Who Owns the Delivery? What AI Roll-Ups Prove and What They Do Not

A Call Worth Taking Seriously

Picture a managing partner at a regional accounting and advisory firm. The scene is a composite, offered for illustration. A broker rings: a venture-backed platform wants to buy the practice, keep the partners and "put AI through the delivery". The price is generous against the multiples the partner has seen. The partner's real question is not whether to sell. It is what the buyer knows that the partner does not.

The public record answers only half of it. In August 2026 Thrive Holdings raised $2 billion at a $12 billion valuation from investors including SoftBank, D1 Capital Partners and Altimeter Capital, and OpenAI has held an ownership stake since December 2025, sending employees to work inside Thrive's companies¹. Thrive's platforms are reported to span more than 70 businesses, with more than 50 firms at Current, the accounting arm, and around 20 companies at Shield, the IT arm¹. Those facts are documented.

The performance claims belong to another category. Thrive says its tax agents processed more than 7,000 returns at 98% accuracy and cut preparation time at participating firms by over 30%; TechCrunch reports, without attribution in the sentence, that Shield's AI products sped up help-desk resolution by 36x¹. One independent compilation finds that every operating number comes from the company or its shareholder, that accuracy appears elsewhere as "up to 97%" from OpenAI and "up to 98%" from Current, and that no baseline has been published for the 36x³. Revenue figures, such as Shield passing $100 million in 2025, are likewise self-reported and unaudited³.

This paper takes a position on what to do with that gap. The form is real and well funded. The margin re-rating that would justify the valuation is not yet evidence. The useful move for a leader is to stop asking whether the model works and start asking which of three different things is doing the work.

Three Explanations for One Valuation

Whatever an AI-native roll-up eventually reports, the result will be produced by some mixture of three drivers. The first is AI-driven change in delivery: fewer hours per unit of work at held quality. The second is ordinary roll-up synergy: procurement, pricing discipline, a centralised back office, shared overhead, the tools private equity has applied to fragmented services for years. The third is access to capital: the ability to outbid rivals for targets and absorb integration costs that a smaller competitor cannot finance.

The public record cannot yet separate them. That matters because each driver implies a different response from the firm across the table.

If AI is the driver, you face a delivery model you must build toward. If synergy is the driver, you face a consolidation, and the choice is whether to join it or price against it. If capital is the driver, you face a financing contest, and the answer lives on your balance sheet and in your alliances. A single label, "AI-native", hides that these are three different strategic problems.

What the Arithmetic Allows

A worked figure shows why the attribution question is not academic. Take a composite services firm with revenue of 100 and EBITDA of 10, inside the 5 to 10 per cent range one investor commentary describes as typical for traditional services⁵. Costs are 90, of which labour is 60 per cent, or 54, within the 55 to 65 per cent labour share that the same commentary assumes without citing a source⁵.

Now automate 40 per cent of that labour, the kind of premise that commentary uses. If every unit of saving were kept, the saving is about 21.6 and EBITDA rises to roughly 31.6, which is inside the 30 to 40 per cent software-like range the thesis targets⁵. That is the investor case.

Three leaks sit between that figure and a reported margin. Clients may take back part of the saving in price; if they take half, margin lands near 21. Tool, model and platform costs reduce it further. And quality control consumes people: if acquired firms cut staff, TechCrunch notes, there are fewer people to catch AI errors, and if they keep staff, the margin gains investors are counting on may not be realised⁷. Meanwhile, three points of overhead and procurement savings, entirely ordinary, would lift the same firm from 10 to 13 without any AI at all. A reported margin of 20 could therefore be mostly AI, mostly arithmetic, or mostly price discipline. The number alone does not say.

The Strongest Case Against This Caution

The best argument on the other side deserves its full weight. Owning delivery, the argument runs, captures value that selling software to the same firms never can. The investor Elad Gil made the point that an owner of the asset can transform it much more rapidly than a vendor selling software to it⁷. A model vendor taking equity is a vendor betting on outcomes: according to Reuters as relayed by Caritas, OpenAI provides research resources in return for an ownership interest, and Thrive owns the intellectual property and products created³. Gartner, for its part, says up to $234 billion of enterprise application spending is exposed to agentic arbitrage between now and 2030, roughly 20 per cent of enterprise SaaS spending by then⁴. And the reported count of Current's firms moved from "almost 30" at its June 2026 rebrand, as Caritas records it, to "more than 50" by August³¹.

