Combinely (YC Spring 2025): teardown vs XMarts

In one line: Combinely sells "Coco", an AI coworker with 50+ ready-made agents that do bookkeeping, VAT, payroll and tax prep inside UK accountancy firms' own tools, for £180 per user per month [site] [yc]. Threat: 3 because it is almost exactly our "one layer for one type of firm" idea (W5, accountancy), live in the UK with named customers, a public price and a funded seed [site] [press].

1. Fact file

Item	Finding	Source
What they say	"Accountants, meet your new coworker." YC line: "AI Coworker for Accountants".	[site] [yc]
Founded, team, funding	YC Spring 2025 (their own post says X25), London, 8 people. Seed in June 2025; Pioneer Fund named as new investor. Another listing names Flex Capital, Lombardstreet Ventures, Pioneer Fund and YC (UNVERIFIED: a check found Pioneer Fund on Preqin but no link to Flex Capital or Lombardstreet). Amount not disclosed. Dealroom shows $125k and $500k, but that page looks auto-generated, so UNVERIFIED.	[yc] [press] [blog]
Founders	Tom Invernizzi (CEO): ex-Deloitte accountant and auditor, then investment banking analyst, LBS Master's 2021. Arthur Granacher (CTO): wrote Shazam's iMessage app, then four years at Google DeepMind automating doctor workflows. What YC likely saw: a founder who lived the pain plus a CTO who has automated a regulated profession's workflow (inference).	[yc] [press]
Customers	They claim 1,000+ accountants and 70+ firms. Named: Blick Rothenberg, Moore, McCabe Ford Williams, Blu Sky, Oury Clark, Incra (Australia), LedgerPlus. Burgess Hodgson is UNVERIFIED (not on the customers page or blog).	[site] [blog]
Price	Public: £180 per user per month if paid annually upfront, £200 annual paid monthly, £220 monthly. All 50+ agents, unlimited use (fair use), up to 10 mailboxes per licence. Enterprise is custom.	[site]
Path	LBS Incubator 2024, then YC. Launch posts moved from "AI coworker for tax teams" (90 minutes a day saved, they claim) to "AI Coworker for Accountants" with a "doubles firm capacity" claim. Product named Coco. No rename found. A US site is UNVERIFIED (a July 2026 Blick Rothenberg partnership post only mentions US expansion). SOC 2 Type II announced 12 Dec 2025.	[press] [yc] [blog]

2. Their proposition, reverse engineered

Part	What they do
Who buys	Partners and owners of UK accountancy practices, mostly mid-sized and "tech-forward" firms [press] [site]. Big names (Blick Rothenberg, Moore) appear as logos or partnerships [site].
The pain	Manual work: data entry, rewriting the same client emails, chasing documents. The founder says 95% of his work was heavily manual; the LBS article puts this in his investment banking job, not accountancy [press]. Partner review is the bottleneck on tax returns [blog]. "The question was never willingness, it was capacity" (McCabe partner) [site].
The promise	"Doubles firm capacity, not individual productivity" [yc]. Work is prepared in the firm's own tools and a person approves it before it is final [site].
How it works	Pick an agent from a library (bookkeeping, VAT, payroll, personal tax, corporation tax, accounts, practice admin), connect Xero, QuickBooks, BrightPay, Outlook and others, review the output [site]. Describe a task not in the library and Coco turns it into an agent [site]. Output is source-backed with a record of what the agent did [site].
The output	A finished first draft for a person to sign off: a payroll run, a VAT working, a tax return review, a workpaper [site].
Proof style	Hours saved and usage counts, set by the customer: 20 hours a month on payroll, 78 agents in 8 weeks, 2,000+ tasks in two months, 39% less manual processing [site]. All self-reported. No fee or profit number.
Who sells it	Mix. Self-serve sign-up and demo booking on the site [site]. Onsite staff for the big pilot: Combinely people sat with each user for 8 weeks [site]. A conference stand at the Digital Accountancy Show, 7-8 Oct, is UNVERIFIED (not found on the home, customers or blog pages, and a search found nothing). A partner page invites others to build and resell agents or refer firms for pay [site].
Price model	Per seat, flat, unlimited use. Seat price does not move with the result [site].
The story	AI works "with accountants, not around them": it does prep, the firm keeps judgement [site]. Their enemy is the manual grind and the junior-skills gap. A February 2026 post argues AI should coach juniors, not cut headcount [blog].

