Let’s talk about the fact that the company selling your accountant its AI now owns a piece of your accountant.
Picture the phone call. A partner in a forty-person practice, second generation, name still on the glass by the door. He’s fifty-eight. Half his senior staff retire inside five years and he cannot hire juniors fast enough to replace them, because nobody wants to do three years of trial balance work anymore.
The offer on the table is friendly. A bit over one times revenue. He keeps twenty percent, rolls it into the group. The name stays. Nobody gets fired. And there’s a team — actual engineers, not a support line — who will sit inside his firm and rebuild how the work gets done.
Most partners take that call.
I’m not here to tell you it’s a bad deal for him. It might be a great deal for him. I’m here to tell you what he’s actually selling, because it isn’t the practice. It’s the position.
Here’s what happened, cold. In December 2025 OpenAI took an ownership stake in Thrive Holdings — the vehicle Josh Kushner’s Thrive Capital set up to buy service businesses and run AI through them. OpenAI didn’t pay cash. It paid in access: its models, its product people, its engineers embedded inside the acquired firms, in exchange for equity, and for access to those companies’ data to fine-tune its own models. That’s not my reading, that’s how it was reported at the time, and the Financial Times filed it under one word: circular.
In August this year that vehicle raised two billion dollars at a twelve billion valuation, with SoftBank, D1 and Altimeter in. Over seventy businesses on the platform. The accounting arm, Current, is more than fifty firms and more than two thousand professionals. The IT arm, Shield, is around twenty companies. And the third platform they’re standing up is regulatory services for physical assets — permits, inspections, certification.
The performance numbers are theirs, so treat them as claims: seven thousand tax returns through their tax agents at ninety-eight percent accuracy, prep time down more than thirty percent, help desk resolution thirty-six times faster. Maybe. What I’d want to know is who checked the ninety-eight and against what. But the claims aren’t the story either.
The story is that this is not one deal. It’s a category. Both big labs are building services arms now — the joint ventures with the private equity giants, the embedded engineering teams, the deployment companies. The model layer is commoditising, so the labs are moving down into the work.
Now the part that matters, because there are three things being bought here and only one of them is on the term sheet.
The first is the margin. An accounting firm is a payroll business wearing a professional licence. Sixty, seventy percent of the cost is people. If you can move a meaningful slice of that cost into tokens, the margin doesn’t improve by ten percent, it doubles. And here’s the mechanism nobody says out loud: the same saving gets counted twice. The lab bills for the tokens. The holding company books the margin. Then the whole thing gets valued on the margin that hasn’t arrived yet. The vendor is paid on the meter and paid again on the equity, on the same removed salary. I have never seen a supplier with that structure.
Second thing being bought: the corrections. Two thousand professionals fixing model output on real client files, all day, every day, is the most expensive training data on earth. Nobody can scrape it. You can’t buy it from a web crawl. It only exists inside a firm where a qualified human has to stake their name on whether the number is right. Kushner said the quiet part in the announcement — the businesses they acquire “represent the right reward systems.” Read that sentence twice. He’s not talking about client service. He’s talking about a reward signal.
So your bookkeeper corrects the machine, the correction trains the model, the model gets sold back to the next firm, and your file was the classroom.
Third thing being bought, and this is the one that should worry you: independence.
The accounting profession has an entire ethical architecture built on one idea — that the person signing off has no stake in the thing being signed off. That’s why the code has independence rules, why audit partners rotate, why you can’t own shares in the client you audit. The AICPA’s own guidance this year still says it plainly: the responsibility for the work product stays with the CPA, no matter who built the tool. It also warns that inputs typed into a model can end up training it, which is a confidentiality problem.
Here’s the gap. That code was written when your software vendor was a supplier with a licence agreement. It has nothing to say about a world where your vendor’s parent holds equity in your employer’s platform, and your employer’s margin depends on how much of your work the vendor’s model absorbs. Nothing in that sentence is illegal. That’s the point. It didn’t need to be banned because in 1990 it was science fiction.
And the client — the business owner whose books these are — is never told. When a firm sends work to an offshore centre, the client knows. There’s an engagement letter, a data residency question, a footnote. When the firm’s owner is the model company, the flow of your data is an ownership fact, not a service term. You will not find it in your fee schedule.
That’s what people are missing when they file this under “private equity does accounting now.” Private equity has bought accounting practices for forty years, and the playbook was always the same: centralise the back office, cut duplicate costs, exit on a better multiple. This is different in kind. PE optimised the firm. This owns the firm, and the tool, and the training loop, and the valuation that depends on all three.
Now, the operator lesson, because I’m not interested in a morality play. Incentives made this inevitable. If a lab can capture the services margin on top of the token revenue, it will, and the firms that say no will be the ones losing partners to the ones that said yes.
So three hard things.
If you sell services — accounting, legal, consulting, agency, whatever — understand that your labour cost is now somebody else’s valuation input. The buyer isn’t pricing your client list. It’s pricing the gap between what your people cost and what a model costs. Which means the only durable defence is the thing a model can’t hold: the relationship, the judgement, and the contract. Get your client agreements in your own name. Own the data rights. Price the outcome, not the hour, because an hourly rate is an invitation to be arbitraged.
If you’re the one being bought, the multiple is the least interesting number in the room. Read the data clause. Who owns the fine-tune? Can an affiliate of the buyer train on your clients’ files? What happens to your clients if you leave in year three? Get that in writing before you get emotional about the headline number.
And if you’re the client — ask who owns your firm. Actually ask. Then ask whether your files train a model, and whether you can leave without your history staying behind. If the answer is vague, that’s your answer. Put it in the engagement letter or find someone who’ll sign it.
Because the thing that makes a professional worth paying isn’t that they’re clever. It’s that they’re on your side and structurally can’t be on anyone else’s. That’s not a vibe. It’s a structure.
The machine didn’t take the firm. It bought it, quietly, with equity instead of cash, and kept the humans on as the part that signs.
You’ve got this. Ask the question before you sign.