Lindsay Walker, AI Studio Product Manager at Hedera – Interview Series

Lindsay Walker is AI Studio Product Manager at Hedera, working on tools and infrastructure for AI agents to interact with distributed ledger technology. Her experience at Starling Lab included applying cryptography to establish the provenance of digital evidence for journalism and legal applications. That background informs her approach to agentic commerce, particularly how to verify human authorization, enforce spending permissions, and preserve evidence of an agent’s actions.

Hedera is a distributed ledger platform supporting digital payments, tokenization, smart contracts, and verifiable records through its consensus services. Hedera AI Studio provides tools for connecting AI applications and agents to the network. In this interview, Walker discusses deterministic payment controls, micropayments, delegated authority, dispute resolution, and the technical and commercial standards needed for autonomous commerce.

At Starling Lab, you worked on cryptographic tools to establish the provenance of digital evidence for journalism and legal applications. How has that experience shaped your approach to agentic commerce, where businesses need to verify not just what an AI agent did, but who authorized it to act?

Working at Starling Lab introduced me to the world of applied cryptography. There, I learned about all the ways we apply and use cryptographic information to secure internet traffic, transactions, and more.

At the lab, we used these methods to establish exactly what digital media (such as photos) was created, by whom, and when, which was becoming more and more important in an age where AI could generate images, audio, and video that could be mistaken for the real thing – with real economic and political consequences. Understanding the foundations of cryptography and how it can be applied in the real world, through simple but elegant schemes like PKI and TLS, has prepared me for the world of agentic commerce.

This work made me realize how many technological options we have. It is possible to have a world where AI, security, and privacy coexist. After working on projects where we used cryptography and distributed systems for solving problems, I quickly and clearly saw how this could solve problems in agentic commerce.

In agentic commerce, many different parties are going to have to exchange evidence and information. It’s important to be able to prove who provided this information and to have tamper-proof (not able to be changed after the fact), nearly instantaneous data and evidence about commercial activity available to multiple parties. This is important because AI agents can hallucinate and cannot always be trusted. The patterns we used at Starling Lab – hashing to privatize data, signing information to attest who provided what information, as well as the use of blockchain and other append-only trustless data sources (so there doesn’t have to be one central trusted authority) – make perfect sense in recording both what an agent did, but also who or what gave it authorization to transact.

What changes when the customer initiating a payment is software rather than a person? Which limitations of existing payment infrastructure create a compelling case for blockchain, and where can traditional payment rails still serve agents effectively?

I honestly think that not that much changes. Though software may be initiating transactions, for legal and regulatory reasons, a human still has to be behind that transaction. Those selling services, goods, or data (“merchants”) need to make sure they aren’t liable for ‘making someone whole’ if funds disappear or agents behave in a way the humans who dispatched them didn’t intend. Otherwise, they won’t be willing to sell to AI agents, as the financial risk is too high. There is no jurisdiction or legal standing for AI agents, nor should there be. What does change is the fact that agents will have to have evidence that everyone can trust, which cannot be bypassed, without violating their budgets and permissions.

Industries and governments involved in agentic commerce need to decide what it is an agent needs to carry with them. In other words, what kind of ‘permission slip’ do AI Agents need to have from a human? Which elements of certain purchases must the human explicitly authorize, such as the size, color, and price range of a clothing product? What should the spending limits be? Should agents be able to spend pennies or thousands of dollars autonomously? And last but not least, we need to build universal, deterministic systems that set the rules for AI agents to act within.

Micropayments are often presented as a natural fit for agentic commerce. Which purchases, such as individual API calls, data queries or computing resources, have the strongest commercial case, and what must the total transaction cost look like for those markets to work?

It’s still early, and I wish I had a crystal ball to understand what area of commerce micropayments will be first and most widely adopted for. There is a lot of potential for data to become a real commodity in the age of AI.

Right now, there is a lot of adoption and use of crypto data, however, I think this is just a temporary spike driven by enthusiasm for x402 in the crypto community. With major card networks getting involved in developing this standard, however, x402 will gain prominence.

Traditional subscription models, where an individual or organization pays for access, don’t necessarily work in a world where AI can quickly search for and find the best information. There is also no traditional ad revenue when an AI agent visits a webpage. Perhaps there will be some subscriptions for specific types of data, but writers and other types of content creators will need a new way to monetize their work. It’s likely that non-AI generated content will also become more valuable.

It all depends on who (and what use cases) start adopting AI agents first – those early adopters will naturally shape the first markets for paid data. Say, for example, AI agents start replacing typical retail shopping marketplaces. In this case, real human reviews and feedback would become the most valuable type of data to offer via micropayment. This changes, if say, the legal industry is an early adopter. In this case, court filings and legal research would naturally emerge as the first pay-per-API call or pay-per-query micropayment industry.