3. Positioning

Category they claim: "AI coworker for accountants" [site].
Category a YC partner would put them in: vertical AI for accounting, the "AI-native accountancy tooling" slot next to Basis and Rillet, but UK and sold to firms rather than to their clients (inference; Basis and Rillet named in press search results, not compared in depth).
What they want to be known for: the ready-made coworker that is live in a day and that firms can extend themselves.
Where they sit on the map: L1/L2/L3 = 2/2/2, buyer = mid-market UK accountancy firms, business.

4. Three layers vs ours

Layer	Them (score + evidence)	Us (Marton's logs show)
L1 Company brain	2. YC says it "learns your clients and your firm's standards" [yc]. A per-licence limit of 10 mailboxes gives it firm-wide context [site]. A third-party page says it builds a client profile from past emails and work; that source is a rival's blog, so treat it as weak [press]. Their own site does not explain how the memory works [site].	A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning. One memory shared by all work.
L2 Agent orchestration	2. 50+ agents, custom agents, a firm-wide library and "one policy" [site]. Every output goes to a human for approval [site]. They show many agents per firm (78 at McCabe, 40+ at LedgerPlus) but no claim that agents hand work to each other [site].	About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests.
L3 Execution and learning	2. Agents do the work end to end (a payroll that "ran with no errors") and a person checks it [site]. They count hours saved and usage [site]. No public sign that a rejected output changes the next run: UNVERIFIED either way.	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.

Note: L3 2 is a borderline call. If a reader holds that "measure" must mean money or error rate, this is a 1 and loop_layers is 2. Their strongest L3 sign is a blog line about errors repeating on "the last fifteen returns" and Coco explaining the rule to juniors [blog]. That is coaching, not a learning loop.

5. Same and different

Topic	Same or different	Them	Us
Belief	Same	An AI that joins the firm and does real work, human signs off [site].	Same (brief).
Buyer	Same, with a gap	Mid-sized UK accountancy firms [site] [press].	Target SMEs and scale-ups; accountancy firms are example H2 in the brief. No customer yet.
Owner of the result	Same	The firm keeps judgement and sign-off [site].	Same.
How it learns	Unclear	"Learns your clients and your firm's standards" [yc]; mechanism not published [site].	Marton's logs show rules added from failures and feedback.
Kind of work	Different breadth	One profession: tax, VAT, payroll, accounts [site].	Any one workflow across a business (brief).
Delivery	Different	Software plus onsite staff in the large pilot [site].	The system plus the two founders.
Proof	Different	Named UK firms, a case study each, hours saved [site].	No outside customer yet (brief).
Price	Different	£180 to £220 per seat per month, public [site].	Not set. Brief aims at pricing on the result.
Speed	Same or faster	"Minutes" to set up, one week to a running payroll at LedgerPlus (they claim) [site].	Target days. Unproven.
Starting point	Different	Pre-built library (50+) first [site].	The company's own memory first.

6. Wedge overlap

Wedge	Fit	Why
W5 One layer for one type of small firm	direct	The brief's own example is a layer above Xero for payroll, VAT returns and forecasts at accountancy firms. Combinely sells exactly that, with Xero, BrightPay and HMRC links, in the UK [site].
W3 Company memory from every call	adjacent	They use mailboxes and past work for context [site]. Calls and meetings are not mentioned on their site (UNVERIFIED that they lack them).
W7 Margin leaks	adjacent	"Fee vs scope review" is in their practice-admin agents [site]. It is a feature for the firm, not a product for a finance chief.
W8 The other 90% of an AI skill	adjacent	Custom agents, a shared library and "one policy" for growing agent sets [site]. No public detail on upkeep from real use. A partner page invites others to build and resell agents [site].

