XMarts room

Summer 2025

Magnetic

AI Tax Preparer for CPA firms

Visit website

What they do

AI tax data entry from client documents for CPA firms

How close to us

5, very close

one reusable layer (document to tax software) sold to many small firms of one type; active, direct W5.

Wedges

  • W5 One layer for one type of small firm direct
  • W7 Find and fix margin leaks adjacent
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 and tax firm owners and partners
UK
no
Their own description (from YC)

Magnetic scans client documents - from handwritten charity notes to excel spreadsheet Schedule Cs to multi-state K-1s - and enters them into tax software automatically, saving hours of tedious data entry.

The accounting profession is facing a massive talent shortage, as graduates choose different professions and the old guard begins to retire. As a result, CPA and tax preparation firms are awash in clientele, but unable to keep up. At many firms, phones go unanswered, clients are "fired", and websites are taken down, as the existing workforce tries to handle the grueling 80 hour+ weeks necessary to push through tax season to the April 15th deadline.

We built this because a former colleague who owns her own tax firm asked us to automate her workload, and we quickly realized the massive opportunity. Thomas brings deep domain expertise as the former Head of Product at Keeper (YC W19), where he built tax engine automation systems and an OCR form scanning system that outperformed everything else on the market. At Keeper, he leveraged early access to OpenAI's LLMs (well before ChatGPT) to build systems that automated the entire tax filing process, allowing a tiny team of CPAs to process thousands of returns with near zero busy work. Patrick was the first engineer at FunCraft, shipping multiple mobile games with 1M+ downloads and building scalable shared codebases. We've worked together since 2017, including co-leading multiple teams at Zynga as PM & Tech Lead respectively.

The opportunity is massive - there are 80M US personal tax returns filed via professionals annually, in a $2B outsourcing market where individual CPA firms currently pay $100k-$1M per year for inferior services. We're positioned to capture significant market share by offering better accuracy & document coverage, instant processing, and superior user experience.

We've built this kind of software before, we've worked together for years (and even lived together), we've built products to $10M ARR before, and we have the grit to outwork anyone in this space.

Full teardown

Magnetic (YC Summer 2025): teardown vs XMarts

In one line: Magnetic drafts US personal tax returns (Form 1040) for small CPA firms with AI, has a licensed US tax professional review each one, and sends it back in 2 to 4 days [site]. Threat: 2, because it is the clearest live example of "one reusable layer for one type of small firm" (our W5), but it is US-only, one job (1040 prep) and one tax regime [site] [yc].

1. Fact file

Item Finding Source
What they say Home page today: "AI-Powered 1040 Tax Return Service." YC line: "AI Tax Preparer for CPA firms." [site] [yc]
Founded, team, funding YC S25, San Francisco, 10 people. Their job ads say they raised $5.2M. A third-party list shows only a $500K seed (July 2025, YC and Orange Collective); that figure is UNVERIFIED and may be an earlier round (inference). Backers on their About page: YC, Orange Collective, Formus Capital, Seguin Ventures, Amino Capital, Singularity Capital. [jobs] [site] [press]
Founders Thomas Shelley (CEO): ex Head of Product at Keeper (YC W19), where he built tax automation and form scanning; earlier product roles at Niantic and Zynga. Patrick Fay (CTO): first engineer at FunCraft, led engineering teams at Zynga. They have worked together since 2017. What YC likely saw: a founder who already ran an AI tax system at volume, plus a long-standing co-founder pair (inference). [yc] [site]
Customers Five logos on the home page: Gilbert CPAs (Sacramento), CGC Accountants & Advisors, Cornbelt Financial (Denver), Zhong & Sanchez (San Francisco), LJB CPA. Three named quotes. The customers page says "Placeholder - customer stories coming." They claim "many dozens of firms" and "hundreds of thousands in ARR". [site] [jobs]
Price Per return, paid in credits: a simple return costs 1 credit, a complex one up to 3. Cash price per credit not public. Job ads say CPA firms pay "$100k-$1M per year" for competing services. [site] [jobs]
Path Summer 2025 launch pitch: software that scans client papers and types the numbers into UltraTax, Drake and ProConnect. Today's site: a full-return service where their AI drafts and their licensed reviewers check, with "no new system to learn". Same buyer throughout. Dates of the change not found (inference from comparing the two pages). [yc] [site]

