> ## Content Index > Fetch the complete content index at: https://varops.com/llms.txt > Use this file to discover other available public pages before exploring further. # A YC startup rebuilt an accounting firm around AI, and operating margin jumped from 5% to 70%. Here's how they keep the upside. - URL: https://varops.com/a-yc-startup-rebuilt-an-accounting-firm-around-ai-and-operating-margin-jumped-from-5-to-70-heres-how-they-keep-the-upside/ - Published: 2026-09-14T12:30:24.000Z - Updated: 2026-09-14T12:30:24.000Z - Description: In a randomized experiment, accountants working with AI classified transactions 17.5 points more accurately - not just faster. How far a services firm can actually go on the published evidence, and the one decision that says who keeps the gain. - Author: Ran Aroussi - Tags: Founder Mode *An accountant working with AI does not just go faster - in a randomized experiment the work came out measurably more accurate.* [*Ran*](https://varops.com/columnist/aroussi/) *on how far a services firm can go right now, what the published numbers actually reach, and the one decision that determines whether the gain stays in the business or walks out with a vendor. The four conditions at the end are ours, and he runs them against us too. -* [*Muximus*](https://varops.com/columnist/muximus/) --- A YC startup called Minerva rebuilt an accounting firm around AI this week and put a margin number on it that got a lot of attention. I will come back to that number. The one I keep coming back to is somebody else's, and it is better. In a randomized experiment inside accounting work, published this June in the [*Journal of Accounting Research*](https://onlinelibrary.wiley.com/doi/10.1111/1475-679x.70052?ref=varops.com), accountants given AI assistance classified transactions **17.5 percentage points more accurately** than those working without it - against a baseline of 74.7%. Not faster. More right. That is the part of this that gets undersold. The pitch is usually about cost, and cost is the least interesting thing on the table. A person who was good at their job becomes measurably better at it. That is a different proposition from replacing them, and it is the one with evidence behind it. ## How far this actually goes The same study - Choi and Xie, 277 surveyed accountants, roughly 200,000 transactions across 79 client companies - found that a one-standard-deviation increase in AI use goes with about **19% more clients supported per week**. Books closed 7.5 days sooner. Time spent on data entry dropped by about 3.5 hours in a 40-hour week. Move from the lowest observed AI usage to the highest and the client figure is around 80%, though the authors call that "an extreme comparison" and steer you to the 19%. Fair enough. Nineteen percent more clients, with the books closing a week earlier and the work coming out more accurate, is already a very good year. Their conclusion is the line I would put on the wall: "AI is most effective as a tool that augments - rather than replaces - professional judgment." The accountant does not disappear. The accountant gets a bigger practice and does more of the work that needed a human in the first place. And it scales past small firms. Klarna's [prospectus for its 2025 listing](https://www.sec.gov/Archives/edgar/data/2003292/000200329225000052/klarnagroupplc424b4.htm?ref=varops.com) reports revenue per employee going from about $344,000 in 2022 to about $972,000 in the twelve months to June 2025, with headcount down from 5,527 to 3,098 - attributed in the filing to a decision to "reduce our overall headcount and drive operational efficiency by leveraging AI in our business." Operating margin improved 44 percentage points. Lemonade told shareholders it had [more than doubled in-force premium](https://www.lemonade.com/investor-relations-bo/wp-content/uploads/docs/LMND-Shareholder-Letter-Q1-2026.pdf?ref=varops.com) while shrinking its team 6%. Both of those companies were still losing money, and both metrics are partly self-defined. I am not holding them up as profit stories. I am holding them up as proof that the volume a team can carry has genuinely come unstuck from the size of the team, in businesses that had to file the numbers. ## What it takes to actually get there Here is the honest part, and it is the difference between the firms this works for and the ones it does not. Humlum and Vestergaard studied [25,000 workers across 7,000 Danish workplaces](https://www.nber.org/system/files/working%5Fpapers/w33777/w33777.pdf?ref=varops.com) in eleven AI-exposed occupations, accountants among them. They found precise nulls on earnings and hours - ruling out effects larger than 2%. The reason is not that the tools failed. Eighty-five percent of chatbot users reported reallocating their time savings to other job tasks. The hours came back. They just got quietly absorbed into more of the same work instead of being pointed at anything. So the ceiling is high and the default is nothing. What separates them is whether somebody decided in advance where the recovered capacity was going - more clients, or better work for existing ones, or advisory services the firm could never staff before. Capacity with no destination becomes busywork. That decision is a management one, not a software one, and it is where the gain is won or lost. ## What set this off A startup called Minerva said this week that it bought an accounting firm and took it from 5% to 70% operating margin. I could not verify that figure - Mars is a wholly-owned private subsidiary, so the number appears in Minerva's own launch post and case study and nowhere a third party signed off. That is not an accusation; it is just how private ownership works, and it would be equally true of anything I claimed about my own company. Worth knowing for calibration: a firm reporting 5% has expensed its partners' compensation and a firm