How do you translate a user’s instructions into enforceable spending permissions? For example, how would you give an agent enough flexibility to compare suppliers and negotiate prices while preventing it from splitting purchases or using another route to circumvent a spending limit?

Several standards and approaches are emerging to address these problems. All of these types of enforcement do share one thing: they can’t be probabilistic; they must be deterministic and secure. When I say probabilistic versus deterministic, I’m referring to the nature of the software. AI is probabilistic: given one input, the output will shift and vary. You can never be 100% sure an AI model or agent will give you a certain result. Using deterministic software as a middle layer, however, means that if you give one input such as ‘my AI can spend no more than $100’, you know your agent will always spend $100 or less.

The budget and other spending limits should live on the payment rail, and be programmed into the systems where the payment is carried out. For example, on Hedera, we have deterministic code on-chain (that anyone can audit and verify) that enforces how much an agent can spend, sets caps on per transaction and per day spending, and limits which accounts agents can send funds to. Anything else that you can do on our network is also able to be controlled.

The problem of enforcing limits on what an agent can buy is entirely another problem, and one that all sides must agree on. Payment infrastructure providers and merchants must establish governance and regulation that specifies when and how an agent can spend so that those merchants don’t take on too much risk by enabling AI to make purchases on their platforms. Once these rules are established, merchants and payment processors will be able to assess when to allow AI purchases, thereby capturing the associated revenue opportunities, and protect themselves against loss and liability. Customers using AI to transact can also act in an informed way, putting the controls they need to on their agent’s commercial actions.

An immutable record can prove that a transaction occurred without proving that the purchase was appropriate or the seller delivered what was promised. What additional evidence is needed to make an agent’s transaction auditable, and how do you preserve that evidence without exposing sensitive financial or commercial information?

There are a few key areas that need to be addressed.

The first is disclosures and agreements. E-commerce has a lot of requirements, not only when it comes to agreeing to legally binding things like Terms of Service and Privacy Policies, but also around other disclosures such as return or delivery policies. Currently, it is required to be posted on a website that a human will view before a purchase. If a human is no longer viewing these websites, as is the case with autonomous agentic commerce, access must be proven another way. I’m not sure what this looks like. Perhaps a person must agree to this ahead of time, and only allow agents to transact on websites they have pre-approved. Or perhaps funds could be reserved to secure a purchase and price, but there is a final step a user must take (before which the transaction is able to be cancelled) where these terms are disclosed and confirmed by the user. All of this must also be recorded. Anchoring it on an accessible record will prove immensely helpful, if not necessary, in an age of rogue AI and exponentially expanding security challenges.

The second area is defining which characteristics of different purchases must be specified beforehand. For example, if you are purchasing clothing, do you need to only specify size and price, or is color and material also required so that customers can’t claim a chargeback from a merchant? The options specified by the purchaser ahead of time must be a matter of record that can’t be tampered with by agents or other parties after the fact.

The third is evidence of human permission: when that permission was given, when it is valid for, and what it covers. Not only do we need to record characteristics of a purchase, but also the scope of the human or organization’s permission.

Blockchains are well suited to these requirements because they are deterministic, can be accessed and added to by multiple parties, and can’t be changed by AI or a malicious actor after the fact. Cryptographic methods can also be used to preserve evidence of specific details without revealing the private data publicly.

Where should control of an agent’s funds and signing authority sit? How do you approach custody, delegated permissions and emergency revocation so that a compromised agent cannot drain an account before its owner can intervene?

The total financial control should never, ever be given to an AI agent. Not only are they probabilistic, they are also vulnerable to security threats like prompt injection. Given the nature of AI, we cannot guarantee complete protection against these vulnerabilities. Just like we have tools to give apps and merchants access to only limited amounts of funds, we should treat AI as similarly untrusted, perhaps even more so. What’s different in this new era is that there is a tremendous upside to autonomous action, and we are able to control some of the conditions of that autonomy.

We need tools that enable users and organizations to set limits based on their level of acceptable risk, weighed against the potential opportunity. Perhaps you would want to give your AI agent access to $2000 to purchase World Cup tickets, as long as you know they can only make purchases for tickets on platforms you trust. Perhaps you want to leave purchases completely up to AI agents or speculate on different stocks, bonds, and commodities, knowing they could lose or earn a lot of money, but knowing that the loss will never exceed $500.

Emergency revocation should also always be possible, at any time.

Stablecoins can simplify cross-border settlement, but businesses still face currency conversion, liquidity requirements and compliance obligations. Which of those frictions can agentic payment infrastructure realistically remove, and which still require banks, payment providers or other intermediaries?

The shift to new technologies, whether cross-border settlement, tokenization, or agentic commerce has never been a one-sided change. Our financial systems are complex and include many organizations that are there for a reason. We need to work with, and learn from, the banks, payment providers, and other intermediaries that have been operating in this space for decades. Distributed Ledger Technology (DLT) offers new solutions in terms of speed, autonomy, and interoperability, but it isn’t a one-stop replacement. Existing systems work well in certain situations and have safeguards, economic stability considerations, and regulatory and governance frameworks built on decades, and even centuries, of experience.

