What Is Agent Identity (KYA) and Why Do AI Agents Need It?

What Is Agent Identity (KYA)?

Agent identity — what Sapiom CEO Ilan Zerbib calls a version of Know Your Agent, or KYA — is the framework that lets a merchant or platform trust an autonomous AI agent enough to let it transact. Zerbib breaks it into three parts: the human or business behind the agent, its delegated authority to act, and its own earned reputation.

“As a business or as a consumer, I give authority to that agents to act on my behalf.”

— Ilan Zerbib, Founder & CEO of Sapiom

The first two parts are close to solved, according to Zerbib. Carrying a consumer's or business's identity into an agent's transactions, and proving that identity delegated authority to the agent, just needs a shared standard. The third part — reputation — is the hard one, because the data to support it doesn't exist yet.

The Three Parts of Agent Identity: Entity, Authority, Reputation

Sapiom's answer to the reputation problem is to build a credit-score equivalent for agents. Because Sapiom's runtime sits underneath every agent it hosts, it can monitor and authorize each API call and transaction an agent makes, and use that activity to score how trustworthy the agent has been.

“The agent reputation is much harder because you need new data that don't exist, actually almost like credit score data, and that's what we have at Sapiom because we enable agents to access the economy.”

— Ilan Zerbib, Founder & CEO of Sapiom

Zerbib is candid that this layer is still early. He estimates agent identity is roughly 12 months from becoming something merchants treat as critical infrastructure, and points to Okta's investment in Sapiom through Okta Ventures as evidence that identity companies already see the machine economy as their next market.

Why Sub-Cent Payments Force Agents Onto Stablecoins

Agent identity solves who is transacting. It doesn't solve how the transaction settles — and Mesh CEO Bam Azizi argues there's only one answer once the transaction gets small enough. Machine-to-machine payments, like an agent paying a fraction of a cent for a single API call or search request, fall below what wires, Visa, and Mastercard can process.

“If an agent wants to pay 0.0005 cents, you cannot wire, you cannot use Visa card or Mastercard. You have to use a stablecoin.”

— Bam Azizi, CEO of Mesh

Azizi's point is that this isn't a matter of preference. Anti-crypto sentiment in parts of the AI world, he argues, is really aimed at crypto as a speculative asset, not at stablecoins or blockchains as settlement rails — and once volume and micropayment size are the constraints, stablecoins are the only rail left standing.

The Hidden Cost of Running Autonomous Agents

Identity and settlement aside, Zerbib says the biggest barrier to getting agents into production today is cost. When a human prompts a model, a person decides each action. When an agent runs autonomously, it consumes intelligence continuously, and that consumption doesn't scale the way subscription pricing assumes.

“The cost of inference grew from zero to 1.2 million after three months of operation. They were getting users, but they were not at million user scale.”

— Ilan Zerbib, Founder & CEO of Sapiom

Zerbib cites one Sapiom customer, a mass-market consumer product automating small-business tasks, whose inference bill outran its user growth. Defaulting every task to frontier models, he says, is what turns that math negative — Sapiom's own router instead shifts roughly 95% of tasks to open-weight models running on Sapiom's own GPUs, cutting inference cost by 60 to 90% at comparable output quality.

What AI Agents Are Actually Doing on Sapiom Today

Zerbib says Sapiom is already routing 270 million transactions and 100,000 daily agent runs, spanning coding agents, go-to-market and lead-generation agents, customer support agents, and fully automated ad-buying agents.

“We even have agents that run end-to-end ads for customers, Instagram and Facebook, and all those ads are fully automated by some of the agents running our platform.”

— Ilan Zerbib, Founder & CEO of Sapiom

The ad-buying case is the clearest example of an agent acting as an economic actor rather than a tool: a customer sets a budget, and the agent builds creative, buys placements on Instagram and Facebook, and closes the loop on performance without a person touching the campaign.

Why Consumers Still Won't Let Agents Shop for Them

Zerbib draws a sharp line between machine-to-machine commerce, where agents buy compute, data, and API access from other agents, and consumer commerce, where an agent buys something on a person's behalf. The technology for the second case largely exists — Shopify has partnered with OpenAI to let agents transact, and Google has its own protocol for agentic purchases — but adoption hasn't followed.

“The holdup is actually the consumers, people. They don't really want to delegate purchase decision for most of the things.”

— Ilan Zerbib, Founder & CEO of Sapiom

The holdup, in his telling, is that people still enjoy choosing. Consumers pulled back once ChatGPT was able to complete purchases on their behalf, and Zerbib thinks that's because decisions like which hotel or which color of shoe are personal enough that people don't want to hand them off. He sees the machine-to-machine side of the economy as the far larger opportunity because it's starting from zero, rather than competing for an existing market.

Why Merchant Checkout Conversion Still Favors Humans

Host Simon Taylor points to data on the merchant side that reinforces the same conclusion. For long-tail Shopify merchants, shoppers who arrive from an AI conversation convert far better than typical traffic, because they've already done their research before reaching the site.

“People arrive later in the funnel; they're four to six times more likely to convert because they've done all their discovery already inside the LLM, and they just show up ready to buy.”

— Simon Taylor, Host, Tokenized

But when a large retailer tried to move checkout itself inside a chat interface, the result went the other way: conversion on Walmart's own checkout inside ChatGPT-style flows came in worse than on Walmart.com. Until that gap closes, merchants have little incentive to hand checkout to an agent, even as they benefit from agents sending better-qualified traffic.

• $45 million Series A. Sapiom raised a $45M Series A led by Dragonfly, announced two days before this Agentic Commerce Ep. 11 recording in August 2026.
• 270 million transactions processed. Sapiom's runtime has processed 270 million transactions and now runs 100,000 daily agent executions, per CEO Ilan Zerbib on the episode.
• $1.2 million inference bill in three months. One consumer-facing Sapiom customer saw inference costs grow from zero to $1.2 million in three months before Sapiom's model router cut spend by 60-90%.
• 6 to 12 months to scale. Zerbib estimates agentic commerce is six to twelve months from meaningful volume, with agent identity infrastructure still roughly a year from maturity.
• Sub-cent payments need a stablecoin. Bam Azizi argues machine payments as small as 0.0005 cents have no viable rail besides stablecoins — wires, Visa, and Mastercard can't clear a transaction that small.