The financial services industry spent decades mastering “Know Your Customer” rules designed to verify that human beings are who they claim to be in financial transactions. Now, as AI agents begin shopping, paying and acting autonomously online, a new challenge has emerged: how to know whether the bot knocking at the digital door is trustworthy.
That question is giving rise to what may become one of the next battlegrounds in enterprise AI: “Know Your Agent,” or KYA. The concept, backed by firms including Experian, Visa and Cloudflare, aims to create a trust framework for autonomous AI systems that increasingly behave like consumers, employees and software intermediaries.
Experian, best known for credit scores, is attempting to position itself as a kind of passport office for AI agents. The company recently launched “Experian Agent Trust,” a framework that cryptographically links AI agents to verified human identities in an effort to reduce fraud and establish accountability in what executives believe will become a vast new layer of digital commerce.
“Trust is so important. It’s not just a nice to have, it’s a regulated necessity,” Kathleen Peters, Experian’s chief innovation officer, said in a fireside chat at the recently held AI Hot 100 Summit in New York (The AI Innovator was a media partner). “We have to build that trust. I think that’s going to come from knowing the human behind the agent and knowing their intent.”
Silicon Valley is rapidly pushing AI agents beyond chatbots and into economic activity. OpenAI, Google, Anthropic and others are building systems capable of searching websites, comparing products and completing transactions on behalf of users. Visa and Mastercard are preparing payment rails for machine-driven or “agentic” commerce. Merchants, meanwhile, are cautiously eyeing what could become a lucrative but risky new source of online traffic.
Good bots, bad bots
For years, internet infrastructure firms treated bots as pests to be blocked. Cloudflare and similar firms built businesses around filtering automated traffic from websites. But generative AI has scrambled that logic. Some bots are now expected to arrive bearing legitimate purchase intent.
“In the old days, any bot coming to your website was bad,” Peters said in a separate off-stage interview. “Now with AI, some bots are good.”
That creates an awkward problem for online merchants. A retailer may want the sale generated by an AI shopping agent but may not trust the source of the traffic. Consumers are similarly hesitant. Peters said most people she speaks with remain uncomfortable allowing AI systems to autonomously spend money on their behalf.
They tell her, “no way I would trust it. I’m not going to give some open agent my credit card information or access to my financials until that trust can be established,” she said. “That trust is going to come from experience.”
Experian’s answer is to adapt its longstanding expertise in fraud detection and identity verification to the AI era. Under the company’s framework, a consumer interacting with an AI assistant could authorize an agent to purchase, say, a pair of Bose headphones. Experian would verify the human user, evaluate fraud risk and issue a cryptographic trust token associated with the transaction.
The framework works alongside existing payment systems rather than replacing them. Visa still validates the payment credentials. Cloudflare helps determine whether the AI traffic should be allowed through to the merchant site. Experian’s role is to establish that a verified human sits behind the agent.
Beyond e-commerce
The broader ambition extends beyond shopping. Peters argues that enterprises deploying internal AI agents face many of the same trust and accountability problems. Generative AI systems are probabilistic rather than deterministic; identical prompts may produce different outcomes. That makes autonomous systems inherently harder to govern than traditional software automation.
“The thing about agents and generative AI is that they’re not necessarily deterministic,” Peters said. “Just because you did something at the input one way one time does not mean you’re always going to get the same answer.”
That uncertainty complicates questions of liability. If an AI shopping agent purchases the wrong product, who bears responsibility: the merchant, the model provider, the payment network or the user?
“I don’t know that the industry or the marketplace has fully vetted that out yet,” Peters said.
Trust is so important. It’s not just a nice to have, it’s a regulated necessity.
For now, companies appear eager to preserve the familiar architecture of e-commerce rather than reinvent it. Peters repeatedly stressed that the quickest route to consumer adoption may be making AI commerce resemble conventional online shopping as closely as possible, including the same liability structures, payment protections and fraud safeguards.
“No one wants to go relitigate all this right now,” she said. “So at least in the near term, … as we all figure this out, the more that this can follow the rules and give people comfort in the way we know how things work today, the more likely it is to get adopted, because that’s what we trust together.”
The race to define standards for agentic commerce may also reshape competition in enterprise technology. Peters believes companies will increasingly compete not merely on AI models, but on the sophistication and trustworthiness of their “agent ecosystems” – collections of autonomous systems capable of coordinating work, transactions and decisions.
“I think this will be a way that companies will distinguish themselves,” she said. “It will be how they differentiate.”
Be First to Comment