Most companies racing to deploy AI agents don't actually know what they're replacing. Not really.
That's the uncomfortable premise driving Autostep, a San Francisco startup emerging from Y Combinator that describes itself as "the P&L for knowledge work." Founded by Aidan Pratt and launched in June 2026, the company's workflow discovery platform attempts something decidedly unglamorous: forcing organizations to map the hidden costs of repetitive tasks before automating them away.
The pitch resonates precisely because it stings a little. How many hours disappear into email loops, ticket triaging, or document shuffling each week? What does that time actually cost? Most operations leaders can't answer with precision—and they're making seven-figure automation decisions anyway.
Autostep wants to change the order of operations.
Discovery Before Deployment
The platform's approach diverges from the prevailing wisdom that treats AI agent rollouts as self-evidently worthwhile. Instead of rushing to replace human judgment with software, Autostep installs desktop monitoring across an organization to quietly track every recurring process. Who performs each task? How often? Where does time leak away?
Every workflow gets tagged with a live cost-to-run calculation and projected savings score, translating hours into dollar figures that CFOs and COOs can use to justify—or, perhaps more usefully, reject—automation investments. It's data-driven decision making applied to the one area companies rarely quantify: what their knowledge workers do all day.
Once the mapping wraps, Autostep either hunts down an existing agent to handle the work or builds one directly inside the company's current toolset. The system bakes in rollback capabilities, approval flows, and human-in-the-loop oversight by default—guardrails that acknowledge automation doesn't always go smoothly.
According to the company's site, the platform has mapped north of 1,400 processes and reclaims an average of 38% of costs. Those are vendor-reported figures, of course, lacking independent validation. Still, the underlying logic holds: you can't measure ROI without a baseline.
There's an odd tension in the messaging, though. The Y Combinator launch post emphasizes that teams can "install once across the org. No integrations." Yet Autostep's website prominently advertises "120+ integrations" with CRMs, data warehouses, helpdesks, and internal tools. This could reflect discovery vs. automation phases—desktop capture for discovery, API connections for execution—but it reads like a company still figuring out how to explain what it does.
Backing and Growth Trajectory
Autostep has drawn investment from Y Combinator, Neo, and Walden Yan, co-founder of Cognition, the AI coding startup behind Devin. YC's LinkedIn amplification of the launch explicitly named the investors, though the company hasn't disclosed funding amounts. A Dealroom listing references an early entry showing $125,000 tied to YC, though there is no separate press release confirming this amount and it may simply reflect standard batch participation rather than a separate raise.
The team appears to be scaling quickly. YC's company profile lists two founders, while Autostep's About page claims 11 people as of mid-year. The startup is hiring across founding product engineers, ML engineers focused on process discovery, agent platform engineers, and a founding account executive for remote work—the kind of breadth that suggests ambition beyond a narrow use case.
Fighting for Oxygen in a Crowded Market
Autostep is hardly entering virgin territory. Established players like UiPath and Celonis have spent years building process mining and task discovery tools. UiPath's Task Mining captures desktop activity to identify automation candidates. Celonis touted customer ROI figures north of 380% over three years in a July 2025 Forrester study. Microsoft, never one to ignore a trend, is layering cost measurement features into Power Automate.
The newer wave of "agentic" platforms might be more direct competition. OutSystems launched Agent Workbench for enterprise orchestration. IFS debuted Agent Studio targeting service businesses. Project44 rolled out Autopilot for logistics workflows. The difference—and this is where Autostep plants its flag—is emphasis. Most platforms treat discovery as an afterthought. Autostep argues it should be step one.
Whether that positioning holds up depends on something the company can't fully control: whether operations leaders are willing to slow down long enough to measure what they're about to automate. In an environment where "agent" has become shorthand for competitive advantage, patience isn't always in strong supply.
Invite-Only, For Now
The platform remains invite-only, onboarding design partners through the year. Pricing starts at $0 for the Discover tier during beta, then jumps to $1,900 monthly for the Automate tier, with custom enterprise pricing for dedicated environments, SSO, and security reviews. The company highlights TLS 1.3 encryption, role-based access, full audit trails, and infrastructure built toward SOC 2 and GDPR compliance—table stakes for enterprise software, but important nonetheless.
The launch materials target COOs, CFOs, operations leaders, and founders. Essentially, anyone responsible for justifying the cost of knowledge work or explaining why AI agents will deliver value. It's a practical pitch for a practical problem, one that cuts against the breathless hype around automation: you can't improve what you don't measure.
And most companies, whether they admit it or not, are flying blind.
