Exclusive: Beltic Raises $7.3M to Help Businesses Keep AI Agents in Check

Beltic, a startup that helps businesses decide what AI agents can do on their websites and systems, is emerging from stealth Wednesday with $7.3 million in seed funding led by Norwest, the company tells This Week in Fintech exclusively.

Restive Ventures, Oxford Seed Fund and Collide Capital also participated in the financing, which brings San Francisco-based Beltic’s total funding to about $8.8 million since its 2025 inception. A slew of angel investors hailing from companies such as Coinbase, Legora, and Ramp aso participated.

The company is working to help solve a problem its founders say existing fraud tools were simply not designed to handle: software acting on a customer’s behalf. Tools such as Instinct and Muse are increasingly enabling AI agents to do just that. For example, an AI agent might search for a flight, book it, and pay for it. Before it can allow the last two steps, a business needs to know who sent the agent, what it's allowed to do, and whether the purchase matches what the customer wanted.

That’s where Beltic’s software comes in. It’s designed to not only identify visits from agents, but also to verify who is behind them and check their actions against a business’s rules. It can do things like allow an action, block one, or require another confirmation. The company calls this “know your agent,” or KYA, a variation on the “know your customer” checks used by financial institutions.

“Identity is not enough,” co-founder and CEO Isha Bhatnagar, a Coinbase alum, said in an interview with This Week in Fintech.

Bhatnagar and co-founder Mike Allan teamed up to start Beltic in early 2025 after working on different parts of the same problem: how to let legitimate transactions through while blocking fraud.

Bhatnagar spent nearly six years at Coinbase, where she worked on customer onboarding, fraud prevention and compliance products. Her work included systems that verified customers and assessed whether transactions should go through. Allan founded Atar, a banking-as-a-service provider that built payment infrastructure in Brazil and was acquired by insurance company Porto Seguro in 2021.

The two began talking in late 2024, when both were considering what would happen to those systems as AI agents began acting for customers. Allan moved from Brazil to San Francisco the following year to build Beltic with Bhatnagar. 

The pair bonded over what they both saw as a problem in the checks businesses relied on: the assumption that a human was at the keyboard. 

That assumption led checks to look for things such as a device fingerprint, a selfie, mouse movements, or typing speed, Allan noted. An agent most likely couldn't provide evidence of any of those things. That would leave a business struggling to decide whether to let it proceed. 

“This doesn’t work when you look at agentic traffic,” he said in an interview.

Still, identifying the customer behind an agent is only the first step, the founders said.

To help businesses keep agents in line, Beltic allows them to set rules for what agents can do. Beltic then checks each proposed action against those rules. As mentioned above, businesses can then review why Beltic decided to allow, block, or ask for confirmation on an action.

For example, a customer might authorize an agent to compare flights but not necessarily to buy one. The premise is that even an authorized agent could misunderstand an instruction or exceed its authority. It can also make a mistake that cannot easily be undone (see the case of Better Tomorrow Ventures principal JC Bahn de Stefano, who very publicly got himself temporarily banned on the reservation app Resy when his AI assistant went a little crazy).

“If the agent is behaving unexpectedly, doing something risky or… drifting from the user’s intent, we can flag it and require the agent to reconfirm the action,” Bhatnagar said.

In a video demonstration, Bhatnagar shared a Beltic dashboard that she said can distinguish agent visits from other website traffic, including agents arriving through a website or a connection to another piece of software.

So while a company might let an agent browse for a customer, it may require another check before the agent can actually make a purchase. 

The founders see that focus on individual actions as a distinction from companies that focus strictly on agent identity. While they acknowledge that establishing who is behind an agent is useful, Beltic operates under the premise that businesses also want to know what the agent is permitted to do and how to respond if it takes an unexpected action. He suggested the two approaches could work together.

“How do I monitor those actions? How do I make sure they don’t make those mistakes? How do I prevent this from happening? But at the same time, I allow them to come in,” Allan said, describing the questions he hears from companies.

Bhatnagar put the distinction more bluntly: “If your answer to a hijacked agent is that it had a valid credential, then you haven’t solved the agent risk. You’ve just solved agent identity.”

Beltic says its product is designed to work on both sides of a transaction. A company sending an agent out needs to be specific about what it may do. But the business receiving that agent needs to check its identity and authority before allowing it to proceed. The question becomes more complicated if an agent passes part of a task to another agent.

The founders also believe that because Beltic owns its entire infrastructure, it can move faster. 

“We can truly produce sub-second results associated with agent verification because we own the whole stack all the way down to the source,” Bhatnagar said.

Beltic added to its ability to identify the people and businesses behind agents through an acquisition last year. Shortly after raising about $1.45 million in pre-seed funding in 2025, it acquired Verifiet, a company co-founded by Farhan Afsahi, who became Beltic’s chief technology officer.

According to the founders, acquiring Verifiet gave Beltic a data platform covering more than 500 million entities across multiple countries. The startup uses that information as a starting point to identify who is behind an agent, then assesses what the agent is trying to do. 

The acquisition also combined separate areas of expertise the executive team considers central to its core product.

“We came together with Mike as the payment infrastructure person, Farhan as the data infrastructure, and I'm coming from identity infrastructure,” Bhatnagar said. “So as you can imagine, our product conversations are always very, very interesting. 

Currently, Beltic is running pilots with unnamed companies. Its nine-person team includes six engineers, the two founders and a lead responsible for finding customers.

“We’ve had great results understanding how to manage this traffic and block actions that could result in a mistake. That’s what companies like most, because right now they’re either blocking everything – like what happened with Resi, Instinct, Amazon and Muse – or trying to figure out how to handle it,” Allan said. “Our main goal is to help these companies understand that they can safely accept agentic traffic. That’s a big challenge right now, and I think that’s where we come in.”

Norwest principal Jordan Leite told This Week in Fintech that he was drawn to Beltic because, in his view, the startup sits at the intersection of two areas where his firm has strong conviction: the modernization of financial infrastructure and the emergence of agentic commerce. 

“As money and commerce move increasingly in real time, we believe verification and authorization infrastructure will have to keep pace,” he wrote via email. “What attracted us to Beltic was both the scale of that opportunity and a team that has lived this problem firsthand and is building the underlying infrastructure from the ground up.”

The fact that Beltic “owns the verification stack end to end” was a big differentiator for Leite.

“Rather than layering an agent credential on top of third-party identity and compliance providers, Beltic owns the underlying data, regulatory context and verification engine that determine whether a business or person can be trusted in the first place,” he added. “That foundation allows Beltic to extend verification into the agentic world, answering not only who is behind an agent, but what it is authorized to do and under what conditions.”