Autostep Wants to Tell You Which Tasks to Automate Before You Waste Money Guessing

The Macro: The Automation Cart Before the Horse

There is a strange paradox in the way companies adopt AI right now. Everyone agrees they should be automating more. Consultants are billing insane hours telling executives the same thing. McKinsey publishes a new report about it every six weeks. And yet, when it comes time to actually pick which workflows to automate first, most organizations are guessing.

I have watched this play out dozens of times. A VP of Operations reads about some new AI agent platform, gets excited, buys a license, and then spends three months trying to figure out where to deploy it. Meanwhile, the real time sinks in the organization are invisible because nobody has bothered to measure them systematically. The accounts payable team is spending 12 hours a week copy-pasting invoice data between systems. The sales ops team is manually building the same report every Monday morning. Nobody tracks this. Nobody quantifies it. The automation budget goes to whatever project the loudest person in the room champions.

This is the discovery problem. Before you can automate anything intelligently, you need to know what your people are actually doing all day. Not what their job descriptions say. Not what they report in time-tracking software. What they are actually clicking on, typing into, copying between, and repeating over and over.

The tools that exist for this are mostly enterprise process mining platforms like Celonis or UiPath Process Mining. They work, but they are expensive, complex to deploy, and designed for large organizations with dedicated transformation teams. If you are a 50-person company that knows it is losing time to repetitive work but cannot justify a six-figure process mining contract, your options have been limited.

The Micro: One Founder, One Desktop App, One Big Idea

Autostep is a desktop application that installs across your organization and watches how people work. It surfaces repetitive, high-cost tasks, ranks them by impact and cost, and then builds or recommends the right fix. The founder is Aidan Pratt, who came through Y Combinator’s Spring 2025 batch as a solo founder out of San Francisco.

The product has three layers that I think are worth understanding separately. The first is what Pratt calls “operational intelligence,” which is essentially a passive observation layer. The app monitors work patterns across departments and calculates the financial impact of inefficiencies. Think of it as a flight recorder for knowledge work. It watches meetings, reports, outbound communication, data entry, all the stuff that fills a workday, and identifies what is repetitive.

The second layer is prioritization. Plenty of tools can tell you that something is repetitive. The harder question is whether automating it is worth the effort. Autostep ranks tasks by cost and impact, which means you are not just getting a list of repetitive work. You are getting a ranked list of the repetitive work that is actually costing you money.

The third layer is where it gets interesting. Once the platform identifies a high-priority task, it generates what Pratt calls “auto-agentic agents,” which are AI agents, prompts, templates, and workflow automations tailored to the specific patterns it observed. So the tool is not just telling you what to fix. It is attempting to build the fix.

There is also a compounding data story here. The longer Autostep runs, the more context it accumulates about how your organization operates. That historical data becomes a queryable record of operations, which means the recommendations should get better over time as the system learns what your actual workflows look like, not just what happened last Tuesday.

The go-to-market right now is demo-driven. You book a call with Pratt, he walks you through the product, and presumably you get a custom deployment. No public pricing, which at this stage makes sense. This is a product that needs to prove its ROI on a per-customer basis before it can standardize pricing.

The Verdict

I like the sequencing of this idea. Most automation tools start with a solution and go looking for a problem. Autostep starts with the problem, finds it, quantifies it, and then works backward to the solution. That is a fundamentally better approach for organizations that are drowning in “we should automate this” conversations but have no systematic way to figure out what “this” should be.

The risk is obvious: it is one person. Pratt is a solo founder building a product that needs to be installed across an entire organization’s machines, watch complex workflows, rank them intelligently, and then generate working automations. That is a massive surface area for a single engineer. Celonis has thousands of employees and they are still iterating on the same basic problem.

But the timing is right. AI agent platforms are proliferating faster than companies can figure out where to use them. The gap between “we have agents” and “we know what to do with agents” is real, and it is growing. If Autostep can credibly close that gap for mid-market companies, the demand is there.

What I would watch: how sticky the observational data becomes. If companies start relying on Autostep as their operational intelligence layer, switching costs go up fast. That is the moat. Not the automation itself, but the accumulated understanding of how your specific organization works. Nobody else will have that data.