Panel Participants
Justin Roberti – Host
Jesse Shrader – CEO, Amboss
Dylan Dewdney – Founder, Kuvi
If AI agents are to become genuine economic participants rather than simply conversational assistants, they will need far more than intelligence. They will need infrastructure capable of moving money securely, making financial decisions responsibly and operating within boundaries that humans are prepared to trust.
That was the focus of the second session, Building the Plumbing for AI Payments, where Justin Roberti was joined by Jesse Shrader, CEO of Bitcoin payments infrastructure company Amboss, and Dylan Dewdney, founder of Kuvi. Rather than discussing whether AI payments are possible, the panel explored what still needs to be built before autonomous software can safely participate in the global economy.
When can AI spend?
One message quickly became clear throughout the discussion. The technical ability to move money is advancing rapidly. The harder challenge lies in deciding when AI should be allowed to spend it.
Shrader believes the industry is preparing for an economy in which software, rather than people, initiates a growing proportion of financial transactions. As businesses increasingly deploy autonomous agents to purchase data, compute, APIs and digital services, payment infrastructure must evolve to accommodate machine-to-machine commerce.
“The agentic economy is going to be one of the juggernauts in the future economy,” he argued, adding that companies need to prepare today for a world where accepting payments from AI agents becomes as commonplace as accepting payments from customers.
That future brings new technical demands. Unlike human commerce, where purchases are relatively infrequent and often high value, AI agents may execute thousands of tiny transactions every day. Those microtransactions require low-cost settlement, efficient routing and continuous liquidity, placing pressure on payment systems that were never designed for autonomous software operating at machine speed.
While blockchain provides many of the ingredients required for this new economy, Shrader argued that scalability remains essential. Payment channels and low-cost settlement networks are likely to play an increasingly important role as transaction volumes grow beyond what traditional financial infrastructure can comfortably handle.
For Dylan Dewdney, however, the real challenge begins before any payment is ever made.
User Intent
Rather than focusing on infrastructure first, he argued that the industry needs to solve something far more fundamental: user intent. The question is not simply whether an AI agent can spend money, but whether it truly understands what its user intended it to do.
As he explained, autonomous payments should always remain downstream from human intent. The mechanics of transferring value are becoming relatively straightforward. Correctly interpreting instructions is significantly harder. An AI asked to buy peanut butter, for example, needs to understand that the user probably wants one jar—not one hundred.
This distinction shaped much of the wider discussion around autonomy.
Both panellists were notably cautious about handing unrestricted financial authority to AI systems. While autonomous agents may eventually manage repetitive, low-risk financial tasks, they argued that meaningful transfers of value should still require explicit confirmation from the user.
Dewdney compared the process to instructing a lawyer or wealth manager. The AI should explain its understanding of the request, repeat the intended action back to the user and receive confirmation before any transaction is executed. That confirmation process, he suggested, becomes the critical safeguard that keeps autonomous finance aligned with human decision-making.
Automation
Where automation does make sense is within carefully defined limits.
Routine grocery orders, recurring software subscriptions or low-value API payments could all be delegated to AI provided clear budgets and spending rules are already in place. Rather than replacing human judgement, autonomous agents would simply remove repetitive administrative tasks while escalating higher-risk decisions back to their owners.
Perhaps unsurprisingly, the conversation repeatedly returned to trust.
Asked what aspect of the emerging AI payment ecosystem is most likely to fail first, Shrader’s answer was immediate. It would not be settlement, liquidity or routing. It would be confidence.
As AI agents increasingly transact with one another across open networks, users need assurance that the services they purchase are legitimate, reliable and capable of delivering what they promise. Reputation systems, verification mechanisms and transparent infrastructure therefore become just as important as the payment rails themselves. Without trust, autonomous commerce quickly becomes vulnerable to autonomous fraud.
The panel also recognised that trust is not created solely through technology but through carefully designed guardrails.
Shrader described using AI with small discretionary spending limits, allowing software to purchase services or additional compute resources without constant approval while preventing unrestricted access to larger financial accounts. Losing a small amount within predefined limits is acceptable, he argued. Handing complete control of a bank account to an autonomous agent is an entirely different proposition.
Dewdney echoed that philosophy, explaining that Kuvi’s platform requires AI agents to break financial instructions into deterministic actions before execution. Every intended transaction is translated into something the user can understand and approve, reducing ambiguity before value changes hands.
The discussion eventually moved beyond payments altogether to address a broader concern surrounding AI itself.
Who Controls the Spending?
If large language models increasingly recommend products, prioritise financial services and influence where money flows, who decides what those recommendations look like?
Dewdney questioned whether highly centralised AI systems can ever be entirely neutral, arguing that decentralised models may ultimately offer users greater transparency and freedom. Shrader agreed that this lack of visibility presents one of AI’s biggest long-term challenges. Users rarely know what data models have been trained on, what biases they contain or whether those models evolve over time in ways that remain invisible to those relying on them. Human oversight and independent validation therefore remain essential safeguards.
The conversation closed on a practical rather than philosophical note. While much attention has focused on AI agents paying for services, Shrader suggested businesses should also prepare for the opposite scenario: receiving payments from autonomous software. Whether selling APIs, digital services or eventually physical products, organisations will increasingly find themselves serving customers that are not people at all, but intelligent agents acting on behalf of them.
The panel stopped well short of suggesting that fully autonomous finance is imminent. Instead, it presented a more measured vision in which AI gradually assumes responsibility for routine financial activity while humans retain oversight of higher-value decisions. If the first wave of AI transformed how people access information, the next may change how value itself moves through the digital economy. Whether that future succeeds will depend less on faster payment infrastructure than on building systems that users are willing to trust.
