Three Investment Models for AI-Enabled Professional Services

A $5 billion U.S. accounting transaction and a €36 million financing for a German tax startup do not look like the same story. They may be two different chapters of it.On July 29, Grant Thornton Advisors agreed to acquire CBIZ for $5 billion. New Mountain Capital, which backed Grant Thornton in 2024, is investing additional equity to support the transaction. Grant Thornton explicitly linked the proposed deal to its existing $1 billion investment in AI and advanced tech, arguing that greater scale should allow it to deploy AI-enabled capabilities across more clients, workflows and professionals.Six weeks later, Berlin-based LimeTax announced €36 million of financing—€6 million of equity led by Motive Partners and a €30 million acquisition facility—to acquire tax and accounting practices and operate them on a proprietary agentic platform. LimeTax describes the ambition explicitly: bring established firms into one group and rebuild how their work is produced.One transaction is a multibillion-dollar combination of scaled incumbents. The other is an early-stage technology-enabled buy-and-build.But they raise the same question for private equity:If AI changes the production economics of professional services, what exactly should investors own—the software, the service provider, or both?That question is becoming more important, because “AI in professional services” is no longer one strategy. At least three distinct playbooks are emerging.Archetype 1: Transform the incumbentGrant Thornton/CBIZThe first model starts with an existing professional-services platform and uses AI to make a scaled incumbent materially better. Both Grant Thornton and CBIZ is the clearest current example.Grant Thornton is acquiring CBIZ because of its scale, client relationships, capabilities, market position and strategic fit. But AI has become an explicit part of the rationale for that scale - Grant Thornton committed $1 billion over three years to tech and AI. The program is broader than buying AI vendors' licenses: the firm describes its tech strategy spanning application modernization, data centralization and AI integration. It is rolling Microsoft 365 Copilot across more than 13,500 professionals while also developing proprietary AI solutions.The CBIZ deal then expands the installed base over which that investment can be amortized. A larger professional-services platform can reuse the same data infrastructure, governance, workflow engineering, security, agents and specialist capabilities across more engagements. The economics of AI therefore potentially improve with scale. In fact, CBIZ itself illustrates the consolidation sequence. It acquired Marcum in a transaction valued at approximately $2.3 billion in 2024, with the transaction closing that November. Less than two years later, the enlarged consolidator became the target in the $5 billion Grant Thornton transaction.The playbook here is AI-enabled transformation of an existing institution. And I would challenge one simplification: this is not merely “put AI tools into a traditional accounting firm.” At scale, the harder work is redesigning the operating model underneath those tools—data, workflows, talent, pricing and controls.Archetype 2: Own the software control pointMotive + LawXThe second model does not acquire the professional-services firms at all. Instead, it tries to own the software layer on which those firms operate. Motive's investment in LawX is a good case study.LawX is building an AI-native legal operating system for notaries and law firms. Rather than buying the firms, LawX sells them the system of record and workflow layer: case management, onboarding, documents, communications, billing and AI-driven automation. Motive describes a land-and-expand SaaS model and sees the opportunity ultimately extending into embedded financial services.The strategic thesis is:Own the workflow. Control the data. Monetize the software and adjacent financial flows.A software provider can achieve extraordinary scalability, as it does not need to finance every customer acquisition with M&A capital. But it has less control over how its customers reorganize staffing, change pricing, redesign incentives or monetize the capacity AI releases.The professional firm retains the service P&L.For investors, the core question becomes:Can software win the workflow cheaply enough that owning the service provider is unnecessary?In industries with standardized workflows, strong SaaS adoption and relatively low customer-acquisition friction, the answer may well be yes.Archetype 3: Own the service and rebuild productionLimeTax, Current/Thrive and General CatalystThe third model goes further.Instead of selling software into fragmented professional-services firms, buy the firms—or meaningful ownership stakes in them—and transform production from inside.General Catalyst has articulated the thesis particularly clearly. It describes its model as combining applied AI with strategic acquisitions of service businesses and argues that, in sticky industries, acquisitions can be the fastest route to customers. Its European team puts the point even more directly: this is an AI transformation and product-reinvention strategy in which acquisitions provide distribution. LimeTax applies that logic to German accounting and tax.Rather than selling its in-house software ATLAS as standalone SaaS, LimeTax brings accounting practices into a common group and operates the AI platform across them. The €30 million acquisition facility attached to its latest financing makes the strategy explicit: M&A is part of the technology deployment model.Thrive Holdings has taken an even more vertically integrated approach. OpenAI took an ownership stake in Thrive Holdings in late 2025 and agreed to embed research, product and engineering teams inside its operating companies, initially concentrating on accounting and IT services. OpenAI and Thrive describe the idea as transformation “from the inside out”: own businesses where domain experts, workflows and real-world data can work directly with frontier-model engineers. Current—formerly Crete Professionals Alliance—is the accounting platform inside that model. By June 2026 it reported almost 30 partner firms, more than 2,000 employees and more than $500 million of annual revenue. Its Tax AI pilot processed about 7,000 returns and reported an average 31% reduction in preparation time, with accuracy of up to 98%. Those are company-reported pilot metrics rather than audited evidence of portfolio-wide margin expansion, but they provide a meaningful early operating proof point.OpenAI's engineering write-up is perhaps even more revealing. Thrive practitioners and OpenAI engineers used production failures, accountant feedback and evaluation infrastructure to improve Tax AI, then began reusing the same architecture in bookkeeping, audit and IT workflows.That is the real thesis:The acquisition creates distribution and workflow volume.The workflow generates exceptions and professional feedback.The feedback improves the technology.The