7. Whitespace (where they stop)

Layer gap: none we can prove. They cover all three layers at score 2. We can not claim they stop at one layer. Needed: evidence of how, or whether, rejected work changes the next run.
Buyer gap: one profession only (accountancy), UK and Australia (Incra); a US move is only hinted (UNVERIFIED) [site] [blog]. Every other kind of firm is open. Needed: a named buyer in another profession with the pain on file.
Number gap: their numbers are hours saved and usage, self-reported. No fee earned, margin or error-rate result [site]. Needed: one before/after money or error number a partner would sign.
Price gap: flat per seat, so a 5-person firm pays about £10,800 a year at the annual rate (£180 x 5 x 12; sum is ours) [site]. Pricing on the result is untouched. Needed: a buyer who prefers paying for outcomes.
Speed gap: not a gap. They claim minutes to set up and a week to value [site]. Their large pilot needed 8 weeks of onsite help, so "nothing to set up" holds for small firms only (inference) [site].

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. It also shows a feedback loop on one workflow: offer feedback lines fell from 44 to 18 as 22 rules were added. Nobody outside Marton has used it.
Draft answer to "how are you different from Combinely?": DRAFT. "Combinely sells accountants a seat with a library of agents. We are not an accountancy tool. Our system already does and checks work on our own business, and we want to run one workflow for a firm and own one number." Founders say it their own way.
Future claim (needs proof): "Our agents improve from every checked and rejected piece of work, and we price on the result." Needs one customer workflow run for 4+ weeks with a before/after number, plus a log that ties rule changes to falling rework.
What we understand that they don't: needs proof. Candidate: that price and proof should be tied to one number, not a seat. They sell seats and count hours (self-reported) [site].
Copyability (M3): easy. A learning-from-rejections loop is a feature that fits their buyer, and they already hold the review step where rejections happen [site] (inference). They have a funded seed, a build team of 8 and live customers to learn from [yc] [press]. Results-based pricing would cut against their seat model, but they could add it for the enterprise tier (inference). Not hard, so M3 can not score above 3 against Combinely.
Why now (M4): None found in this teardown. Do not use "agents can now do the work" or "UK firms are short of staff": neither is a dated, sourced change here.
Do not claim: "accountants have no AI coworker" (Combinely is live with named firms). "They stop at one layer." "They do not learn from corrections" (UNVERIFIED). Any accountancy wedge as unique. That Combinely's customer numbers are proven (they are self-reported).

9. Ideas worth taking

Sit with the user for the first weeks and let the people doing the work name the first task. McCabe asked each manager for their most time-consuming workflow, and the pilot grew to 78 agents [site].
One page per customer with 3 numbers (hours, tasks run, share of clients on the agent). Easy for a buyer to repeat to a partner [site].
Show how a failure beat a human-only attempt. LedgerPlus left DIY Claude because it had no system access, no shared context and no schedule [site]. The "why not just use Claude?" answer is worth writing for our own pitch.

10. Unknowns

Real revenue, customer count by paid seat, and funding amount (only a June 2025 seed is shown; no figure). Look for a press release or filings.
How "learns your firm's standards" works and whether a rejected output changes the next run. Ask a customer, or book their demo.
Whether the 1,000+ accountants and 70+ firms are paid seats or sign-ups (they claim both; no breakdown).
Whether they plan to move beyond accountancy. The partner page invites third parties to build agents on their base, which could widen them (inference) [site].
Their sales team size. YC and aggregator job pages show no open roles [jobs].
The AccountingWEB article on Combinely returned an access error. It may hold investor and customer detail.

Sources

combinely.ai home [site]
Pricing [site]
Customers [site]
McCabe Ford Williams story [site]
LedgerPlus story [site]
About [site]
Security [site]
Partners page, from site map description [site]
Blog index [blog]
AI junior to manager, 24 Feb 2026 [blog]
Economics of AI tax review, 23 Jan 2026 [blog]
YC company page [yc]
YC jobs page [jobs]
London Business School incubator article, 28 May 2025 [press]
Aggregator listing investors (Zero G Talent) [jobs] (secondary)
Preqin profile, seed June 2025, seen in search result only [press] (secondary)
eesel AI review of Combinely [press] (written by a rival, one-sided; older view of pricing as hidden)

Combinely

What they do

How close to us

Wedges

Who buys

Full teardown