Stablecoins offer more universal access and greater autonomy, but will face many of the same challenges that banking and payment infrastructure already know how to address. I don’t foresee stablecoins replacing anything wholesale, but instead offering better solutions with greater access, reach, affordability, and security than other solutions. Ultimately, we will continue to need these intermediaries, but we should instead work with them to improve the efficiency and capability of the global economic system.

Through your involvement with the x402 Foundation and the Agentic AI Foundation’s commerce work, where do you see the most consequential gaps in interoperability? Beyond agreeing on how to request and settle a payment, what must different platforms recognize about an agent’s identity, authority and transaction history?

The biggest thing I would like to see is how we can create a system where we can identify the human or organization ‘behind’ or giving permission to an agent to act on their behalf, and how the agent carries the evidence that they are given that permission. We need to agree on a standardized way for an agent to present who it is, who gave it permission, and what it has permission to do, in a format that any merchant or seller can recognize and accept. Existing solutions are very fragmented, where they exist at all. This is a technical standards problem, as well as a software problem.

In addition, in order for autonomous agentic commerce to emerge, we also need to very clearly define who is liable for what, and in what situation. There needs to be a standardized set of requirements and permissions that a human can give to an AI agent enabling merchants to accept payments from autonomous agents without taking on potentially catastrophic financial losses or liability. This is both a regulatory and a technical standards challenge.

We also need tools and systems to deterministically record and send this information. With agents in the mix, security vulnerabilities and the lack of determinism emerge as concerns, and we need a way to check that the right data is being passed along before a transaction, as well as audit when and if something does go wrong.

Payment finality benefits sellers, but buyers still need recourse when an agent orders the wrong service or a provider fails to deliver. How should refunds, disputes and responsibility for losses work in autonomous commerce, and which protections belong in smart contracts versus commercial agreements?

First of all, we need to minimize the possibility that agents can even order the wrong service. It will still happen, of course, but users should be able to set their limits according to their appetite for risk. Merchants should also have protections when they verify and comply with those requirements, and not be held liable when the person dispatching the agent is taking the risk.

Disputes will always be necessary, however, they will become more complex as actions are being taken by an agent, at machine speed, and outside of human supervision. To reduce this complexity and cost to the merchant for addressing refunds and disputes, we should automate what we can. Not everything can be checked automatically in a standardized way but things like spending budgets or transaction caps can be verified automatically.

For this, we need immutable records of the entire commercial flow and checks that all parties can trust and verify. Creating mutually verified shared data between every purchasing platform and merchant isn’t a viable option. Using DLT to maintain a public, verifiable, immutable record of evidence is a much more practical option.

Things that are deterministic and can be put on-chain should be. It costs next to nothing to have an automated check inside a smart contract, and both parties can know exactly what is being checked. But some of it also comes down to company policy and human preferences, and what merchants want to offer their customers.

From your work on Hedera AI Studio, which agentic commerce use cases are closest to sustained commercial adoption? What evidence would convince you that businesses are trusting agents with meaningful spending, and which properties of Hedera would drive their choice of infrastructure?

Honestly, none yet. Some companies and organizations are launching commercially available ‘AI Assistants’, however, many platforms are moving just as fast to block commercial actions. Some merchants have enabled AI-driven purchases, however, it is in a siloed ecosystem. You can only make purchases with certain AI agents, or within a limited network from a payment provider. It’s more of a proof of concept, and not a real, enduring, finalized solution. You can also only complete a ‘human in the loop’ style transaction where a human has eyes and approval at every step.

For truly useful agentic commerce to emerge, we need a situation where an agent can search and transact universally, and take responsibility and the need for constant attention off of the human user. Only then will adopting a new technology be appealing to a consumer. If you can only purchase certain things in certain situations, and with a high touch flow, AI Agents will simply create more noise, rather than become a technology worth adopting, especially commercially.

What would convince me that agentic commerce is working is seeing useful agents equipped with the tools they need, alongside adoption on both sides of the commercial model within a particular industry.

For example, imagine the science community wanted to find a way to fund valuable research and speed up technological development.

Agentic commerce would be able to drive value in this industry if several major players on both sides of the commercial equation agreed on the mechanisms or tools, such as publication and purchasing platforms designed for scientific research, and both sellers and buyers including R&D teams, pharmaceutical companies, and medical technology companies were equipped to conduct meaningful transactions. The final signal is when we start to see real accelerated output as a result of this system. In the medical industry, for example, that could mean new drugs, devices, and diagnostics developed at a faster rate, with transactions enabled by agentic commerce contributing to that progress.

Thank you for the great interview, readers who wish to learn more about agentic commerce and distributed ledger infrastructure should visit Hedera.