Brandon Boehme still remembers the moment the idea crystallized. He and his co-founder Wasi Ahmed were building a voice AI agent for restaurant orders—the kind of conversational bot that takes your pizza toppings over the phone and doesn't accidentally add pineapple when you said pepperoni. Everything worked smoothly until checkout. Then the whole careful choreography fell apart.
"You can't just ask someone to read their credit card number to a bot," Boehme recalls. Not without triggering PCI compliance nightmares, anyway. Not without the card data potentially touching servers it shouldn't. And certainly not without breaking the conversational flow that made the AI worthwhile in the first place.
So they built Maven.
The two-person startup, which emerged from Y Combinator's Winter 2026 batch and launched publicly in late February, offers what its founders describe as a deceptively simple fix: a payment API that lets voice AI agents collect card details without ever handling the sensitive data themselves. It's infrastructure for a narrow problem, perhaps. But it arrives just as commerce giants are scrambling to retrofit checkout for a new era of conversational AI.
Everyone wants agentic commerce. No one's quite sure how to build it.
January was a busy month for payment announcements. Google unveiled its Universal Commerce Protocol at the National Retail Federation's big show, positioning it as an open standard for AI shopping assistants. Microsoft and PayPal, not to be outdone, introduced something called Copilot Checkout three days earlier. Mastercard told Axios that "agentic commerce will only scale at the speed of trust," while Visa quietly confirmed it had completed secure AI transactions with unnamed partners.
By March, mid-tier players like Spreedly and Oceanpayment had launched their own agentic commerce offerings. The pattern is consistent: established infrastructure providers are racing to adapt checkout flows for chatbots and voice agents, betting that conversational commerce is more than hype this time.
Maven's wager is different, though. While most of these solutions target text-based interactions—where you can embed a payment link in a chat thread or pop up a widget—Boehme and Ahmed focused on voice. Specifically, the kind of phone-based interactions that restaurants, home service companies, and (somewhat grimly) debt collectors still rely on. Places where you can't just paste a Stripe checkout URL.
"IVRs are dead," their homepage declares, with the confident absolutism of founders who've spent too many hours in interactive voice response hell.
The handoff
Maven's technical approach sidesteps the compliance headache through what the company calls a session handoff model. When a voice agent reaches the payment moment, it pings Maven's API, receives a temporary phone number, and transfers the caller. Maven's infrastructure then takes over—capturing card details through speech recognition or good old DTMF tones, tokenizing everything in-memory via Very Good Security, and handing control back once the transaction clears.
Sessions expire after five minutes, a PCI requirement that the founders can't negotiate around. But most complete in 60 to 90 seconds, according to their documentation. The webhook fires with gateway-specific metadata—PaymentIntent IDs for Stripe users, transaction codes for Authorize.net, status tokens for Braintree. Card data never touches the customer's servers, which is the entire point.
Developers who use VAPI, one of the voice AI platforms Maven integrates with, get auto-generated tools—collect_payment, transfer_to_payment, get_session—that plug directly into the agent's function-calling layer. There's a catch, though: you have to preserve caller ID using your own Twilio or Vonage numbers, not VAPI's telephony, to ensure Maven can match sessions correctly. Small detail, potentially big headache if you miss it.
The company supports Stripe, Authorize.net, Braintree, Shift4, and Adyen on the payment side. Voice platform integrations include VAPI, Retell, LiveKit, and Twilio. There's also an SMS fallback for when voice capture fails or customers simply prefer tapping numbers into their phone.
It's a narrow product solving a specific problem. Whether that problem is big enough to build a company around is another question entirely.
The incumbents were here first
Secure voice payments aren't exactly new territory. Twilio has offered a <Pay> verb for years. Companies like PCI Pal, Sycurio, Eckoh, and Paytia have built entire businesses around DTMF masking and speech recognition for contact centers—intercepting keypad tones or routing audio away from human agents during the payment moment.
But those products were designed for a different world. IVRs. Call centers staffed by actual people. The assumption was always that a human would be orchestrating the conversation, not an autonomous AI navigating multi-turn dialogue.
Maven's session lifecycle—with its API-driven handoffs, automatic transfers back to the agent, and webhook-based state management—assumes the agent is code. The documentation talks about transfer logic and callback URLs in the language of developers building conversational applications, not telephony admins configuring PBX flows.
Whether that distinction creates enough daylight between Maven and the legacy players is unclear. As of late March, the company has no public customer logos. No published pricing. The product is live, the documentation is thorough, and Y Combinator partner Ankit Gupta gave the launch a LinkedIn boost. But early traction? Anybody's guess.
A bet with two people
Boehme and Ahmed—both UC Berkeley computer science grads who spent time at Meta and Amazon—founded Lambda Systems, Inc., Maven's legal entity, sometime in 2025. The team size, per YC's directory, remains stubbornly at two.
No external funding beyond Y Combinator's standard investment has been disclosed. The Privacy Policy lists only Lambda Systems and a contact email. The Terms of Service were last updated December 31, 2025, which either suggests careful attention to legal detail or that not much has changed since launch. Hard to say.
There's a certain Silicon Valley romance to the two-person technical team tackling infrastructure, and Boehme and Ahmed clearly have the credentials. But infrastructure plays often require not just engineering chops but also the ability to navigate enterprise sales cycles, partnership negotiations with payment processors, and the slow grind of developer adoption. Maven will need to do more than build a good API.
Where does payment infrastructure end?
Here's the broader question Maven's launch surfaces, intentionally or not: Where does payment infrastructure stop being a feature and start being a product?
Stripe, Adyen, Braintree—they process transactions but don't handle voice capture. Twilio and VAPI manage telephony and conversational AI but weren't built for PCI-scoped payment collection. VGS tokenizes card data but doesn't orchestrate session handoffs or manage agent transfers.
Maven sits in the gap, stitching these pieces together for one specific use case. If agentic commerce takes off the way Google, Microsoft, and the card networks seem to believe it will, that gap might widen. Voice agents could proliferate across industries that still rely heavily on phone interactions—healthcare appointment scheduling, financial services, government hotlines. All of them need to collect payments eventually.
Or voice could remain a niche channel, overshadowed by chat and embedded UI where the checkout experience is cleaner and the infrastructure already exists. In that scenario, Maven's narrow focus becomes a limitation rather than an advantage.
For now, the product exists. Developers building voice AI for restaurants or service businesses can add payments through a session API that abstracts away the compliance mess. Whether that's enough to build a sustainable company—or whether Stripe or Twilio eventually absorbs the functionality into their own offerings—depends on two things: how quickly autonomous voice agents move from clever demo to actual deployment, and whether the compliance friction stays steep enough to justify a dedicated abstraction layer.
Boehme and Ahmed are betting yes. It's a small bet, at least for now. Two people, one product, a slice of the conversational commerce stack that most developers probably haven't thought about yet.
But sometimes that's how infrastructure gets built—quietly, by people who ran into a problem and decided the existing solutions weren't quite good enough.