2. Their proposition, reverse engineered

Part What they do
Who buys Owners and partners of small CPA and tax firms that cannot hire fast enough [site]. Named customers are small firms in California and Colorado [site].
The pain "If I still have to review everything, it doesn't save me any time." (CEO, About page.) Staff shortage, 80-hour weeks, clients turned away [yc] [site].
The promise Extra return capacity "exactly when you need it", back in 2 to 4 days [site].
How it works Firm uploads client papers. Their AI reads them (vision models plus OCR), reasons across all documents, and enters the data into the firm's tax software using a "computer-use protocol". A second AI pass re-checks figures against the source papers. A US-based EA or CPA reviews. The firm imports an upload-ready file and does the final review [site].
The output A drafted return file with review notes, flagged assumptions and a list of missing items [site].
Proof style They count accuracy and time: "99.9%" on 94+ document types, 80%+ prep time saved (one firm owner), 2 to 4 day turnaround [site]. Their YC launch post said "90%+" field-level accuracy, so their own accuracy claim moved between pages [yc] [site]. All self-reported.
Who sells it Demo requests and self-serve sign-up on the site [site]. Hiring a first Account Executive ($90K to $300K) [jobs]. Comparison pages against Juno and Filed look aimed at search and sales [site] (inference).
Price model Credits per return, by complexity. An error that moves the client's tax by more than $10 gets the prep fee credited back [site].
The story Enemy: tools that only extract some fields, so the firm still does a "detailed review". Belief: AI alone is not enough in tax; human plus AI, so firms do only the "high-level review" [site].

3. Positioning

  • Category they claim: AI tax prep agent and "AI-Powered 1040 Tax Return Service" [site].
  • Category a YC partner would put them in: an AI-native service company for tax prep (AI does most of the work, licensed humans finish it), with software roots (inference).
  • What they want to be known for: finishing the whole return, including the hard ones, so the firm only does the final review [site].
  • Where they sit on the map: 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. Uses the client's documents and prior-year files for each return [site]. No public claim of a lasting memory of the firm that other work reads (UNVERIFIED that none exists). A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning.
L2 Agent orchestration 2. One agent reads, reasons, enters data and re-checks, then hands to human reviewers who work from "confidence scores and agent notes" [site]. They are hiring someone to build "routing systems" for reviewers [jobs]. Several steps with handovers, not a team of specialist agents. About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning 2. Does the work and checks it twice (AI re-check, then licensed human) [site]. The Head of Operations role is to "reduce unnecessary human review over time" and turn "operational learnings" into better AI [jobs]. That is an intent in a job ad, not a shown loop, so not 3. 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; merges were unreliable in October.

5. Same and different

Topic Same or different Them Us
Core belief Same Do the whole job, not a thin slice. AI plus a human check [site]. Agents do and check the work (brief).
Buyer type Same Small professional-services firms short of staff [site]. Target: SMEs and scale-ups; W5 uses accountancy firms as the example (brief). No customer yet.
Owner of the result Same They deliver the drafted return; the firm owns the final sign-off [site]. Aim: we run the work and own one number (brief). Not proven.
Kind of work Different One job: US 1040 personal tax prep [site]. W5: a layer above Xero for payroll, VAT and forecasts (brief).
Country and rules Different US only; IRS rules; US-based reviewers; US-only data [site]. UK mid-sized firms (brief).
Where the human sits Different Their own paid EAs and CPAs sit inside the product [site] [jobs]. The founders and the firm's own staff; Marton's logs show a cross-engine AI review gate, not licensed professionals.
How it learns Different Reviewers "help train our Agent" [site]. Learning from review is the stated aim [jobs]. Marton's logs show rules added from failures and feedback, on his own business only.
Proof Different Accuracy %, time saved %, turnaround days; five logos, three quotes [site]. No customer number yet.
Delivery Different A service with a web portal and file import [site]. The system plus the two founders.
Price Different Per-return credits; error credit-back [site]. Not set yet.

