XMarts room

Fall 2025

Cranston AI

AI Tax prep and accounting software for CPA firms

Visit website

What they do

AI tax prep and bookkeeping agents for CPA firms

How close to us

5, very close

agents plus reviewer handover for one firm type (CPA); close to W5 shape

Wedges

  • W5 One layer for one type of small firm direct
L1 Company brain (knows the organisation) 1, light
L2 Agent team (runs agents, tools, approvals) 2, clear
L3 Does and learns (finishes, checks, measures, updates the brain) 2, clear

Who buys

Firm size
sme
Buyer type
business
Role
CPA firm owner or managing partner
UK
no
Their own description (from YC)

Cranston is AI tax prep and bookkeeping for CPA firms.

Our tax agents prepare returns the way a strong staff accountant would: they read the source documents, work the return in your tax software, and hand your reviewer a return that holds up, with workpapers attached. Our bookkeeping agents keep client books current all year, so tax season starts from clean numbers instead of cleanup.

Prep and books are the start. We're building Cranston into the system a firm runs on: the place where prep, review, client requests, and the rest of the practice get done.

Full teardown

Cranston AI (YC Fall 2025): teardown vs XMarts

In one line: Cranston sells AI agents that prepare tax returns and keep the books for US CPA firms, inside the tax and accounting software the firm already uses, with a human reviewer approving everything [site] [yc]. Threat: 2, because it is the closest live example of W5 (one layer for one type of small firm, with accountancy as the example), but it is two people, US only, with no named customers and no result numbers [yc] [site].

1. Fact file

Item Finding Source
What they say "Tax prep and accounting software that runs itself." YC line: "AI Tax prep and accounting software for CPA firms." [site] [yc]
Founded, team, funding Founded 2025, San Francisco, YC Fall 2025, team of 2. YC primary partner: Tom Blomfield. Funding is UNVERIFIED and sources conflict: Simplify says $130K total; aggregators list a seed (Oct to Dec 2025) with Palm Drive Capital, Rebel Fund and YC, size reported as about $500K; one aggregator says $4M total. No company announcement found. [yc] [press]
Founders Max Minsker (CEO on the YC page; one profile says COO): owned a tax practice that filed 10,000+ returns, and the company page says it grew to 3,000 clients. Sean O'Bannon (CTO): Stanford CS and AI, founder and CTO of ReMatter (an ERP for scrap yards), worked at Databricks. What YC likely saw: a founder who lived the CPA pain plus an engineer who has built a vertical ERP (inference). [yc] [site] [press]
Customers None named on the site, the demo page or the YC page. A job-ad summary says they are "deployed with accounting firms across the country" (UNVERIFIED, seen only in a search summary). One Product Hunt comment says "happy customers" with no firm named. [site] [yc] [jobs] [press]
Price Public. Bookkeeping: $30 per client book per month. Tax prep: $40 per prepared return, any form. "No credits or tokens." Past 2,000 returns a season, price is agreed one to one. [site]
Path Early 2025 listings call it "a full-stack AI accounting firm" for startups and SMBs, with licensed CPAs supervising AI (a firm, not software). Product Hunt launch Nov 2025 under that framing. By 2026 the site and YC page say software and agents for CPA firms, and "building Cranston into the system a firm runs on". [press] [yc] [site]

2. Their proposition, reverse engineered

Part What they do
Who buys Solo practitioners, growing firms and multi-office CPA firms in the US [site]. The signer is the firm owner (inference).
The pain Prep work and bookkeeping eat staff time. Missing client documents hold up returns. The site frames it as "serve more clients with your current team" [site]. No buyer quotes found.
The promise AI prepares the work and reviewers "make the calls". Nothing posts, files or goes to a client without approval [site].
How it works Agents read source documents, key the return into Drake, Lacerte, CCH Axcess or UltraTax, run document checks, and draft client emails for missing items. Bookkeeping agents classify transactions, reconcile and draft journals in QuickBooks or Xero, and leave an exception list. A tax research module answers questions with citations. One workspace per client holds documents, books, returns and email [site].
The output A prepared return with workpapers attached for the reviewer [yc]. Books kept current with an exception queue [site]. Draft client requests [site].
Proof style None public. No time saved, accuracy or revenue numbers on the site [site]. One third-party profile repeats "about 99% accuracy" for reconciliation; that is an aggregator line, not a company claim we could find (UNVERIFIED) [press].
Who sells it Founder-led. Book a 30-minute intro call [site]. A Chief of Staff ad says the hire will talk to accounting firms and "help shape our GTM motion" [jobs].
Price model Flat per unit: per client book per month, per return [site]. Not seats, not hours.
The story "Accounting software that does the work." Work inside the tools firms already use, nothing to replace, and the product "improves with each job" [site]. Every feature is kept or cut on live client feedback [site].

3. Positioning

  • Category they claim: "AI agents for accounting firms" and "tax prep and accounting software" [site] [press].
  • Category a YC partner would put them in: vertical AI for accounting firms, an AI-native back office for one profession (inference).
  • What they want to be known for: the system an accounting firm runs on, starting with tax prep and bookkeeping [yc].
  • Where they sit on the map: layers L1/L2/L3 = 1/2/2, buyer = sme, business.

