There's a peculiar irony in the rush to deploy AI agents across corporate America. Everyone agrees repetitive work is everywhere. Few can pinpoint exactly where it lives, what it costs, or which tasks would actually benefit from automation rather than just sound impressive in a slide deck.
Autostep, a small outfit fresh from Y Combinator, wants to solve the part that comes before the agents themselves: figuring out what's worth automating in the first place. The company's pitch is less about the sophistication of its AI and more about the mundane reality that most organizations are flying blind when it comes to operational waste.
"Even the smallest team we've worked with uncovered $100K+ in operational waste with 10 people," the startup claims on its homepage. No customer names accompany that assertion, and the company hasn't published case studies. But the underlying thesis is difficult to argue with—somewhere between consultant fees and employee surveys, there ought to be a cheaper way to map how work actually happens.
Desktop Surveillance, Rebranded
What Autostep built is desktop monitoring software with ambition. Install it across an organization's computers and it watches how people work: task patterns, workflow sequences, departmental rhythms. The software then ranks repetitive activities by dollar value and automation feasibility, surfacing what the company calls "operational intelligence" in queryable form.
Once high-cost targets emerge, Autostep generates the fix—an AI agent, a process template, a workflow automation, sometimes just a recommendation to change how something gets done or which vendor to use. The agents deploy inside existing tools, ostensibly maintaining themselves without much human oversight, though the company hasn't detailed what happens when things go sideways.
Founder Aidan Pratt studied machine learning and computer science at Georgia Tech before stints at early-stage startups. In launch materials, he frames the problem as a discovery bottleneck rather than a technology gap. Plenty of companies can build agents. Fewer know where to point them.
The platform integrates with Model Context Protocol, letting agents access data inside enterprise systems without rebuilding connectors from scratch. Autostep displays SOC 2 and HIPAA compliance logos, though publicly available audit reports or certificates don't appear to exist yet.
Crowded Timing, Different Angle
Autostep entered the market during what might generously be called a busy stretch. In May alone—around the time the startup was making its public debut—Microsoft made computer-using agents generally available in Copilot Studio, UiPath folded coding agents into its automation platform, and Automation Anywhere rolled out updates to something it calls Agentic Process Automation. Celonis added MCP-based agent tools to its process mining software. Camunda announced ProcessOS, positioning it as an orchestration layer with discovery baked in.
The list goes on. The difference, perhaps, is one of sequencing. Established players added agent capabilities to existing enterprise platforms. Autostep positioned discovery itself as the product, with automation following behind. Whether that resonates with operations leaders who've watched expensive agent projects fizzle because they targeted the wrong workflows remains an open question.
The company is backed by Y Combinator and Neo. Third-party profiles mention Walden Yan from Cognition as an angel investor, though this remains unverified in fully accessible sources. No formal seed round has been announced, pricing isn't public, and sales appear to run through demo requests via Calendly.
The Questions That Follow
With a small team in San Francisco and a launch measured in months rather than years, Autostep is still taking shape. There's no public customer roster, no open roles listed, limited information on how long the software retains activity data or what redaction mechanisms exist.
LinkedIn comments on the company's launch posts surfaced the predictable concerns: who sees the desktop-level monitoring data, how consent works, what happens when employees realize the software is cataloging just how repetitive their days actually are. Autostep hasn't published detailed documentation on role-based access controls or user consent flows.
Every desktop-monitoring tool eventually confronts these tensions. The value proposition—stop guessing, start measuring—is straightforward enough. But discovery has a way of revealing more than organizations bargained for. Tasks that seemed critical turn out to be make-work. Roles that appeared full-time contain hours of repetition no one wanted to name out loud.
That cuts both ways. For a startup selling operational transparency, the real test isn't whether the software can identify waste. It's whether companies are ready to act on what it finds, and whether employees will tolerate the watching that makes it possible. Autostep is betting the answer to both is yes. The market, as always, will have the last word.
