Autostep (YC Spring 2026): teardown vs XMarts

In one line: Autostep is a desktop app that watches how work gets done across a company, puts a dollar cost on each repeated task, and then recommends or builds a fix (an agent, a process change, or software the firm already has) [site] [yc]. Threat: 2, because it targets the same ops and finance buyer with a "find the cost leak, then fix it" story, but it is one founder, $500k raised, and no named customers [yc] [jobs].

1. Fact file

Item	Finding	Source
What they say	Home page: "Discover every process. See costs. Fix what matters." YC line: "The P&L for knowledge work".	[site] [yc]
Founded, team, funding	Founded 2025, San Francisco, YC Spring 2026 (partner David Lieb). YC lists team size 1; other profiles say 2 to 10 (UNVERIFIED which). They raised $500k pre-seed on 1 Jun 2026, backed by YC and Neo plus angels, including Walden Yan (Cognition co-founder).	[yc] [press] [jobs]
Founders	Aidan Pratt, solo founder and CEO. Georgia Tech, machine learning and computer science. 8VC Fellow, cohort 2023. Built and shipped many products since childhood and worked at seed to Series A startups. What YC likely saw: a young, fast builder with an unusual measurement angle on AI spend (inference).	[yc] [press]
Customers	None named, no case studies. One claim: the smallest team they worked with found "$100K+" of waste with 10 people. The YC launch post says they are "already working with companies". One press profile says he claimed revenue tripled (UNVERIFIED, page not opened, seen in a search summary).	[site] [yc]
Price	Not public. The "Pricing" link goes to a booking calendar. Job ad names "mid-market and enterprise customers". A Founding GTM/SDR role is open.	[site] [jobs]
Path	Company footer reads "accrued, inc." and the About link goes to accrued.ai, which calls Autostep "an applied lab working on automating automation". Earlier pitch: "the last program", an always-on human-AI feedback loop. The YC page lists "Accrued" as a former name. No date for the rename. Pitch now: ROI of work, human and agent.	[site] [yc]

2. Their proposition, reverse engineered

Part	What they do
Who buys	COOs, CFOs, ops leaders and founders who must prove AI paid off [yc]. Mid-market and enterprise, per the job ad [jobs].
The pain	"Most companies know what they spend on AI but not what it returned" (paraphrase of the launch post). A board asks if AI paid off and nobody can say [yc].
The promise	See the cost of every repeated task, human and agent, then fix the costly ones. No interviews, consultants or integrations needed [site].
How it works	Install a desktop app across the firm (Mac and Windows) [site] [jobs]. It maps work, lists each process, prices it in dollars and ranks by impact. Then it recommends changes, deploys them into existing tools and measures against a baseline (they claim) [site].
The output	A ranked list of tasks with dollar cost and ease of fix. Agents, prompts and process changes built from observed patterns ("Auto-Fix"). A searchable, daily-updated record of how the firm runs, with sources ("Compounding Context") [site].
Proof style	Dollars of waste found ("$100K+" at a 10-person team, they claim) [site]. No saved-money result, no named case.
Who sells it	Founder-led demos through a calendar link [site]. Hiring a founding sales rep and an engineer who works directly with customers [jobs].
Price model	Not public [site].
The story	AI spend is unmeasured, so agent roll-outs are random. Long term: a "system of record for work", and one day "paying people for working smarter" [yc].

3. Positioning

Category they claim: "The P&L for knowledge work", operational intelligence [yc] [site].
Category a YC partner would put them in: task mining plus AI-ROI measurement for ops leaders, with an agent-building upsell (inference: the desktop watcher is the same method as task-mining tools).
What they want to be known for: the number that tells a firm which work, human or AI, earns its cost.
Where they sit on the map: layers L1/L2/L3 = 2/1/1, buyer = mid-market (inference from job ad), business.

4. Three layers vs ours

Layer	Them (score + evidence)	Us (Marton's logs show)
L1 Company brain	2. "Compounding Context" is a searchable record of operations, updated daily, answering with sources. Built from watching screens, not from calls or documents [site].	A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning.
L2 Agent orchestration	1. Builds single agents and prompts from observed tasks. No public claim of running a team of agents with approvals and handovers [site].	About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning	1. They claim to deploy fixes and measure against a baseline [site]. Nothing public shows a built fix running or a measured saving. Agent-ROI measurement of other people's agents is "checks only" [yc]. Would be 2 if one customer result is shown.	Every failure gets a record and an automatic check (about 160). Offer feedback lines fell from 44 to 18 as rules were added. Only about 15% of jobs reached the main code, and merges were unreliable in October.

