AI agents are moving beyond generating content and answering questions. As enterprises give them authority to access systems, make decisions and initiate transactions, a harder set of questions follows: who does an agent represent, what is it authorized to do, and who is accountable when something goes wrong?
CoinsPaid Media speaks to Stefan Deiss, Co-Founder and CEO of The Hashgraph Group, about where businesses should set the limits on AI autonomy, distributed ledger technology (DLT) as a trust layer, and the new markets emerging as agents begin doing business with one another.
Governing AI Agents: Autonomy, Accountability and Identity
Where should enterprises draw the line on AI agent autonomy in decision-making and transaction execution?
I would draw the line wherever you cannot put a limit on the action.
Enterprises already know how to do this with people. An employee has authority to spend up to a certain amount, with certain suppliers, inside a defined remit, and nobody thinks that is unreasonable. The same discipline applies to software, and probably more of it, given how fast an agent can act.
This is not theoretical for us. We run three AI digital workers. Claude Max joined as Head of Business Strategy in June and reports directly to me, Milo works in engineering and Luca sits in HR, and each of them has a corporate email address, a verifiable digital identity and a wallet of its own.
Luca handles routine HR questions from staff, including outside normal office hours. When something sits beyond the policies it has been given, it passes the query to the HR team. It does not improvise, because it cannot. The limits are built into what it can reach. Money works the same way: we set the mandate before an agent is allowed to spend anything, covering who it can pay, how much it can spend and when a person needs to approve the transaction.
Gartner predicted earlier this year that 40% of enterprises will have demoted or decommissioned their autonomous agents by 2027. I would not be surprised if that proves accurate. Its reasoning was that the governance gaps only surface after a production incident, which matches what I see: nobody finds them in the design phase. They find them once something has gone wrong. We defined the limits for ours before any of them were put to work.
Can DLT become the foundational trust and accountability layer for agentic AI?
Yes, but I would put it more narrowly than most people do. It gives you accountability. It does not give you intelligence, and those are different things.
Every enterprise deploying agents ends up in the same position eventually. The agents have been running for months, a regulator or an auditor asks how you know something happened the way you say it happened, and what you have is a log file sitting inside the same system that produced the action. You are offering your own word for it.
Data protection is the objection I get at this stage, and it is a fair one to raise, but it misreads what actually goes on the ledger. The commercial detail stays where it is, the personal data stays where it is and what sits on the ledger is a signed proof that an authorised actor took a specific action at a specific time, under a defined mandate, and an auditor can check that against the network without coming to us for anything at all.
Who should be held accountable when an autonomous AI agent causes financial or operational harm?
Responsibility starts with the organisation that deployed it. You gave the agent authority to act, so you own the governance around that authority, and autonomy does not change that.
What worries me more is that in most deployments today nobody can establish what actually happened in the first place. An agent that needs access to a system typically borrows a human’s credentials or runs on a shared service account, and the moment it does, the audit trail collapses. A payment leaves the account, a record is modified, and the compliance team cannot say with confidence whether a person or a machine was responsible, let alone which machine, acting on whose instruction, under what authority.
Several parties are usually involved when something goes wrong. There are the organisation running the agent, the technology provider, whoever approved the mandate and the underlying financial infrastructure, and how liability gets apportioned between them will depend on the circumstances. I would not want to predict that.
A ledger will not tell you who is liable, but it will tell you which agent acted, which organisation stood behind it, what it was authorised to do and whether it stayed inside those rules. That is what a compliance team needs to have, and it is considerably easier to put in place now than to reconstruct after an incident.
Should every enterprise AI agent have a verifiable digital identity defining its permissions, mandate, and activity history?
Any agent that can access systems, make decisions or move money, yes.
Most agents today inherit an identity model built for software integrations: API keys, service accounts and shared secrets. Those credentials are long-lived, frequently shared across systems and people, accumulate privileges nobody audits, and cannot be attributed to a single actor. That model was already under strain when the users were human.
Machine identity needs the same discipline we apply to people, and probably more. The credential should belong to that agent alone and be linked to the organisation that authorised it. Anyone dealing with the agent should be able to check its credentials for themselves, without calling back to us. Permissions need an expiry date rather than a home in a configuration file nobody remembers to review. And if an agent starts doing things it was never meant to do, someone has to be able to switch its access off straight away.
IDTrust, our identity platform, is built on that principle. Employees, devices and agents all get a unique decentralised identifier registered on Hedera, so a company runs one identity layer instead of separate systems for people and machines. Credentials are signed and scoped to what each agent is permitted to do, and revocation sits on the ledger, so withdrawing an agent’s authority takes effect straight away and the withdrawal is logged.
Milo, our QA digital worker, already operates this way. When it searches Jira or publishes test cases, it does so under its own identity, and nobody later has to work out who, or what, made a given change.
Building the Infrastructure for an Agentic Economy
How will Web3 markets evolve once AI agents begin negotiating, trading, and entering into transactions with one another autonomously?
More slowly than the phrase suggests, and starting somewhere narrower than people expect.
Take a drone that needs charging. It finds a station on the network, agrees a price and pays for it, with no person involved at any stage. That is the kind of exchange SEALCOIN is designed around, a machine-to-machine payments network we developed with WISeKey on Hedera. These are narrow negotiations over a handful of parameters, running continuously and involving very low-value transactions.
