AI and the PE roll-up: where the real value creation has moved

The classic private equity roll-up has one piece of maths at its heart. Buy small businesses at 4–5x EBITDA. Combine them. Sell the platform at 10–12x. The gap is the return, juiced by leverage.

That maths still works. What's changed in the last two years is where the operational story underneath it sits. AI hasn't replaced the roll-up thesis. It's quietly relocated the value creation from multiple arbitrage to margin expansion — and that changes which roll-ups are worth doing.

The historical playbook

For decades the value creation story in a roll-up has run something like:

1. Buy fragmented operators in a sector with weak central infrastructure.
2. Consolidate back office — finance, HR, procurement, IT.
3. Negotiate better supplier terms with scale.
4. Apply some commercial discipline that small operators lacked.
5. Sell the platform to a bigger PE firm or strategic acquirer.

Gross IRRs of 20–30% are typical when executed well. EBITDA margins might lift a few points. Most of the return comes from the multiple, not the operating performance.

This works when the back-office consolidation actually pays — i.e., when the cost stack at the small-operator level is dominated by duplicated administrative function.

What's actually different now

The new generation of AI-enabled roll-ups — Accrual rolling up accounting practices, Dwelly in UK lettings, Savvy Wealth in advisory, Carr Riggs & Ingram with Centerbridge — share a different thesis.

They're not just consolidating overhead. They're targeting sectors where the operator's daily workflow itself is automatable. The accountant's data-entry hours. The lettings agent's tenant-comms hours. The advisor's client-prep hours.

That's a categorically different lever. Back-office consolidation might pull 2–4 points of EBITDA margin out of a fragmented sector. Workflow automation at the operator level — done well — can double margin in 12–18 months. Not in every sector. But in any sector where the work is high-volume, document-heavy, and rules-bounded, the maths is meaningfully different from the old playbook.

Carr Riggs & Ingram, post the Centerbridge/Bessemer investment, is publicly targeting growth from ~$500M to $1.2B over five years on the back of this. That's the kind of number that only works if the AI margin thesis is real.

Where it pays — and where it doesn't

The roll-up sectors where AI materially changes the economics share three properties:

• The work is mostly information, not physical
• The expert spends real hours on repeatable, rules-bounded tasks
• The customer doesn't mind, or doesn't notice, if a model is doing the first pass

Accounting, customs brokerage, conveyancing, insurance broking, parts of legal services, certain medical specialities, property management, certain B2B service categories. These are the obvious candidates and they're the ones currently attracting capital.

Where it doesn't change the maths much: HVAC, roofing, landscaping, dental practices, restaurants. The work is physical, the customer is local, and AI helps at the edges — routing, scheduling, marketing — but doesn't restructure the unit economics. These can still be good roll-ups. They're just not AI roll-ups. The capital chasing them on an AI thesis is mispricing the opportunity.

What this changes about diligence

If the margin thesis is the source of return, the diligence has to validate the margin thesis. That sounds obvious. In practice, we've seen PE deals priced on AI-margin assumptions that nobody had stress-tested operationally.

The questions worth asking before pricing in the AI uplift:

• Has anyone in this sector actually achieved this margin profile, or is it a model? If the answer is "consultants modelled it," derate.
• What does the workflow actually look like today, at the operator level? Site visits, not just data rooms.
• Where does the work currently fail when a model gets involved? The failure modes are usually where the real cost sits.
• Who in the target organisation will own the implementation post-close? If the answer is "we'll bring someone in," add 12 months to the timeline.

The AI-enabled diligence tooling — Blueflame, Third Bridge, the cohort of vendors selling into PE — is genuinely useful for the first pass. It compresses screening cycles from weeks to days. But it doesn't validate the margin thesis. Only operational work does.

The integration trap

Roll-ups fail in integration. They've always failed in integration. The PE literature is consistent on this — 70%+ of value destruction happens in the first 100 days of post-merger integration.

AI doesn't fix this. In some ways it makes it worse, because the temptation is to roll out the automation before the integration is stable. The pattern looks like:

1. Acquire operator. Promise the team the AI tooling will free them up, not replace them.
2. Discover the tooling needs custom workflow design per operator, because none of them did things the same way.
3. Run two parallel workflows for six months. Both more expensive than the original.
4. Either kill the project or kill morale. Sometimes both.

The discipline that works: stabilise the operating model first, standardise the workflow second, then automate. Reversing the order is where the value destruction sits.

The PE-AI joint venture moment

The recent moves by OpenAI and Anthropic — both launching enterprise services arms with PE backing in May 2026 — are worth paying attention to. The thesis is that the deployment problem in mid-market PE portfolios is significant enough to be a business in its own right.

We think they're right about the deployment problem. We're less sure that the model providers are the right vehicle to solve it. The mid-market gap isn't AI capability — it's the operational translation between capability and a specific sector's workflow. That's a different skill set, and it's the one most PE firms are underweight on internally.

What we'd tell a GP

Three things, in our view.

The AI margin story is real, but it's bounded. Apply it where the work is informational, rules-bounded, and high-volume. Don't pay for it where the work is physical or relational. The capital allocation discipline matters more than the AI thesis.

Underwrite the operating team, not the technology. The deals that compound are the ones where a sector-experienced operator partners with a small AI implementation team. The deals that struggle are the ones where the technology team is in charge and learns the sector on the job.

Earn the margin before pricing it in. If your model assumes EBITDA doubles in 18 months, build the deal with margin of safety. The maths only works if execution lands. Most of the time, it lands slower than the IC memo assumed.

The AI roll-up is a real strategy. It's just not the strategy most of the marketing makes it out to be — and the difference matters.