This case is right about direction and wrong about inference. Gartner's release discloses no methodology and defines no "at risk" beyond "exposed", describes the shift as "less an apocalypse and more of a metamorphosis", and concerns software spend rather than services margins⁴. The firm counts are company-sourced, and the basis of counting is not explained³. Investor projections are what they are: General Catalyst's thesis of re-rating margins from 5 to 10 per cent toward 30 to 40 per cent is reported by a commentator whose own figures are, in its words, reported by the firms themselves and not independently audited, and which notes that none of the portfolio has operated through a recession⁵.

Yet the argument does sharpen the conclusion. Even if the margins turn out to be roll-up arithmetic, the competitive threat does not disappear. A vendor-owned, venture-funded buyer can still take your targets, your talent and your clients. The caution is about what to believe, not about whether to prepare.

Where the Model Is Most Likely to Break

Integration is the first break point. A sceptical investor analysis argues that integrating AI into acquired business N+1 is not much easier than doing so for business N, because data models, technology stacks, processes and cultures are likely to be completely different⁶. The same analysis observes that cost reduction means continuous personnel cuts, and that how this affects the ability to attract and retain the necessary talent remains to be seen⁶. Note the interest: its author's firm invests behind building services businesses from scratch, which is the alternative to acquiring them⁶.

Quality is the second. TechCrunch cites a Stanford Social Media Lab and BetterUp Labs survey in which 40 per cent of 1,150 employees reported shouldering extra work from "workslop", AI-generated output that looks polished but lacks substance⁷. That research concerns AI use generally, not roll-ups. It matters here because it names the mechanism by which a labour saving can quietly convert into a review burden.

Regulation is the third, and it is live. The AICPA proposed revisions to its independence rules for CPA firms with outside investors last December, and a task force has made whether investors hold significant influence or control a critical factor in whether an attest entity is independent². According to a former PCAOB board member quoted by Accounting Today, none of the 81 comment letters came from technology companies, ambiguity remains about when licensing AI audit tools creates a conflict, and the rule could favour large incumbents that build AI in-house². The committee indicated it plans to re-expose the guidance in October or November, with adoption next May, although the article's own timing language is not fully consistent². Audit-firm ownership is therefore neither permitted nor settled. How far the model scales beyond attest work is not shown in the evidence available.

A Decision Test

For an incumbent, the choice between rebuilding, partnering and selling should follow from three questions, asked in this order.

1. Could we reproduce the target's reported gain without AI? Run your own delivery economics against a conventional roll-up benchmark: pricing, procurement, overhead. If the gain appears without AI, the label is adding a valuation story, not a cost base.
2. Does the gain repeat in the second acquisition? Ask for the margin at business N and at business N+1, by name, before and after. A pilot result at the first firm is evidence of a pilot.
3. Who holds the data, the IP and the exit? Where a model vendor is also an owner, the terms on data access inside delivery are undisclosed in the sources opened. Buyers of these services, CFOs and CIOs among them, should contract for data ownership and exit rights and ask for disclosure of the vendor's shareholding.

The answers map to action. Where the labour share is high and clients will accept AI-assisted delivery at a lower price, rebuild. Where capital is the real advantage, partner before consolidation narrows your options, or arrange financing of your own. Where the offer values synergies you cannot capture alone, selling can be rational, so long as you know what you are being paid for.

The useful question is a subtraction: remove the AI from the pitch and see what margin remains.

What Would Change Our Mind

The main uncertainty is stark. No audited or independently verified revenue, margin, organic-growth or client-retention data exists, in the public record we could inspect, for any Thrive platform, and it is unknown whether any gains come from AI or from standard roll-up synergies. The key challenge follows from it: a matched comparison of an AI-integrated acquired firm against a conventionally rolled-up one, with audited margins, would settle the argument.

The thesis would weaken if filed or audited financials showed margins in line with conventional roll-ups, if acquired firms lost clients or staff at a high rate, if regulation limited investor control of attest entities, or if model-vendor terms restricted portability. It does not apply to attest work where independence rules bind, and the evidence here is centred on the United States. Several analysts cited also sell adjacent services or invest in alternatives, so their scepticism and enthusiasm both carry an interest³⁵⁶.

Hold the subtraction in mind when the next call comes. If most of the margin survives the removal of the AI, you are competing with private equity that has a better story. If little does, the evidence has not yet arrived, and your own books may produce it before the buyer's do.