Dr. Tonya M. Evans: Picture this. It's Thursday afternoon. You're between meetings, or between clients, or between school pickup and your next deadline. Your phone buzzes. It's a text from the attorney who handled your last big contract. Someone you've worked with for years. "Need to move fast on something confidential. Can you talk?" Thirty seconds later, your phone rings. It's her. Her voice. Her rhythm. The way she always says your name.
She tells you there's a time-sensitive payment to close out a deal. The paperwork is coming by email, but keep it quiet until it's done. The invoice lands, and it looks exactly right — like every other invoice you've received before. And here's the part that's new. You don't even have to pay it yourself.
A few months ago you set up an AI assistant to handle your vendor payments — anything under ten thousand dollars, it pays automatically. That's the whole point. It saves you hours every week. The invoice is for $9,800. Your assistant pays it in seconds. Except your attorney never actually texted you. She never called. That voice was built by AI — from a few minutes of her speaking on a webinar. A fake person. A real payment. And no human checkpoint in between.
Now, if that sounds like science fiction, here's the honest answer: that exact combination — a cloned voice fooling an AI agent into paying — hasn't hit the headlines yet. But both halves already have.
The cloned voice: just last month, CNN reported on a Michigan couple who lost $66,000 in a suspected AI voice-cloning scam days before closing on their home. In 2024, an employee at the engineering firm Arup's Hong Kong office sent about $25 million after a video call where the executives on screen were all deepfakes. And earlier this year, a cloned voice helped trick a private bank in Italy into wiring about 95 million euros. More than half was recovered — but the rest is still missing.
The AI agent: just this past May, someone posted a hidden, coded message on X that tricked an AI chatbot into relaying a transfer order to a crypto trading bot. The bot carried it out, and somewhere between $150,000 and $200,000 worth of tokens went out the door. Much of it was later returned — no cloned voice required, just words the machine took as instructions.
Put those two halves together and you have the risk at the center of today's episode. Today we find out who's on the hook when the machine holds the wallet. If you're ready to grow, let's go.
Hey everybody, welcome to Confidently Crypto and More, the show where we make sense of the technologies rewriting money, business, and law — crypto, AI, intellectual property, and Web3 — and what they mean for your bottom line and your future. I'm your host, Dr. Tonya M. Evans. If you are new here, hit subscribe right now so you never miss an episode. If you're watching on YouTube, subscribe, tap the bell, and drop a comment because I read them. And today I want to know: Would you let an AI agent spend money on your behalf? And if so, what's the first limit you'd set? If this show makes you smarter, share it with one — or ten — people who need to hear it, and leave us that five-star rating wherever you listen.
And everything we cover today — the transcript, the glossary, the state of crypto summary, and all the extras — live in one place: confidentlycrypto.com.
Before we get going, some news I'm so excited to share. I will be speaking at Freedom Rising, a conference hosted by Middle Church, this October 23rd through the 25th. My panel explores how the financial future is female. Head to FreedomRisingConference.com to register. More on that at the end of the show.
Now, last week in part one, we asked what AI really is. We learned the name is 70 years old. We broke AI into five types. And we drew the most important line in the whole conversation: the line between AI that talks and AI that takes action. This week we cross that line — because two things are true right now that were not true a few years ago. Number one, the machine can pay. And number two, the machine can pretend to be a person. Put those two together and you get the question every business owner, every compliance officer, and every person with a saved card in an app needs to answer: When something goes wrong, who's on the hook? You can't manage a risk you can't name. So let's get started.
Story one: Crypto Mom has left the building, but the rulemaking hasn't. Last week I told you SEC Commissioner Hester Peirce — the industry's so-called Crypto Mom — was leaving the commission. On October 2nd, she did just that. The SEC is now down to two sitting commissioners: Chairman Paul Atkins and Commissioner Mark Uyeda.
And the day before she left, the SEC proposed yet another crypto rule. This one is about custody — how investment advisors and funds are allowed to hold crypto for their clients. Among other things, it would let state trust companies serve as crypto custodians and, in limited cases, let advisors hold client crypto themselves when no eligible custodian is available. That's on top of the SEC's proposed framework for crypto offerings, Regulation Crypto Assets, which is open for public comment until October 20th.
As I wrote about in Forbes last week, with the CLARITY Act stalled in the Senate, the agencies are still writing the rules — because that really is the status quo. A patchwork of agency actions, just under a new administration. And now a two-member commission is holding the pen. The liability here: assuming the rules you're planning around today will look the same a year from now. Agency rules can shift when the people running the agencies shift. The opportunity: these are proposals, which means the comment windows are open. So if you have a stake in how these rules land, this is the time to be heard.
