Executive Summary

Key Numbers

The first two cards show the scale and the results the strategy has produced so far. The last two show the gates those results have to clear before they turn into price.

Sources: TechCrunch (2026-08-12), The Finance Story

Thrive Holdings is a holding company spun out of Joshua Kushner's Thrive Capital. Instead of building AI to sell, it buys the companies that AI will be planted inside. Current, the accounting arm, gathers more than 50 firms and more than 2,000 professionals. Shield, the IT services arm, holds around 20 companies. Current started life in 2023 as Crete Professionals Alliance.

Part of the new capital goes toward a third vertical: regulatory and compliance services attached to physical assets such as buildings and infrastructure. The reason accounting and IT came first lines up with conditions the industry has flagged for years. Volume is high, rules are fixed, and the output leaves a paper trail. These businesses also sell human hours, so any drop in processing time moves the margin directly.

None of that is unusual for a private equity roll-up. What differs is the merchandise that travels with each deal. Buying an accounting firm brings more than a client roster and staff. It brings the returns, workpapers, review memos, and revision histories the firm produced over years. Buying an IT shop brings helpdesk tickets and support transcripts. None of it is listed as an asset in the purchase agreement, but the raw material for building an agent to do the work, and for checking whether the agent got it right, is on that side of the ledger.

That is also where this structure separates from a firm that licenses the same software from outside. A bought tool has no idea what this practice missed last year, for which client, on which line item. Bring 70 companies under one roof and those records pool in one place, and the pool becomes the training and validation material for the next round of automation. Call the purchase a set of companies or a set of work records; the price attaches to the second one.

OpenAI took a stake in Thrive Holdings in December 2025. Terms were not disclosed, and the arrangement reportedly grows that stake as the portfolio companies perform. In return, OpenAI places its own staff inside those companies to build products and install systems. Boris Power, who leads applied research at OpenAI, holds a role at Thrive Holdings alongside it. The tax filing agent was built jointly using Codex, OpenAI's coding agent.

The arithmetic differs from a software company charging per seat or per call. The seller only earns when the customer actually earns more. That makes real-world accuracy a revenue metric, and the only place to confirm it is in the customer's own work records. The returns 50 accounting firms produce every year double as evaluation data in this arrangement.

Deals of the same shape are stacking up. OpenAI is tied to The Deployment Company, backed by TPG, Brookfield, and Bain Capital, and Anthropic has taken a similar position through Ode. General Catalyst built an organization that buys professional services firms outright, including the accounting firm Accrual. Reading this as a shift from competing to sell models toward competing to buy the work itself is not a stretch.

The criticism arrived alongside it. When Thrive Capital sits on both sides and OpenAI staff work inside the companies, outsiders have a hard time telling whether results came from market demand or from internal support. The company's response is that the unmet demand was real and that several portfolio companies were already looking for AI tools. Which account holds up will be settled by numbers, and the first number in line is 98%.

Three results were announced: more than 7,000 tax returns processed, 98% accuracy, and preparation time cut by more than 30% at participating firms. On the IT side, helpdesk resolution got 36 times faster and the number of custom agents deployed doubled in the past month. Measured by throughput and speed, the automation worked.

Convert the remaining 2% into a count and it is 140 returns. Since 7,000 is a floor, the real figure is higher. But the conversion rests on an assumption nobody has confirmed, because the denominator behind 98% was never published. Counting whole returns that came out right is a completely different number from counting individual line items that came out right. If a single return carries hundreds of line items, 98% per line item means virtually every return contains an error.

The company's statement in the same announcement, that AI does not replace field work, local judgement, or professional sign-off, is worth reading against this. The accountant whose name goes on a return carries legal liability. Signing requires knowing what is right and what is wrong, and an accuracy rate of 98% does nothing on its own to lighten that check. If the 140 wrong returns are not flagged, the review population is still 7,000.

Verification labor did not disappear. It moved. The hours once spent producing a return are now spent inspecting one that has been produced. The 30% cut in preparation time may or may not account for that shift, and the announcement alone does not say.

