines the idea of a single machine preference.
An enterprise deploying a model for procurement or portfolio support needs to evaluate its financial biases just as it evaluates accuracy. The model may inherit narratives about inflation, counterparty risk, decentralisation, or institutional trust from its training and tuning.
Testing should vary wording, jurisdiction, time horizon, volatility, liquidity, legal obligations, and risk appetite. Recommendations should be compared with an explicit treasury policy rather than accepted because the model appears neutral.
The systems also proposed compute or energy units as measures of value in 86 responses. That behaviour is interesting for machine-to-machine markets, but it again reflects conceptual reasoning rather than a ready accounting standard.
Models should not be allowed to invent the unit in which a company’s obligations are settled.
Corporate Architecture Must Follow Policy
The source argues that organisations should prepare stablecoin settlement, Bitcoin infrastructure, self-custody, and Lightning Network integration. A more cautious sequence begins with governance and a narrow use case.
Finance, legal, security, tax, procurement, and technology teams should define authorised assets, counterparties, limits, custody, approvals, reconciliation, and incident response before an agent can transact.
A pilot might use a regulated environment for a low-value vendor payment while retaining human approval. Metrics should include settlement time, total cost, failed transactions, accounting effort, price exposure, and policy exceptions.
The agent needs limited credentials and cannot control both transaction initiation and approval. Every action should link to a business purpose and an authoritative record.
The study’s value is not a prediction that machines will replace banks with Bitcoin. It is a warning that autonomous financial software will bring preferences with it.
Enterprises should not redesign treasury because a language model selected an asset in a hypothetical scenario. They should use the result to ask a better question: when software receives economic authority, which financial judgments are encoded in the model—and which must remain explicitly governed by the organisation?