Next up: the rails for machine money are going up fast. Two announcements in the last two weeks show how quickly the payment infrastructure for AI agents is being built. On September 30th, Cloudflare — the internet infrastructure company — announced tools that will let businesses charge AI agents in USDC, which is Circle's dollar stablecoin. The idea: instead of a human signing up for a subscription or typing in a card number, an AI agent pays a small fee every time it uses an online service. It's in a closed beta for now, not yet broadly available, but this is worth paying attention to. And on September 28th, a company called Verona announced a new dollar stablecoin it says is built specifically for AI agent payments, with more than $100 million in commitments. One important caveat: commitments are not the same as coins actually in circulation. That number is about promises, not usage — at least not yet.
The liability here: these rails are being built for machines that act on your behalf, and every one of them raises the same question. What can your agent spend, and who approved it? The opportunity: understanding these rails now, while they're still in development, puts you ahead of the people who'll be scrambling to understand them once they are everywhere.
Now let's start with the first new reality. Last week I told you about the fifth type of AI — agentic, AI that doesn't just answer you, it takes action on your behalf. And here's how an AI agent pays for something, in plain language. You give the agent a job — like "find me the best data on this market," "book my travel," "keep our software subscriptions running." You give it permission to spend, and that might be a saved card, a digital wallet, or a stablecoin account. And then it goes to work. It finds what it needs and it pays. Sometimes a few cents at a time, sometimes many times a minute, without stopping to ask you.
That's why stablecoins keep coming up: when a machine needs to make lots of tiny payments, at any hour, to services anywhere in the world, a digital dollar that settles in seconds is a natural fit. That's the core of the BlackRock research we covered in the last episode. And that's why the card networks are building their own versions, too.
But now, the reality check. This is early. When the blockchain analytics firm TRM Labs looked at agent payment activity this year, it first filtered out everything that didn't look like real commerce — and what was left, it estimated, only somewhere between 0.6% and 7.5% of the value likely came from actual AI agents. A lot of what gets called "AI payments" is really just ordinary, automated scripts. So the volume is small. But remember the line from the BlackRock paper I told you to hold on to. It noted that anti-money-laundering checks, know-your-customer checks, and know-your-agent checks generally happen off-chain.
Know your agent. We have decades of rules about knowing your customer — KYC. A bank has to know who you are before you open an account. But who verifies that an AI agent is who it claims to be, that it's acting for the person it says it's acting for, that it's staying inside the limits it was given? Right now, there's no single settled answer to that question.
Now the second new reality. In part one, I told you that generative AI creates new things — text, images, audio, video — and I told you to remember that deepfakes live there. Here's why that matters for your money. For years, proving who you are online has rested on two things: a real identity document, like a driver's license, and a live check that the person holding it is really you — a selfie, a video, a voice. Both of those are now under attack at the same time.
On the document side: in early September, the security journalist Brian Krebs reported that a dark web marketplace was advertising more than 153 million U.S. and Canadian driver's license scans. He traced it to IDScan, a Louisiana company whose technology businesses use to verify IDs. IDScan has confirmed that an unauthorized party may have accessed customer data, including names and driver's license numbers. The company hasn't said how many people were affected. The FBI has been looking into it.
On the live check side, AI can now generate a face that blinks and moves, and clone a voice from a short recording. And U.S. regulators have been warning about this since at least 2024. That November, FinCEN — the Treasury Department's financial crimes unit — issued an alert to banks. It said financial institutions were reporting more and more suspected use of deepfakes, including AI-generated documents, photos, and videos used to get around identity checks.
Now put those two together: a real stolen license paired with an AI-generated face and voice to match. That's not a fake ID. That's a real identity wearing a synthetic face.
Remember what Bill Gates said on Meet the Press — the story we covered last week? He said the most dangerous thing right now is people with bad intent using AI, including to defraud others. Whatever you think about the bigger AI regulation debate, on that point, the evidence is already in.
Here's where the two realities collide. Go back to the cold open. The voice was fake. The agent was real. And the payment was real. In the Michigan case, people wired the money themselves — so humans were the checkpoint, and the humans were fooled. In the world we're building, the checkpoint may not be a human at all. It may be an agent following the rules you set months ago. And an agent doesn't get a funny feeling. It doesn't notice that the attorney sounded a little rushed. If the request fits the rules, it pays. If, then.
So who's on the hook? These are issue-spotting questions, not legal conclusions, and the answers depend on the facts, the contracts, the jurisdiction. But here's the map of who could be in the frame. Number one: you — the person or business that set up the agent and gave it permission to spend. Number two: the company that built the agent. Number three: the platform or payment providers that processed the transaction. Four: the bank or vendor whose identity check was fooled, if one was involved.
And here's the uncomfortable truth for a lot of us: many of us have never read the terms of service for the AI tools we're already using. Those terms may say a great deal about who carries the risk when something goes wrong. Find out what yours say before you find out the hard way.
Now let me take the strongest version of the skeptic's argument in this cross-examination. Maybe agent payments are tiny, so why should we care? The argument goes like this: Agent payments are a rounding error. TRM's own numbers show most so-called agent activity is just ordinary automation. BlackRock itself calls this early. So why spend a single minute worrying about it?