5. Same and different

Topic	Same or different	Them	Us
Core belief	Same	Work, human or agent, must be measured and fixed with a number [yc].	Same (brief: "pitch the full loop on one workflow with one number").
Buyer	Same	COO, CFO, ops leader [yc].	Target: SMEs and scale-ups, commercial or finance leaders (brief, W7). No customer yet.
Result they sell	Different	Cost of waste found, then a fix [site].	The brief's aim is a money result on one workflow, measured before and after (not yet proven).
How they capture	Different	Desktop app watches every machine [site].	Marton's logs show calls, documents and checked work feed the brain. No install on staff machines.
Kind of work	Different	Repeated computer tasks (reports, email, data entry, slides) [site].	W7: margin leaks in pricing and demand (brief).
Who does the fix	Same, by claim	They say Auto-Fix builds agents in existing tools [site].	Marton's logs show agents do and check work, on his own business only.
Size of team	Different	One founder, hiring [yc] [jobs].	Two founders and a running system.
Proof	Same gap	"$100K+" found at one 10-person team, no name [site].	No customer number yet.
Delivery and price	Unknown	Self-serve software plus forward-deployed engineer; price not public [jobs] [site].	Not set yet.

6. Wedge overlap

Wedge	Fit	Why
W7 Find and fix margin leaks for commercial or finance leaders	direct	Same buyer kind (finance and ops leader), same story: find the cost leak, then fix it [site] [yc]. Difference: they price time spent on repeated tasks, W7 targets money leaking from pricing and demand (brief). A YC partner may still call them a rival (inference).
W2 Seeing whether your AI agents produce	adjacent	Their "agent ROI" shows which AI workflows pay and which lose money [yc]. Their buyer is an ops leader; W2 targets teams running coding agents (brief).
W3 Company memory from every call	adjacent	"Compounding Context" is a company record built from screen watching, not calls [site].
W8 The other 90% of an AI skill	adjacent	They generate agents and prompts from observed work [site]. No public claim of keeping them fresh from real use (UNVERIFIED that none exists).

7. Whitespace (where they stop)

Layer gap: they find, rank and recommend; the "fix" and "measure" steps have no public proof [site]. To own it we need one customer workflow run by our agents for 4+ weeks with a before/after number.
Buyer gap: the only named size is a 10-person team found waste, while the job ad says mid-market and enterprise [site] [jobs]. A desktop app on every machine is a heavy ask for a small firm (inference). To own it we need SME or scale-up buyers who would rather start from a workflow than install a watcher.
Number gap: "waste found" is not money saved [site]. To own it we need a result a finance person would sign: money moved, not hours identified.
Price gap: price not public, booked through a call [site]. To own it we need a published, result-linked price a 50-person firm can pay.
Speed gap: time to first value not stated; "no interviews, consultants or integrations" implies days (inference) [site]. To own it we need first value in under 14 days on one workflow, and to show it.

8. For the YC application

Proven today: Marton's logs show his system already runs agents that do and check work: about 5,600 jobs, each checked by a different AI, with about 160 failure records that each added an automatic check. Autostep's public pages show cost found and fixes recommended, with no named customer and no saved-money result [site] [yc].
Draft answer to "how are you different from Autostep?": DRAFT. "Autostep measures the cost of work and points to fixes. Our system already does the work and checks it, every day, on our own business." Founders say it their own way.
Future claim (needs proof): "We find the leak, fix it and show the money back." Needs one customer workflow run for 4+ weeks with a before/after money number a finance person would sign.
What we understand that they don't: the number a buyer trusts is money that came back after a checked fix, not waste identified by watching screens. needs proof on a customer.
Copyability (M3): easy. "Auto-Fix" and "measure against a baseline" are already on their site [site], they target the same ops and finance buyer, and they are hiring a forward-deployed engineer to work with customers [jobs]. It is a feature for them, so M3 can not score above 3 against Autostep. What slows them is a one-person team and a product not yet shown with a customer result (inference).
Why now (M4): None found in this teardown. Do not use general AI progress.
Do not claim: "we are the P&L for work" (their YC line). "Company brain" as a headline. Any saved-money result or customer. "Nobody measures agent ROI": they say they do.

M2 (what YC funds now) is NOT judged here. It comes from the themes of many companies in Pass 3.

9. Ideas worth taking

Say the buyer's board question in one line: "did AI pay off?" It gives a CFO a reason to act [yc].
Rank every fix by dollar impact and ease. A buyer can act on the top of the list [site].
Sometimes the answer is "no agent". Telling the buyer so builds trust [site].

10. Unknowns

Whether any customer pays, and how much. Ask on a call, or check the product video linked from the YC page.
What the desktop app records (screens, keystrokes, apps), and how staff data is protected. The privacy policy covers the website only. SOC 2 Type II is "in progress" [site].
Real team size and whether the "tripled revenue" claim is true. Look at later launch posts and press.
UK or Europe presence. Nothing found; San Francisco only.
Date of the rename from Accrued. Look at accrued.ai history and registry records.

Sources

autostep.ai [site]
accrued.ai [site]
autostep.ai privacy policy [site]
YC company page [yc]
YC launch: Autostep, Measure the ROI of Knowledge Work [yc]
YC jobs page [yc] [jobs]
Founding Engineer, ML + Forward Deployment job ad [jobs]
Clay community hiring post by Aidan P., 25 Aug 2026 [jobs]
Fast AI Jobs company profile [press] (funding: $500k pre-seed, 1 Jun 2026)
Extruct company profile [press]
8VC fellow profile, Aidan Pratt [press]
Yespress profile [press] (page blocked, 403; claims on revenue seen only in a search summary, UNVERIFIED)

Autostep

What they do

How close to us

Wedges

Who buys

Full teardown