4. Three layers vs ours

Layer Them (score + evidence) Us (Marton's logs show)
L1 Company brain 1. One workspace per client with documents, books, returns and email [site]. Claims it "improves with each job" but shows no firm-level memory, rules or knowledge base (UNVERIFIED). A vault of 1,396 pages, 112 call transcripts and 8,763 customer quotes, searchable by meaning. Built for one business, not a firm's clients.
L2 Agent orchestration 2. Separate tax, bookkeeping and research agents, document chasing, and reviewer approval before anything posts or files [site]. Handover to a human is the core design. 5,641 jobs run by agents from three makers (Codex, Claude, Gemini), each checked by a different AI before merge.
L3 Execution and learning 2. Does the work (returns, journals) and checks it: diagnostics for missing or wrong-year items, source links on every output, reviewer sign-off [site]. No measured result feeds back that we can see. Scored 2 on the "checks" wording; the checking is mostly the human reviewer. About 160 failure records, each with an automatic check. Offer feedback lines fell from 44 to 18 as 22 rules were added. Only about 15% of jobs reached the main code, and merges fell to 43 so far in October.

5. Same and different

Topic Same or different Them Us
Core belief Same Agents do the work, people approve it [site]. Same: work is built by one AI and checked by another (inventory).
Where it runs Same Inside tools the firm already owns [site]. Marton's logs show agents work in the owner's own tools and repos.
Buyer Different US CPA firms, solo to multi-office [site]. Target is UK SMEs and scale-ups (brief). No customer yet.
Owner of the result Same The firm's reviewer signs the return [site]. Not fixed. W5 says the owner sets the rules.
Kind of work Different One profession: tax prep and books [site]. W5 could be a layer over a firm type (brief); accountancy is only the example.
Learning Different Says it improves per job, no mechanism shown [site]. Marton's logs show rules added from failures and feedback.
Proof Same (both thin) No customer, no number [site]. No outside customer pays yet (inventory).
Delivery Different Software with a demo-led sale [site]. System plus the two founders (brief).
Price Different Public, per unit: $30 a book, $40 a return [site]. Not set yet.
Size of team Same 2 people [yc]. 2 founders (brief). Xavier does not code; Marton built the system.

6. Wedge overlap

Wedge Fit Why
W5 One layer for one type of small firm direct The brief's own W5 example is a layer above Xero for accountancy firms. Cranston is that, in the US: one firm type, agents on top of the firm's tools, "the system a firm runs on" [yc]. Different country and tax rules, same shape.

7. Whitespace (where they stop)

  • Buyer gap: US only. The product, the research module (IRC, rulings), the compliance claims (GLBA, IRS Pub. 4557) and the tax software list are all US [site]. No sign of UK work (HMRC, VAT, payroll, Companies House). To own it we need UK accountancy firms that say they have the pain, plus one run on UK rules.
  • Number gap: no result is public: no hours saved, no returns finished per reviewer, no error rate [site]. To own it we need one firm run for 4+ weeks with a before/after number a partner would sign (for example returns per reviewer per week).
  • (Not a gap) Layer gap: they already do work and check it, so we can not say "they only map". Weak spots: L1 is a per-client workspace, not a firm brain, and learning is claimed, not shown [site].
  • (Not a gap) Price gap: $30 and $40 per unit are low and public [site]. A smaller firm can pay. We can not win on price without a result-based price.
  • (Not a gap) Speed gap: unknown. Demo call first, no stated time to first value [site]. Worth asking a customer.

8. For the YC application

  • Proven today: Marton's logs show his system already runs agents that do work and have a different AI check it: 5,641 jobs, about 160 failure records that each added an automatic check. Cranston's site shows the same pattern (agents plus human sign-off) in tax and books, with no public result number [site].
  • Draft answer to "how are you different from Cranston?": DRAFT. "Cranston is a good example of the shape: agents on top of a firm's own tools. They are US tax only and show no results yet. We would do it for UK firms and publish a number from the first run." Founders say it their own way. Only say the UK part if W5 is picked.
  • Future claim (needs proof): "Our agents run a UK accountancy firm's routine work and the firm's reviewer time per client falls by X." Needs one UK firm run for 4+ weeks with a before/after number, and a named person with the pain on file.
  • What we understand that they don't: a firm layer should learn from checked results, with the rule change logged, not just from reviewer sign-off. needs proof on a customer; Marton's logs show it only on his own work (offer feedback lines fell from 44 to 18).
  • Copyability (M3): medium. They already name Xero and QuickBooks [site], and the site says they are "building Cranston into the system a firm runs on" [yc], so a UK step is possible. It needs UK tax software links, HMRC rules and a UK buyer, and pulls two founders in San Francisco away from US busy season (inference). Not hard, so M3 can not score above 3 against Cranston for W5. A rival marked hard for the learning-loop edge is still needed.
  • Why now (M4): None found in this teardown. Cranston cites no dated change. Do not use "agents can now do the work": that is general AI progress.
  • Do not claim: "nobody does AI for accountants" (Cranston does, with public prices). "Our agents learn from every job" for a customer (no customer yet). "Faster or cheaper than Cranston" (we have no price, no speed number and no tax work). Any accounting or tax capability: the inventory shows none.

9. Ideas worth taking

  1. Publish a flat price per unit (per return, per book) that a buyer can do the sums on at once [site].
  2. Block anything from posting or going to a client until a reviewer approves, with source links on every output [site]. It answers the trust question first.
  3. Sell by finding the missing input before the work starts (missing documents, chase emails) [site]. A visible first win.

10. Unknowns

  • Real customers, how many, and what they pay. Look at the founders' X posts and any YC launch post; ask a CPA who has seen a demo.
  • Funding amount and date. Look for a Form D filing or a company post.
  • Whether the "improves with each job" claim has a mechanism (rules, memory) or is marketing.
  • Whether any UK or non-US plan exists. Check the careers page and founder posts.
  • The FAQ answers for "how is Cranston different?" and "who signs the return?" were blank when fetched, so their own defence line is unknown [site].
  • Product Hunt launch date of 12 Nov 2025 comes from a search summary; the page itself showed "11mo ago" [press].

Sources

Technical detail
Slug
cranston-ai
Lens hash
179096c479
Pass
deep
Stale
no