StackAI

The agent that never asks for a raise

Somewhere inside a Nubank back office, a software agent is reading a 40-page policy document, cross-checking it against a customer ticket, and drafting a reply. Nobody coded it from scratch. Somebody dragged a few boxes onto a canvas, connected them, and pressed deploy. That somebody was probably not an engineer.

This is StackAI in mid-2026: a San Francisco company whose product quietly runs inside other companies. It builds the workers that read the documents, file the tickets, and run the due diligence that nobody enjoys. In May 2026, Asana bought it for a reported $75 million. The two founders packed up three years of work and walked into a much larger building.

Everyone wanted AI. Nobody could wire it up.

In 2022, large language models arrived with a flourish and a problem. They were brilliant in a demo and useless in a vault. A bank's most valuable information sat behind firewalls, in formats nobody had touched since 2009, governed by rules that lawyers cared about deeply. The models couldn't reach any of it.

So companies did what companies do. They hired consultants, ran pilots, and produced slide decks about "the AI journey." Most of those journeys ended at the demo. The gap between "this is impressive" and "this is in production" turned out to be enormous, and it was made almost entirely of plumbing: connectors, permissions, retrieval, governance. The boring parts. The parts nobody wanted to build.

Two robotics PhDs change their minds

Antoni Rosinol and Bernardo Aceituno met at MIT, where they were both earning doctorates in robotics. They had spent years on the hard problem of getting machines to act in the physical world. Then they looked at the messy, document-choked physical world of the enterprise and reached a slightly heretical conclusion: the more useful robot might not have arms at all.

They started StackAI in 2022 and joined Y Combinator's Winter 2023 batch. The bet was specific. Instead of selling another model or another chatbot, they would sell the assembly line - a no-code canvas where the person who actually understood the work could build the agent that did it. The lawyer, not the lawyer's engineering ticket.

Drag. Drop. Deploy. Govern.

StackAI's interface looks deceptively like a flowchart tool. You pick a trigger, add a model, wire in your data, connect the systems where the work lives, and ship. Underneath that simplicity sits the plumbing nobody wanted to build: 100-plus integrations into Salesforce, Slack, Google Workspace, Microsoft 365, and the ERP and ITSM systems that run the enterprise.

The agents do real jobs. They redline contracts and accelerate due diligence. They triage support tickets and answer them at 3am against official policy. They run financial research and draft RFP responses overnight. And because the buyer is a bank or a defense contractor, StackAI shipped the unfashionable features first: role-based access control, SOC 2 Type II, on-premise deployment, and local LLM endpoints for data that legally cannot leave the building.

Receipts, not adjectives

A startup can say "enterprise-grade" all it likes. The customer list is what counts. StackAI's includes BAE Systems, the defense and aerospace group; Nubank, one of the world's largest digital banks; and LifeMD, a telehealth provider. These are organizations that read security questionnaires for sport. They are not in the habit of trusting a three-year-old company with their documents unless it works.

By the time Asana came calling, more than 100 enterprises were running StackAI agents, and the platform had processed over a million documents. The investor list told the same story from a different angle: the Series A pulled in Guillermo Rauch, CEO of Vercel, and Bob van Luijt, CEO of Weaviate - two people who build the internet's plumbing and tend to recognize good plumbing when they see it.

An AI agent for every job

StackAI's stated goal was deliberately broad: put an AI agent in every job. Not to replace the people, but to hand them the dull half of the work - the reading, the cross-checking, the form-filling - and let agents and humans split the labor. IT support, sales prospecting, due diligence, financial research. The jobs that have a clear answer buried in a stack of documents nobody has time to read.

That mission is also why Asana wanted them. Asana's pitch is the "operating system for human-agent teams," and StackAI gave it the part it was missing: agents that don't just chat, but reach into other systems and actually do things. "This acquisition accelerates our roadmap and takes us into the next phase of human-agent work," said Asana CEO Dan Rogers.

The boring parts, finally usable

The hard truth StackAI noticed early is that AI's future inside companies was never going to be decided by who had the smartest model. It was going to be decided by who made the plumbing usable - who let the person who understood the work build the thing that did it. That is a less exciting story than artificial general intelligence. It is also the one that pays the bills.

Now go back to that Nubank back office. The agent reading the policy document is still reading. The person who built it has moved on to building the next one, and the one after that. The work that used to require an engineering ticket and a quarter of waiting now takes an afternoon. That is the change StackAI made - quietly, inside other people's buildings, where almost nobody saw it happen.

See it in motion