6. Wedge overlap

Wedge Fit Why
W5 One layer for one type of small firm direct Same buyer type (small accounting firms). They already sell one reusable layer (document to finished return) to many firms of one type [site]. They stop at US 1040 prep; our example is UK payroll, VAT and forecasts (brief). A YC partner who knows Magnetic would name it.
W7 Find and fix margin leaks for finance leaders adjacent Finance-side buyer, but a different problem: they add capacity, they do not find leaks [site].

7. Whitespace (where they stop)

  • Layer gap: no public memory of the firm across returns (UNVERIFIED that none exists), and learning is a job-ad intent [jobs]. To own it we need a brain that shows rework falling after rules are added, on one customer workflow.
  • Buyer gap: US only, and 1040 only; no business, payroll, VAT or advisory work [site]. To own it we need UK accountancy firms who say the pain is theirs, and one workflow run for 4+ weeks.
  • Number gap: their numbers are accuracy and time saved, one customer quote each; no public money result and no case study ("customer stories coming") [site]. To own it we need a before/after number a firm partner would sign.
  • Price gap: per-return credits with no public cash price [site]. A smaller firm may find it fine; we do not know. To own it we need to learn what a UK 10-person firm pays and tie our price to the result.
  • Speed gap: 2 to 4 days per return [site]. Not a gap for tax drafts. Skipped.

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. Magnetic is a funded, live US example of the same shape on one job; their own proof is accuracy and time saved, self-reported [site].
  • Draft answer to "how are you different from Magnetic?": DRAFT. "Magnetic does one job, US personal tax returns. We are building the layer for the other work in a UK accountancy firm, and our system already runs agents that do and check work every day on our own business." Founders say it their own way.
  • Future claim (needs proof): "We run payroll, VAT and forecast work for UK accountancy firms and own one number." Needs a UK accountancy firm, one workflow run for 4+ weeks, and a before/after number.
  • What we understand that they don't: a firm's work is many jobs that share the same data, so one brain across jobs beats one agent per job. needs proof (no customer yet; Marton's logs show the data-reuse idea only on his own business).
  • Copyability (M3): medium. Adding other US returns (business, partnership) is a feature for them, but UK rules, UK software and UK buyers would pull a 10-person US team off its tax-season core [site] [yc] (inference). Their reviewers are US licensed professionals, which does not carry over to the UK [site]. Not hard, so M3 can not score above 3 against Magnetic alone.
  • Why now (M4): None found in this teardown. Their stated reason (a talent shortage in accounting) is a trend with no dated change in the sources [yc].
  • Do not claim: "no one does one reusable layer for small firms" (Magnetic does, in the US). "Our brain learns from checked results" as a customer fact (no customer run yet). That we beat them on accuracy (no test; their numbers are self-reported and shift between pages).

9. Ideas worth taking

  1. Sell finished work, not software. Their copy says the problem with partial tools is the firm "still has to review everything" [site]. The same line fits W5 and W7.
  2. Pages that compare them to a named rival (Juno, Filed), with a row for who runs the AI and who reviews [site]. A cheap way to own a buyer's question.
  3. A written error guarantee with a clear limit (credit the fee if a mistake moves tax by more than $10, report within 60 days) [site]. It turns "can I trust it?" into a rule.

10. Unknowns

  • Cash price per credit. Look: their pricing page failed to load for us; ask a buyer who has seen a quote.
  • Real funding: $5.2M (their ads) against $500K (third-party list). Look: Orange Collective's announcement or an SEC Form D.
  • Real ARR and number of firms. Look: a Demo Day recap or a founder post.
  • When they moved from tool to service, and how many returns are fully automated today (they claim "over 25%" of the return process) [jobs].
  • Whether the AI re-check, the EA/CPA review and the 1-day reviewer turnaround ask in a reviewer job ad [jobs] mean heavy human cost per return. Look: reviewer job ads and the Head of Operations hire.
  • Any plan for the UK or other tax regimes. Look: careers and press in 2027.

Sources

Technical detail
Slug
magnetic
Lens hash
179096c479
Pass
deep
Stale
no