The durable asset in an AI roll-up should be a governed way of delivering work, not merely a collection of acquisitions equipped with agents. The operating principle is straightforward: agents perform work; accountable people decide which rules become active and which outputs reach clients.
Greg Isenberg gives that principle practical form in his post on AI roll-ups: the reviewer can block but never ship, and the preparer can ship nothing on its own. His rulebook process requires practitioner approval before activation and turns approved recurring corrections into test cases. These checkpoints make an important distinction visible: capability does not confer authority.
For private equity and roll-up operators, the question is how to preserve that distinction as more work and more businesses enter the system.
Two ways into the operating problem
“The Roll-Up Thesis Lives in the Dependencies” contributes the architectural test: distinguish useful output, authorized action and improved business outcomes, then examine the information access, exception handoffs, measurement and transfer conditions that connect them. “Underwrite the Operating Platform” contributes the investment and operating test: identify accountable owners, examine actual work, cost the review burden and make further commitments conditional on readiness.
They differ in where they put the reader’s attention. “The Roll-Up Thesis Lives in the Dependencies” asks whether an operating gain can be produced and transferred under known conditions; “Underwrite the Operating Platform” asks what an operator must demonstrate and resource before taking the next step. Read “The Roll-Up Thesis Lives in the Dependencies” for the causal chain and transfer test, and “Underwrite the Operating Platform” for the underwriting discipline.
The useful tension is between operating coherence and evidence-led expansion. A platform needs enough common structure to carry learning across firms, but its scope should grow through workflow evidence rather than architectural confidence. Reusable output is not enough: the value must survive the permissions, handoffs and costs required to produce it in the receiving business.
An operating gain and a justified acquisition remain separate tests. A promising intervention does not settle the purchase price, risks or alternatives; an attractive purchase does not demonstrate repeatable AI improvement. When the investment case materially depends on that improvement, the operating evidence belongs inside the underwriting rather than alongside it as an independent promise.
Folders organize the starting point
Isenberg’s common folder structure centers on acquisition criteria, operating rules and a corrections log. It gives a solo buyer or small team a starting place for buying discipline, practitioner knowledge and lessons from completed work.
But a shared folder shape is not proof of integration. Matching files do not settle which rules apply, who may activate them or how a local exception changes the work. When acquisitions outrun the capacity to absorb their workflows and decision requirements, another company can add unresolved operating problems rather than operating leverage.
A platform should therefore be tested through one bounded workflow before its reach is widened. The test is not whether the software can be distributed. It is whether a particular piece of work can be delivered within understandable checks and accountable decisions.
There is a serious counterargument to building a platform first: construction can become another layer of cost before the operator has understood the work. A larger specification can organize uncertainty without removing the integration bottleneck. A bounded workflow makes the platform answer to an operating problem rather than allowing the architecture to define its own success.
Let the work earn the next step
Begin with real, appropriately anonymized work samples. Map the inputs, checking standards and consequences of error before deciding what agents should take on. Isenberg distinguishes tasks that can be automated, tasks requiring human review, tasks suitable only for assistance and tasks that should remain human.
That classification matters because a firm is not one automation opportunity. A repeatable task with checkable output presents a different decision from work whose correctness depends on a practitioner’s judgment or a particular client relationship.
Isenberg recommends shadow mode: agents do the work in parallel with the existing human process, and the outputs are compared. He also recommends tracking minutes of human attention per job each week. The comparison needs both dimensions. Acceptable output with an unsustainable review burden leaves the operating problem unresolved; fast review without acceptable output leaves the quality problem unresolved.
The production transition should be explicit: after shadow-mode comparison supports acceptable output and sustainable human review, move a suitable high-volume, low-risk task into production only with accountable human authorization and every draft reviewed.
Give a human workflow owner responsibility for tracking review effort each week. Expand only when output reliability, review cost and client and staff retention support the decision. Readiness belongs to the scope and conditions actually evaluated; a successful test of one workflow does not settle readiness for the next.
Isenberg’s proposed first-hundred-days timetable moves from shadow comparison to reviewed back-office work and then careful expansion. Use that sequence to organize the work, not to substitute a date for the evidence and authorization each transition requires.
The same discipline belongs in acquisition underwriting. Allow for tooling, integration and sustained human review rather than treating faster draft production as the whole delivery model. Isenberg also cautions against relying on headline results elsewhere and recommends underwriting a deal on the buyer’s own numbers.
Make integration readiness part of the next acquisition decision. The prospective business has to be assessed against the operating capacity it will require, not just the opportunity it appears to offer.
Reuse the lesson, preserve its limits
The corrections log can connect a local improvement to a wider operating platform. It can also spread a mistake more efficiently.
Isenberg distinguishes factual errors, client preferences, missing information and style changes when comparing agent drafts with human-approved work. Those categories help expose an essential difference: a correction records what changed in a particular job, a proposed rule asks whether the lesson should govern another job, and an active rule has passed the relevant approval decision.
A portfolio platform should make corrections reusable where they apply, with permissions, human approval and local exceptions intact. Before extending a lesson, preserve its originating context, intended scope, relevant test and any conflicting local requirement. Permission to share the material and authority to adopt the rule are separate questions.
A client’s preferred wording may belong to that relationship. A missing-information check may apply to a particular service. A factual correction may support a broader rule only where the conditions that made it correct still hold. Without that distinction, shared learning can become shared error.
The opportunity is to make an applicable lesson available beyond the firm that discovered it—not to make every accepted correction universally active.
Keep the acquired business inside the test
The operating platform should improve delivery without consuming the relationships and practitioner knowledge that make the acquisition worth pursuing. Isenberg recommends measuring client and key-person retention alongside margin and warns against degrading relationships while changing back-office work.
That also constrains the meaning of greater autonomy. A lower correction rate may support reconsidering a checkpoint, but it does not itself transfer release authority to an agent. Changing the boundary remains an accountable decision.
Governed growth is therefore a continuing discipline: preserve decision rights, demonstrate a bounded workflow, account for the attention it requires and extend the lessons only where their conditions hold.
For owners who do not sell, federation could offer access to shared operating capabilities while preserving ownership and decision authority. The governing requirement would be accountable participation, not common ownership.
Source note
Based on personal communication with Breyden Taylor and Sabeel Ahmed, retrieved October 4, 2026.