reporting 70% probably has not. The AICPA's 2025 MAP survey puts net remaining at 36.9% of net client fees across all respondents, 38.4% for top performers, and that is *before* partner compensation comes out - figures I have via [Auxo Capital Advisors](https://auxocapitaladvisors.com/accounting-firm-profit-margins/?ref=varops.com), an M&A advisory that sells accounting firms, since the AICPA's release is members-only. Their note is that comparing firms this way needs a "market-pay adjustment for partner labor." Margin at a private firm is a definitional choice before it is a result. Also worth knowing: Mars is the rebranded myPocketCFO, a venture-backed bookkeeping startup [Forbes covered in 2024](https://www.forbes.com/sites/elainepofeldt/2024/07/30/a-startup-wages-a-war-on-messy-bookkeeping-its-primary-weapon-an-easier-way-for-product-based-startups-to-record-inventory/?ref=varops.com). It already had software and structured data - an easier starting line than most firms have. None of which is a knock on them. Minerva's own post says the thing I most agree with: "Humans are not going anywhere," and staff shift to higher-value advisory work while agents take the routine. That is exactly right, and it is what the research shows too. I just cannot plan against their number, and neither can you. ## The decision that decides who keeps it Minerva is acquiring firms and also, in their words, "working with firm owners to make their operations AI native." Selling is a perfectly good outcome, and for an owner near retirement it may be the best one available. But if you are not selling, one question decides whether any of this compounds: [who owns the thing that did the work](https://varops.com/karp-asked-the-right-question-the-answer-is-an-architecture-not-a-vendor/). The model is the commodity part. Everyone gets roughly the same one. What produces the gain is the system built around it - the data, the workflows, the accumulated understanding of fifty specific clients and how their money actually moves. That understanding is the most valuable thing a services firm has ever had, and for the first time it can be captured in something other than a senior person's head. If it lives somewhere else, the gain shows up in your P&L this year and belongs to someone else next year. We wrote a while back about [not outsourcing the thing that made you different](https://varops.com/dont-outsource-the-thing-that-made-you-different-and-call-it-ai-adoption/) and calling it AI adoption. This is that, with money attached. > Disclosure: VarOps is published by me, and [Resident AI](https://varops.com/resident-ai/) is what we build. That turns the skepticism up, not down. ## The four conditions An AI is resident in your company when all four are true. Fail one and it is bolted on. **It runs where the work already happens.** No second window, no new login, no habit anyone has to be trained into. Adoption dies quietly a few weeks after the pilot team stops trying. **It sees the whole operation.** Not the slice somebody thought to paste in. An AI with no data asks a question, which is safe. Give it half the picture and it answers confidently instead, which is worse: people trust it once, get burned, and then stop trusting the system at all. **It runs on your own infrastructure.** Your hardware, your network, your security perimeter. No vendor-hosted service in the middle holding a copy of your operating record. **What it learns belongs to you.** In a store you own and can export in a format you choose. Otherwise you are renting access to a picture of your own business. Now the part that makes this a test rather than a pitch. These conditions rule out work we do ourselves: a deployment that ends up sitting beside the business has failed the first one, and the fact that we built it does not change the result. A definition that only ever disqualifies other people is a marketing claim wearing a lab coat. They rule out plenty of good software too. A chat product with an excellent interface fails the first. A platform that ingests your systems into its own environment fails the third. A well-integrated agent whose context cannot leave fails the fourth. Those are fine products - they are just not the thing that keeps the gain inside your company. And there is a case where owning it is the wrong call. If a firm is being sold in eighteen months, building the system is an expensive route to an asset the buyer will replace. Sell, or partner, and take the cash. Four out of four is resident. Three out of four is bolted on with better marketing. ## Write down three numbers first Not margin. Margin is the figure a definitional choice can move on its own, which makes it useless as a starting line. Clients per person. Days to close the books - Choi and Xie found adoption associated with a 7.5-day reduction, so there is a published delta to aim at. And the share of hours going to data entry rather than advisory work. All three are operational, hard to fudge, and a stranger could check them. Then decide, before the capacity arrives, where it is going. That is the step the Danish workers' employers skipped. ## What came of it The ceiling is real and it is documented where it counts. Seventeen and a half points of accuracy in a randomized experiment. Nineteen percent more clients per accountant. A 44-point margin swing in a securities filing. A doubled book with a smaller team. None of that is a projection. What none of it is, is automatic. The firms that got there decided where the recovered hours were going and built the system themselves. The gain is available to any firm willing to do both. So we would start with the unglamorous part. Write down clients per person and days to close, today, before anything changes. Then settle who is going to own the system that moves them, because that is the difference between a good year and a business worth more than it was.