What actually shortens review time is not another decimal of accuracy but knowing where the errors cluster. If the error rate can be broken out by filing type, client size, and line item, the range a human has to inspect narrows. That requires processing histories preserved alongside their ground truth. Which version was the final submission, where a human changed it, and why, all have to be in the record before an error map can be drawn. A firm with those histories and a firm without them are not companies you can buy at the same price.

The economics of a roll-up are simple. Buy small companies at a low multiple, bundle them, and sell the bundle at a higher one. Where those multiples land in accounting is broadly known.

Company size is not what separates the bands. Two firms earning the same profit can be priced differently depending on how fast they are growing and whether the bundle looks like it will run as one business afterward. The gap between the bottom and top bands is roughly double, and the spread a roll-up is chasing lives inside it.

Firmlever data. Source: The Finance Story

Add AI and expectations climb another step. In a 2026 survey of AI roll-up investors, more than 90% of the 102 respondents named a 2x improvement in EBITDA, meaning 100% or better, as the meaningful bar. Given that traditional levers such as back-office consolidation and centralized overhead typically deliver 10% to 15%, AI roll-ups are making a different kind of promise than the roll-ups that came before them.

Two recent cases show how that promise gets tested at exit. Xeinadin, which bundled 122 accounting practices across the UK and Ireland, has been marketed at £1 billion on revenue above £100 million and EBITDA around £60 million. Exponent, which came in during 2022, targeted 15x to 16x and has not found a buyer willing to pay it. On the other side is Citrin Cooperman. New Mountain Capital bought it in 2021 at roughly 11x and sold it to Blackstone in 2025 for $2 billion, around 15x.

The question that separated the two was integration. Is this genuinely one business, or a list of companies filed together? And the only way an outsider confirms a claim of integration is measurement. Before and after the acquisition have to be measurable on the same yardstick.

Using the same yardstick is heavier than it sounds. If 50 firms each ran their own chart of accounts and stored returns and workpapers in their own formats, simply putting pre-deal and post-deal costs on a common basis takes real work. Until those different systems are mapped onto standard fields, there is no basis for calculating how much the automation actually lifted the margin. If the multiple breaks, the likelier cause is not that AI failed to work but that the work could not be proven.

The playbooks AI roll-up operators consult list six things to look for when picking a target sector: fragmented ownership, labor-intensive operations, room for automation, recurring cash flow, aging owners, and customers who rarely leave. All six describe market structure. None of them asks what condition the target's work records are in.

In accounting and regulatory work, where being wrong is expensive, that gap matters. Written out as data diligence items, the questions look roughly like this.

• Are the last three years of output searchable in a consistent format, or scattered across folders by staff member?
• Can final submissions be told apart from drafts? The final version is what becomes the ground truth set.
• Does the revision history record who changed what, and why?
• Are the client-specific exceptions written down, or held in the memory of a long-tenured staffer?

All four ask about the possibility of verification rather than the performance of the automation. Without the first two, there is nothing to compare an agent's output against. Without the last two, there is no way to trace back why an error happened. In the room where target lists are reviewed and multiples are negotiated, these four answers usually go unchecked.

Go back to the 140 returns and the difference is clear. A firm with those histories can work out later which filing types and which line items the 140 clustered in, and the range a human inspects next year gets narrower. A firm without them is left with the number 98%, and the review scope resets to everything, every year. That is why the same agent installed at two firms cannot fetch the same price.

The same problem shows up wearing other faces. How training data provenance and ownership come back as a liability when acquiring an AI-native company is covered in The Hidden Data Debt Surfacing in AI Acquisition Diligence, and the data-side reading of why pilots stall on the way to production is in The Model Was Never the Problem.

Primary Sources

• 1.TechCrunch. (2026-08-12). "OpenAI-backed Thrive Holdings raises $2B to bring AI to the enterprise."
• 2.TechCrunch. (2025-12-01). "OpenAI's investment into Thrive Holdings is its latest circular deal."

Industry & Market Sources

• 3.The Finance Story. (2026). "Private Equity Bets on Accounting Roll-Ups Are Hitting a Reality Check."
• 4.Tenet. (2026). "AI Roll-up Investor Sentiment Report 2026." AI Roll-up Nexus.
• 5.Capital Founders OS. (2026). "Founder's Guide to AI-Enabled Roll-Ups."