Franklin Templeton Digital Assets estimates that “agentic commerce” — where autonomous AI software completes economic transactions — could grow to between $3 trillion and $5 trillion by 2030. Aptos co-founder and CEO Avery Ching said the forecast depends less on AI model capability and more on whether the underlying infrastructure can reliably execute payments, enforce boundaries and provide an auditable record of what agents do.
Ching argued that for agents to move beyond recommendations into actual settlement at machine speed, the payment and identity layer must be fast and low-cost, while also offering programmable controls such as spending limits and approved counterparties. He also said accountability becomes a core requirement as agents gain autonomy, particularly for enterprises that need proof of authorization and outcomes.
Key takeaways
- Price move: No market price movement was reported in the source material.
- Catalyst: The discussion centers on how agentic AI could enable autonomous payments and the infrastructure needed to support it.
- Key implication: Enterprises may rely on blockchains to provide auditable, policy-driven payment capabilities for software-to-software transactions.
- Risk focus: Adoption hinges on confidence in confidentiality, reliability, traceability and safeguards for value-holding and spending.
What has to change for agentic commerce to scale
Ching said the central challenge is whether AI agents can be trusted with real authority by people and businesses. In his view, models are already strong enough for many day-to-day workflows, but the remaining gaps are largely infrastructural — including secure handling of data, reliable long-running execution, and the ability to prove what an agent did.
For commerce in particular, he described requirements that go beyond generic automation. Agents need their own identity and payment capabilities, paired with enforceable boundaries. The goal is to allow users or enterprises to grant capital access with strict spending limits, specify which counterparties the agent can transact with, and ensure there is a provable record of actions taken and the policy under which they were authorized.
He also emphasized an economic constraint for high-frequency activity: if agents generate large numbers of very small payments, transaction systems must be fast and cheap enough that the payment mechanism itself does not become the bottleneck.
Why throughput is only one part of the infrastructure equation
When asked whether transaction throughput is the deciding factor for AI commerce, Ching said throughput is only one element of overall system performance. He pointed to the need for properties that include security, speed, scalability and low cost across the entire decentralized financial system.
Ching compared the infrastructure required for autonomous agents to mature cloud and database platforms, highlighting expectations such as durable state, recovery, failover and predictable performance. In his view, Aptos was designed for a high-throughput, low-latency environment, aligning with the requirements of machine-driven activity.
The first scalable use case: agents paying directly
Ching said the earliest large-scale path for autonomous payments is when agents use the same interfaces humans use today to pay for goods and services — but with transactions completed by software rather than requiring a separate human or system step.
He described today’s baseline as largely “read-only” in economic terms: agents can search, recommend and identify actions, but another party typically finalizes the transaction. In his scenario, enabling direct per-request payment would change the structure of the network. Agents would be able to discover the service they need, make the payment and continue the task without setting up accounts, negotiating contract details for every spend, or waiting for approval on each transaction.
As adoption grows, he said machine-to-machine commerce would become more realistic when agents negotiate and transact with each other at machine speeds.
Accountability, programmable money and audit trails
On who should be accountable if an agent makes a bad purchase or is compromised, Ching said the system should be capable of showing exactly what happened and under what authority. He argued that enterprises would need traceability from the agent’s actions back to the policy that authorized them, supported by an immutable ledger.
Ching said accountability is not one-size-fits-all, but the architecture should make it possible to attribute actions. He described a control model where users or enterprises define allowed behavior, developers enforce those controls, and the infrastructure provides a record that cannot be quietly altered later.
For payment safety, he said the agent should not have unrestricted access to a wallet. Instead, the agent should operate with its own keys and limitations such as spending caps, approved counterparties, and categories of activity. He added that auditability could become as important as the payment itself, because if organizations cannot prove what an agent did and why it was allowed to do it, they are unlikely to grant meaningful autonomy.
Why blockchain settlement may matter for open, per-request payments
Ching said blockchain is not required for every AI interaction, but he framed the blockchain question around scenarios where software needs to transact with software that does not already have an established commercial relationship. Existing payment systems work best when there is an account holder identity, a payment processor and an established counterparty.
Agents, by contrast, may need to pay fractional amounts for services, receive results and move on repeatedly across many counterparties. Setting up accounts, API keys, invoices and bilateral relationships for each interaction, he said, does not scale well.
In that environment, Ching argued that open payment rails become useful. He pointed to the role of stablecoins as a digital unit of account and referenced protocols such as x402 for per-request payments. He said blockchain can provide a common settlement and accountability layer for unrelated machines without requiring a single company to run the entire network.
Bigger picture: managing timelines and stablecoin reliance
Ching cautioned against assuming agentic AI adoption will follow a predetermined timeline. He said the difference from past blockchain narratives is that demand for agents exists independently of crypto: enterprises are already deploying AI systems and model quality is improving rapidly. Still, he said adoption will accelerate once systems solve confidentiality, scaling, audit trails and ease of use — with the harder problem being turning agents from experiments into tools ordinary people and enterprises can trust and depend on.
On resilience, he argued the system should not rely on a single stablecoin or assume one asset fits everywhere. Instead, he said agents should follow a clear policy governing what assets they are allowed to hold and spend, and under what conditions activity should stop if an asset or counterparty falls outside the policy. He described agent wallets conceptually equipped with approved assets, spending limits, approved counterparties and automatic safeguards designed to prevent actions when constraints are violated.
While stablecoins can enable software access to digital money that can move globally and programmatically, Ching said they are only one piece of the architecture. The more important principle, he argued, is that agents need a secure way to hold and exchange value without receiving unlimited financial authority. He added that if these safeguards are put in place, payment rails underneath can evolve over time.
For investors watching the agent economy, the next phase will likely come down to whether infrastructure can deliver the combination Ching outlined: policy-based authorization, auditability, reliable execution and safe value handling. Key developments to monitor include further progress in production-grade agent deployments, regulatory clarity affecting tokenized payment rails, and major releases or upgrades from networks positioning themselves to support high-frequency, programmable settlement.
