Here's the thing about enterprise AI agents: most companies have no idea which tasks are actually worth automating. They hire consultants who charge six figures for process audits. They interview employees who describe what they think they do all day—rarely what actually happens. Months later, armed with binders full of flowcharts, the moment has usually passed.
Autostep, a startup fresh out of Y Combinator's Spring 2026 batch, thinks it's found a shortcut. Install desktop monitoring software across your organization. Wait some period of time. The platform will tell you exactly which repetitive tasks are costing real money and, crucially, how to fix them.
Whether that sounds like a revelation or a privacy nightmare probably depends on where you sit.
Watching Work Happen
The pitch is straightforward, perhaps deceptively so. Autostep's software runs on employee desktops, observing work activity across the company. After some period of monitoring—the company doesn't specify how long—it surfaces repetitive tasks ranked by dollar cost and ease of resolution. Then it recommends specific fixes: process tweaks, AI agents, task reassignments, or outright elimination of busywork.
The company calls this "discovery-before-automation," positioning itself as the layer that runs before enterprises commit budget to building or buying anything. The output isn't just a list of problems. Autostep generates custom agent prompts, artifacts compatible with the Model Context Protocol specification, vendor recommendations, and templates that supposedly plug into existing tools. Over time, the system builds what the company terms "Compounding Context"—a queryable record of how the organization actually operates, growing more useful each week.
Founder Aidan Pratt, who studied machine learning and computer science at Georgia Tech, frames the bottleneck differently than most automation vendors. The problem isn't building agents or finding tools, he argued in the company's Y Combinator launch announcement. It's knowing what to automate in the first place.
That seems... obvious? And yet.
The Quantified Waste Promise
"No interviews. No consultants. No integrations," the homepage declares. The product calculates operational waste in dollar terms rather than abstract efficiency scores—a smart positioning move in an industry drowning in vague productivity promises.
One claim stands out: "Even the smallest team we've worked with uncovered $100K+ in operational waste with 10 people." That's a striking number—a self-reported claim with no external validation published. For a company with no public customer references yet, it's the kind of statement that demands follow-up questions.
The desktop-monitoring model gives Autostep visibility into meetings, reports, administrative tasks, and outbound work across departments—not just workflows already logged in enterprise software. Traditional process mining tools analyze structured data from ERP or CRM systems. Task mining software like UiPath's product captures application usage. Autostep positions itself a step earlier, discovering repetitive work that hasn't been formalized into documented processes yet.
Which raises the obvious question: how do employees feel about software watching their every click?
The company acknowledges the sensitivity, at least somewhat. Its homepage footer displays SOC 2 and HIPAA compliance badges alongside promises that users control data access and can "turn off anytime." No audit reports or detailed documentation are publicly linked. How the platform handles PII redaction, device coverage, and opt-out mechanics remains opaque on public pages. Pricing isn't disclosed; prospects are directed to book a demo via Calendly.
A Crowded Moment
Autostep's launch coincided with what might be called the Great Enterprise Automation Reckoning of May 2026. Cresta released Automation Discovery to identify which customer service conversations need AI agents. Glean introduced a seven-stage agent development lifecycle starting with "Opportunity" identification. Microsoft updated Power Automate with object-centric process mining. UiPath announced integrations for coding agents like Claude Code. Automation Anywhere enhanced its Context Intelligence Graph for AI-driven processes.
The pattern is hard to miss: enterprises need to know what to automate before they build. If 2025 was the year companies experimented with AI agents, 2026 appears to be the year they're asking which experiments deserve real budgets.
Traditional RPA and process mining platforms still dominate with mature discovery tools and execution platforms. Autostep differentiates by positioning as the prerequisite layer—the thing that tells you which agent or process change to pursue before you write a check.
Adjacent products address different lifecycle stages. ActivTrak launched AI Insights in May to measure how AI changes productivity after deployment. Honeycomb added Agent Observability to monitor agentic workflows in production. Autostep focuses on the front end, before the build decision gets made.
Whether there's room for a standalone discovery layer, or whether this eventually gets absorbed into existing platforms, remains an open question.
Early Days
The company launched publicly around February or March 2026, based on timestamps from its Y Combinator announcement. When Y Combinator shared the launch on LinkedIn, it drew more than 160 comments—discussion centered predictably on compliance concerns and the appeal of quantifying waste before committing to automation projects.
Autostep runs a Discord community and collects partner interest through a Google Form. The site displays a "Recognized by Vercel" logo but lists no named customers or case studies. LinkedIn shows a team size of two to ten employees, while the Y Combinator profile lists one person. The discrepancy likely reflects standard lag in self-reported headcount, or perhaps rapid recent hiring.
The company's About page, hosted at accrued.ai, hints at a more ambitious vision: "build the last program: a human-AI feedback loop" that automates automation itself. Whether that's inspiring or unsettling probably depends on your worldview.
The Unanswered Questions
For a product that installs monitoring software on every employee's machine, several details matter more than usual. The SOC 2 and HIPAA compliance badges appear as marketing statements without linked audit reports. The platform mentions generating Model Context Protocol artifacts—a specification that faced security scrutiny in May 2026 after reports of remote code execution risks in some implementations. How Autostep addresses those risks isn't discussed publicly.
The $100,000 operational waste claim needs context. What industry? What kinds of tasks? Over what period? Without customer references, those numbers float in a vacuum.
Still, the timing feels right. Enterprises spent the past year spinning up proof-of-concept agents, many of which now sit idle or underutilized. The harder question—which workflows actually justify the investment—is precisely the kind of unglamorous infrastructure problem that venture capital occasionally backs.
Autostep is betting that answering that question is itself a business. Whether it's a venture-scale business, or a feature that gets folded into larger platforms, is the kind of thing we'll know in a year or two.
For now, it's a YC-backed startup with monitoring software and a sharp premise, launching into a market that's finally asking the right questions. Even if some employees might wish those questions didn't require quite so much... observation.
