Andrew Berman has spent a decade making streams of activity easier to understand. First it was a camera watching a room. Then it was a meeting that could be searched after everyone left. Now it is the work of AI agents, those tireless new colleagues who can retrieve data, call tools, spend money and create an administrative headache before lunch.
The products are different enough to look like pivots in a particularly restless career. Nanit is consumer hardware. Vowel was collaboration software. Runlayer is enterprise infrastructure. But the recurring instinct is clear: when technology creates a flood of new information or action, Berman looks for the interface that makes it useful, visible and accountable.
He is now CEO and co-founder of Runlayer, a New York company that sits between enterprise AI agents and the systems they want to touch. Its job is part switchboard, part security desk and part record keeper. Employees get an approved way to use agents. Technology and security teams get identity controls, permissions and a trail of what happened. It is a sensible answer to a very modern office question: if software begins acting for everyone, who keeps the keys?
The useful habit of arriving one problem early
Berman did not begin as a product builder. He worked in technology investment banking at Lehman Brothers from 2007 to 2009, then moved to venture capital at Norwest Venture Partners. He earned an MBA from Wharton. Those years put him close to products, markets and founders, but observation eventually gave way to operation.
In 2014 he co-founded Nanit with Assaf Glazer and Tor Ivry and served as its operating chief. The idea was to make a baby monitor do more than stream video. Computer vision could turn the picture into practical information for parents. The technical ambition lived inside a familiar object, which is often where ambitious technology earns trust. Machine learning disappeared behind a familiar glance at the crib.
The company also gave Berman an organizational problem. Nanit's team worked across locations and time zones. By 2018, it was looking for a better way to combine live conversation with the notes, decisions and context people needed later. Existing tools divided the work into calls, recordings, transcripts and follow-up documents. The inconvenience became the seed of his next company.
What happens after everyone leaves the meeting?
Vowel treated a meeting less like a calendar event and more like a piece of company memory. It combined video, transcription, agendas, notes, action items and search. Someone arriving late could catch up. Someone absent could find the exact moment a topic appeared. A decision did not have to evaporate when the call ended.
The timing was awkward, then suddenly exact. Early prospects thought the idea sounded strange. Distributed work was still a choice for many companies. The pandemic turned it into daily life, and Berman's phrase for the new condition was “death by meeting.” Vowel raised a $13.5 million round in 2021 and built for a future in which conversations became durable, shareable knowledge.
“Our goal is to make meetings more inclusive and worthwhile, which includes the prep, the meeting and the follow-up.”Andrew Berman on Vowel
There was a practical marketing lesson in the product too. Berman has said that lofty, futuristic language did not convert. Showing a user the useful moment quickly did. It is an unfashionably concrete point: people may enjoy a vision, but they adopt an advantage they can feel. Vowel's advantage was the moment a person searched a conversation instead of scheduling another one.
Zapier acquired Vowel in 2023. Berman moved inside the automation company as Director of AI, where he worked on AI products with teams at OpenAI and Anthropic. That placed him near the emergence of Model Context Protocol, or MCP, the standard that lets an AI system connect to tools and data. Once again, a capability arrived before the office had decided how to manage its consequences.
The control room era
Berman left Zapier in August 2025 with Tal Peretz and Vitor Balocco. Four months later, Runlayer launched from stealth with $11 million in seed funding. Its early customers included Gusto, dbt Labs, Instacart and Opendoor. The founders focused on the neutral layer around models, tools and company systems rather than inventing another model.
The need becomes clearer when an agent stops merely answering and starts acting. A chatbot can draft a response. An agent might enter Salesforce, pull customer records, query a warehouse, open a ticket and trigger a workflow. Each step may require different permission. Each action may cost money. Each mistake may need to be reconstructed. Traditional software assumes that users and requests behave in fairly predictable ways. Agents bring initiative, which is useful right up to the moment it is expensive.
“Everyone talks about AI, but AI is really only as useful as the tools and the resources it has access to.”Andrew Berman on connected agents
Runlayer's pitch is that governance can make adoption easier. If a company has one approved path, employees do not need to improvise connections while security teams chase invisible use. It is a careful balance. Too much friction sends useful experimentation elsewhere. Too little control gives a responsible buyer good reasons to wait.
Runlayer capital raised - cumulative
In June 2026, Runlayer raised a $30 million Series A led by Felicis with Khosla Ventures participating, bringing its total funding to $42 million. Berman's forecast is larger than a single gateway. He believes each employee will eventually supervise many agents, not as a novelty but as a normal way work gets done. If he is right, the quiet infrastructure around those agents will matter as much as the agents themselves.
The company as its own laboratory
Berman's public writing is especially revealing when he describes what Runlayer changed after trying it. Every new employee receives an onboarding agent organized into 16 sections. It carries role context, an interview recap, product strategy, customer context, the current sprint, a team directory, first-week milestones and the useful trails through company systems. The agent's value comes from context that somebody packed with care.
The company also pushed agents to write and maintain internal documentation, then pulled back. A document could be half-right and still look official. Another agent could read it as fact, repeat it, and slowly turn a small error into institutional memory. Runlayer shifted toward deterministic, human-generated context. The distinction is a good one: permission to read a company's memory is different from permission to rewrite it.
This is the operator inside the evangelist. Berman is plainly enthusiastic about AI, but his useful observations concern the point where enthusiasm meets permissions, incentives and ordinary human inattention. Being early does not require pretending the first version was right. Sometimes it means noticing the error before everyone else repeats it.
A public collision, quickly closed
Runlayer's year also supplied a blunt lesson in enterprise selling. In July 2026 the company sued Rippling after an extended product trial, alleging that the larger software company had misused access while developing a competing gateway. Rippling denied the allegations and filed patent claims of its own. In August, both companies dismissed their cases with prejudice. No compensation changed hands, and each side kept its own legal costs.
The brief fight ended without a courtroom answer, but the business question remains. A young infrastructure company must reveal enough during a trial to prove that its product works. Its prospective customer may have the engineers and incentive to build something adjacent. AI shortens build cycles while enterprise sales cycles remain stubbornly long. Trust, here too, needs architecture.
One long argument in three products
Berman's career can be read as a series of category changes, but it makes more sense as one long argument. New technology becomes valuable when people can understand what happened and decide what happens next. Nanit interpreted a stream. Vowel preserved a conversation. Runlayer records and controls the work of agents.
He has also remained an investor and adviser, with more than 40 investments across AI, security and robotics. His stated reason for advising founders is simple: many people helped him build Nanit and Vowel, and he wants to pay that forward. It gives the financier at the beginning of his career a pleasing return, this time measured in pattern recognition passed along.
A human detail in the Runlayer story is one Berman offered himself. He signed the term sheet on the day his first child was born. New family, new company, same calendar square. It is the kind of collision no productivity tool can tidy completely. The founder's answer was apparently to begin anyway.
Runlayer now faces the ordinary test behind a considerable amount of AI talk: whether a product can become dependable plumbing. Berman's bet is that the office of the near future will contain more agents, more connections and more delegated action than any one person can follow. His work is to make that invisible organization legible. The clever part may be the AI. The useful part is knowing who gave it the key.