better technology increases the economics of the next acquisition.That is fundamentally different from a conventional roll-up.The three models differ because they control different parts of the value chainThe distinction between these archetypes is ultimately not about who uses the most sophisticated AI. It is about how much of the production system the investor controls.A software-led company such as LawX controls a technology layer, but its customer still controls the workforce, pricing, operating model and service P&LAn incumbent such as Grant Thornton already controls the client relationship and the people delivering the service, but must modernize a large existing organization around new technologyAn AI-enabled roll-up such as LimeTax or Current goes a step further: it uses ownership itself as the mechanism for changing how the service is producedThe more of the service business an investor owns, the more important it becomes to decide which parts of the AI stack should be rented, which should be shared centrally and which must remain close to the operating company.Where should an AI-enabled professional-services platform actually differentiate?The temptation is to think that owning better AI means owning a better model.For most professional-services investors, that is probably the wrong layer to compete on. Frontier intelligence is increasingly available from OpenAI, Anthropic, Google and others. Training a proprietary foundation model would require enormous capital while exposing the platform to rapid technological obsolescence. The more defensible assets are likely to sit above the model.Think of the professional-services AI stack as four economic layers.1. Foundation intelligence: rent itAt the bottom are frontier models, which supply increasingly powerful general reasoning, language and tool-use capabilities. For most PE-backed professional-services platforms, these should be treated as infrastructure rather than proprietary IP. The objective should be model flexibility—not dependence on a single vendor.2. Professional intelligence: teach the models what “good” looks likeGeneral intelligence is not the same thing as expert professional judgment. This is where companies such as Scale AI, Mercor and Handshake increasingly fit. They began with data labeling but are moving toward expert datasets, model evaluation, reinforcement learning, red-teaming and realistic professional environments.Mercor's APEX work, for example, evaluates AI on multi-step tasks drawn from investment banking, consulting, law and increasingly accounting. Scale provides expert-data generation, RLHF, evaluation and safety testing. Handshake recruits academic and professional specialists to train and validate models.But there is an important ownership distinction. A platform can outsource the labor-intensive process of recruiting experts, structuring tasks and building evaluations. It should not outsource the proprietary knowledge that emerges from its own operations.The roll-up should retain ownership of professional rubrics, exception histories, correction data, workflow-specific evaluations and the definition of what acceptable work looks like - those are potentially much more durable assets than the foundation model itself.3. Enterprise context: own itA frontier model may understand U.S. tax law. But it doesn't know which subsidiary belongs to this particular client, which prior filing is authoritative, what happened in last year's engagement, which professional has sign-off authority, which document superseded another or which local workflow exception matters. That is the enterprise-context problem.For a professional-services platform, this may become one of the most important sources of differentiation. The relevant asset is not simply “more data.” Knowledge graphs, semantic layers and retrieval architectures can help create that structure.The strategic value does not come from owning a particular graph-database technology. Instead, it's from owning the ontology and institutional memory encoded within it. Over time, the platform can begin to capture why a professional overrode the default answer in an unusual situation.4. Agentic workflow: turn intelligence into productionThe final layer is where AI stops generating answers and starts producing work.An agent can: retrieve client documents, reconcile accounts, classify transactions, populate workpapers, request missing information, draft outputs, route exceptions and potentially write results back into enterprise systems. This is the layer that turns AI from a productivity tool into a new production system. And this is also where the three archetypes diverge most sharply.A software company can build the workflow and sell it. An incumbent can deploy that workflow across its own organization. An AI-enabled roll-up can build the workflow once, deploy it across acquired firms, collect exceptions and corrections from each deployment, improve it and then use the enhanced system in the next acquisition.That creates a potential flywheel:More acquisitions → more workflow volume → more exceptions and corrections → better workflows → faster transformation of the next acquisition.That is the deeper logic of an AI-enabled roll-up.Conclusion: the AI roll-up thesis is really about where value compoundsThe evolution of professional-services consolidation is moving beyond a simple question of whether firms should “use AI.” The more important question is where an investor wants to own the economics of AI-enabled value creation.In the first model, AI is layered onto an existing scaled services platform. In the second, value is captured through the software and workflow layer. In the third, the investor owns the service business itself and uses M&A to acquire customer relationships, workflow volume and professional expertise—then attempts to turn those assets into a continuously improving production system.That third model has the greatest potential for compounding, but also the highest execution burden. The winning platform does not need to own the frontier model. It needs to know what intelligence to rent, what professional context and workflow knowledge to own, which capabilities to centralize, what judgment and trust must remain local, and how to convert every new acquisition into reusable learning for the next one.For private equity, that changes the underwriting question. The test is no longer simply whether AI can reduce hours or whether a consolidator can acquire more firms. It is whether the platform can repeatedly turn acquired distribution into better workflows, stronger same-store economics, more institutionalized client relationships and a faster, more scalable integration engine.In that sense, the evolution is straightforward:Roll-up 1.0 bought EBITDA.Roll-up 2.0 built scale and operating platforms.Roll-up 3.0 is trying to acquire distribution and rebuild the production system itself.If that learning loop becomes real, AI can make each successive acquisition more valuable. If it does not, the result is simply a traditional roll-up with a larger technology budget.