For decades, technology made paying faster while leaving the decisive act in human hands. A customer clicked “buy” or approved a transfer. Artificial intelligence could now remove that final gesture. Software agents can interpret objectives and initiate transactions under delegated authority—a profound change for banks, merchants and regulators.
Agentic payments introduce a new participant into financial infrastructure: software able to act rather than merely advise. It might find and order a product, launch a compliance process or allocate liquidity across markets. The recommendation engine becomes an economic actor.
That prospect shaped a BFRR conversation between hosts Jonas Gross and Michael Blaschke and Sonja Davidovic of the IMF. Davidovic co-authored the IMF note “How Agentic AI Will Reshape Payments” with Hervé Tourpe. Its central problem is architectural as much as technological.
When Probabilistic Intelligence Meets Financial Finality
Modern AI is probabilistic: it interprets context and adapts. Payment systems follow a different logic. An authorized instruction must produce a predictable result, responsibility must be identifiable and settlement legally final. A model that invents an answer may be tolerable in an email. It is unacceptable when moving money.
Connecting a general-purpose AI model to a bank account is no solution. Autonomy needs boundaries: enough latitude to fulfil a task without unrestricted access to funds, credentials and personal data.
Davidovic proposes three layers separating reasoning from control and settlement. In the first, an agent interprets intent and compares alternatives. A request for a shirt may specify brand, size, price and delivery date. The agent evaluates the market against those preferences.
The Architecture Agentic Payments Need
The second layer turns intention into a rule-bound instruction. It verifies identity and authority, checks the mandate and enforces limits on price, merchant, timing or liquidity. Routine purchases might proceed automatically while unusual transactions require human approval.
Only then does the instruction reach settlement. Existing rails execute it with the determinism, auditability and finality expected today. AI remains adaptive upstream while money stays controlled downstream.
The separation helps localise failure: a misunderstood request concerns intent, an action outside the mandate concerns authorization, and a failed approved transaction concerns the payment rail. That distinction will matter when institutions design controls and courts assign liability.
Commerce First, Treasury Later
E-commerce is the clearest early market. An agent can manage search, checkout and dispute resolution. Yet processing a return may require access to email, order histories and payment credentials. The more useful the agent, the larger the attack surface.
Institutions may find equally valuable applications away from checkout. Agents could initiate predefined compliance procedures. In treasury, autonomous allocation across currencies and venues could improve efficiency in fragmented markets. If many agents respond to the same signal, however, they could amplify volatility and systemic risk.
Opacity compounds the problem. Explaining a shirt purchase may be manageable; reconstructing a cross-border transaction involving several agents is harder. Audit trails must capture intent, authority, data and machine actions in a form supervisors and courts can understand.
No Single Form of Money Will Win Every Use Case
The debate quickly reaches the settlement asset. Stablecoins appear naturally suited to machines: they operate continuously, interact with programmable wallets and move across borders without traditional operating hours. Michael Blaschke argued that agents need money that behaves like software.
Davidovic offered a more cautious view. Credit cards already support e-commerce, while bank deposits, tokenized deposits, CBDCs or stablecoins may suit other contexts. Retail purchases and high-value cross-border settlements impose different requirements. The instrument is likely to follow the use case.
Smart contracts may offer a bridge by mapping agent mandates to programmable controls attached to digital assets. Whether this improves security and interoperability remains open.
Adoption Will Move at the Speed of Trust
Forecasts for 2030 or 2035 conceal the harder work. AI capability is unlikely to be the main constraint. Authorization, cybersecurity, privacy, explainability, accountability and legal liability will determine whether users and institutions delegate control.
The systems must also evolve. Security that is robust today may be vulnerable a year from now. Modular design, upgradeable controls and separated functions offer more resilience than a monolithic architecture.
Agentic payments may make buying and liquidity management almost invisible. Their success depends on making authority visible: who granted it, what its limits were and who answers when the agent gets a decision wrong. Technology can make a payment autonomous. Credible governance must make it acceptable.
Paper: How Agentic AI Will Reshape Payments
Previous BFRR Episode with Sonja Davidovic
Bitcoin, Fiat & Rock’n’Roll Website
Bitcoin, Fiat & Rock’n’Roll Telegram Channel
Relai*: Buy Bitcoin with Relai—you can do a one-time purchase or savings plan: Click here. Use the referral code „ROCK“ to reduce transaction fees by 0.1% while supporting Bitcoin, Fiat & Rock’n’Roll.
Value4Value Podcast Streaming: Support our podcast by listening to our episodes on the Fountain Podcast App. This way, if you wish, you can support us „Value4Value“ while listening to the podcast. You can find us on the Fountain Podcast App here: Click here
Disclaimer: The content of this podcast reflects the private opinions of the hosts, serves exclusively for general information purposes and does not constitute investment advice. Always remember: Do your own research—inform yourself before making any investment decisions, such as buying Bitcoin. First try to understand what Bitcoin is and how to store it. This podcast does not provide financial advice. Note that the co-hosts might be invested in crypto assets. Read more on our website: Click here
All links marked with „*“ are affiliate links. If you use these links for purchase, the podcast receives a small share of the revenue without any additional costs to you. On the contrary, affiliate links often include discount promotions, so you can even save money. We would appreciate it if you use these links to support us. Thank you very much!