And here's the thing. On the facts, the skeptic is partly right. The volume is small. Much of the hype is ahead of current reality. But the argument falls apart on three points. Point one: the cheapest time to build guardrails is before the volume shows up, not after. Nobody installs smoke detectors after the fire. Point two: the deepfake threat isn't small, and it's not waiting for agents. That Michigan couple didn't need an AI agent to lose $66,000. It took one convincing voice at the most stressful moment of a home purchase. And point three: the identity layer that agents will eventually depend on is under attack right now. A breach involving the data of a major ID verification company isn't a future risk — it's literally a current one.
So the verdict: the agent payment volume is hype-sized, but the identity risk is not. Treat them differently.
And now for your guardrails. Last week I gave you three questions that start your AI policy. This week, let's add the guardrails specifically for your money. None of these require you to buy a single piece of software.
Number one, set hard limits for any AI that can spend. A dollar cap per transaction, a cap per day, a short list of who it's allowed to pay. Anything outside those limits stops and waits for you.
Number two, verify out of band. If a request to move money arrives by text, email, or phone call, confirm it through a different channel, using contact information you already had. Hang up and call back the number you know — not the number in the message.
Number three, agree on a code word. With your business partner, your bookkeeper, your family — something an AI couldn't learn from a webinar or a social media post.
Number four, treat urgency and secrecy as huge red flags, not reasons. "Move fast and keep quiet" shows up in these scams again and again. And never move money based only on an email, a text, or a voicemail.
And number five, write down who's accountable. If you have a team, who approves exceptions? If you're a team of one, what will you never let an agent do without you? Put it on one page, and that's your AI money policy.
Again, these are starting points for your own issue-spotting, not legal advice. A lawyer can help you pressure-test them against your contracts and your specific situation. The liability: an agent with your saved card and no limits, in a world where voices can be faked. The opportunity: the people who can show clients, partners, lenders, and regulators exactly how their AI is controlled are the ones who are going to earn trust faster than everyone who can't. That's a competitive advantage — and it costs almost nothing to build.
Now, everything we covered today comes back to one skill: spotting the issue before it becomes the problem. And that's what I do every week in the Clarity Briefing Series — my live working session on digital asset regulation. Every Thursday at noon Eastern, one hour. What changed, who it touches, and the questions worth asking before they're asked of you. You bring your questions, I take them live. And this Thursday, October 8th, we are covering the CLARITY Act and where market structure actually stands. And on October 15th, we're digging into that brand-new SEC custody proposal — because the rails these AI agents run on, stablecoins and digital assets, are governed by rules that are still being written. Treasury expects its stablecoin licensing requirement to take effect January 18, 2027. The SEC's crypto framework is still a proposal. And market structure decides which rules apply to your product, your partners, and your disclosures. You register once and you're covered for the whole fall series. So head to web3readyclarity.com/webinar. That's web3 — the number three — readyclarity.com/webinar.
So here's what I want you to walk away with from this two-part series. AI isn't new, and it isn't one thing. The line that matters most is the line between AI that talks and AI that takes action on your behalf. The machine can pay already, and the machine can pretend to be a person as well. And when those two meet, the question isn't whether the technology works — it's who's accountable when it's used against you. The volume of agent payments is still small. The identity risk is not. And the guardrails that protect you — spending limits, out-of-band verification, a code word, a one-page policy — cost almost nothing to put into place today.
Next week, on October 13th, I'm bringing in a guest to look over the horizon with me: Anne T. Griffin. Anne Griffin is an AI product leader, educator, and fellow Forbes contributor who teaches AI through O'Reilly and Pearson. We're talking about what's coming next in AI — and the liabilities and opportunities for business leaders, whether you are racing to adopt AI or trying to avoid it altogether. You won't want to miss that. So October 13th, be sure to tune in. In the meantime, join me this Thursday for the Clarity Briefing on where the CLARITY Act actually stands — register at web3readyclarity.com/webinar.
Before I let you go: subscribe if you haven't, so you don't miss the conversation with Anne. Share this episode with the person or people in your life who have an AI assistant connected to their bank account and have never thought about any of this. And if you're on YouTube, leave a comment and tell me: what's the first limit you'd put on an AI agent with your money? Be sure to leave us that five-star rating wherever you listen. Get the transcript, the glossary, and all the extras at confidentlycrypto.com.
And one more thing. I'm so excited to share that I will be speaking at Freedom Rising, a conference hosted by my church, Middle Church, this October 23rd through the 25th. My panel will be exploring how the financial future is female — and that's just one of the amazing conversations that will be happening at the intersection of justice, social change, and faith. If you've been longing for a community where justice, joy, and faith meet, I'd love to see you there. So head to FreedomRisingConference.com to register.
I'm Dr. Tonya M. Evans, and until next time: stay informed, stay prepared, and stay confidently crypto and more.