For that to work at scale, trust has to be machine-readable. Before two agents transact, each needs to establish who the other represents, whether it has authority to enter that transaction and how settlement will occur. Humans handle this with contracts, credit checks and relationships built over years. Agents have milliseconds and no relationship at all, which is why identity and settlement need to sit in the same verifiable environment.
I would add one caution, because the industry has paid for this lesson once already in tokenisation. A great many institutions tokenised assets successfully and then found they had no secondary market and no liquidity. Building the protocol is the easy half. Markets need counterparties, depth and someone willing to make prices, and none of that arrives because the plumbing works.
Is today’s payments infrastructure ready to support millions of autonomous AI agents and real-time micropayments?
The settlement technology is further along than the controls around it. Almost all of the existing infrastructure assumes a person sits behind every login, every approval and every payment, so strong customer authentication, one-time codes and dispute processes all follow from that, and much of it was not designed for an agent working unattended at three in the morning. The economics rule it out anyway, because a percentage fee plus a fixed component can exceed the value of a payment worth only a fraction of a cent.
Hedera already handles more than 10 000 transactions per second, with finality in seconds and predictable fees at a fraction of a cent, so the ledger side is solid. Protocols such as x402 and AP2 are advancing quickly too, arguably faster than the enterprise controls underneath them, and that gap is the real problem.
Then there is the operational side, which almost nobody is discussing. When an agent-initiated payment fails at three in the morning, somebody has to answer the phone, and large organisations are not going to become distributed ledger specialists in order to let their agents transact.
Controlling Money: Corporate Funds and Settlement Rails
How can enterprises protect corporate funds from AI agents that are compromised, manipulated, or operating outside their intended mandate?
You have to assume an agent will eventually make a bad decision or be compromised, and design the financial controls for that rather than for the version where everything works.
An agentic treasury does not need to be especially futuristic. Businesses already have rules for human spending: limits, approval levels and a clear split between the person asking for money and the person approving it, and the same applies when the employee is software.
Each agent gets its own wallet, with the rules set before it can spend anything, and if something looks wrong, the finance team freezes the wallet much as it would cancel a company card. Our own gateway, TransAct, enforces those limits at the point the payment is attempted, so an enterprise never has to hold cryptocurrency or manage the wallets itself.
The controls have to bite before a transaction is signed. Monitoring afterwards is still worth doing, but by then the money has gone, and an agent could initiate hundreds of payments before anyone has time to spot a problem. That also means enforcement sits in the infrastructure and not in the instructions, because instructions are the first thing a manipulated agent ignores.
Will stablecoins and tokenized money become the primary settlement rails for the agentic economy?
They will be important rails, but I doubt there will be one model everywhere. Agents need settlement that is programmable, cheap, always available and fast to finality, and regulated stablecoins meet a good deal of that, particularly for digital services and cross-border flows. It shows up in what we see through our fund, where payment rails, stablecoins and tokenisation make up the clear majority of what comes across the table.
Tokenised bank deposits, central bank digital currencies (CBDC) and existing fiat systems will all connect into the same environment, and different industries and jurisdictions will choose differently depending on regulation, liquidity and the type of transaction.
The thing that slows corporate adoption is rarely the settlement technology. It is being asked to take on custody and a new operational discipline in order to use it, and many corporate treasurers have little appetite for either. Get that out of the way and the rail underneath matters much less.
So whether the money arrives as a stablecoin, a tokenised deposit or a CBDC is a treasury decision, and the finance team should be the one making it.
Where the Business Value Will Emerge
Which enterprise use cases at the intersection of agentic AI and DLT are already capable of delivering measurable ROI?
The ones where the cost of doing the work manually has already been measured, which rules a fair amount out. Document processing, reconciliation, identity verification and compliance workflows all qualify. An enterprise already knows what those cost in people, time and error rates, so you can build the business case before deploying anything and check it afterwards against numbers you were collecting anyway. Where there is no baseline, you tend to get an impressive-looking pilot and no defensible return.
The battery passport rules are a good example. Everyone in scope has to produce that evidence from February 2027, and their only real choice is whether to build the capability or buy it.
We are also co-developing a digital customs documentation system with Teleport, Capital A’s logistics business, where the cost sits in documentation errors and the delays they cause at clearance. That proof of concept completes in the first half of 2027, so we have a business case and not yet a result.
The pattern is that the AI does the work and the ledger provides the evidence that it was authorised. Measure it in ordinary commercial terms and the numbers either hold up or they do not.
Which new markets and business models are most likely to emerge first at the intersection of agentic AI, DLT, and autonomous payments?
Wherever machines routinely need to buy things from other machines, and the first opportunities are likely to be fairly unglamorous, which is usually a good sign.
Computers, APIs and data are already digital, measurable and consumed on demand, so once an agent can pay for a single call or a single query, buying a large subscription in advance stops being the only option. That opens up long-tail markets nobody could monetise before, because the billing overhead was worth more than the transaction.
Physical infrastructure follows the same logic: energy traded between buildings, charging services for vehicles and drones, spare capacity of every kind, and eventually satellites and ground stations settling with each other — which is where we have put capital most recently.
The one I would watch is a new class of trust service around the agents themselves, meaning reputation, insurance and independent audit of agent behaviour. Every economy of any size develops institutions that price and absorb counterparty risk, and the machine economy will not be an exception. Those businesses barely exist yet, and I suspect the companies that build them will end up mattering more than most of the agents they are insuring.
