How a Private Equity Firm Gets AI Into Its Portfolio Companies: The Operator’s Playbook

Seventy-three percent of private equity executives expect AI to raise the value of their portfolios in the next twelve months. Eight percent say it is materially moving EBITDA today. Those are Alvarez and Marsal's numbers from their 2026 survey of a hundred PE executives and operators, reported by PitchBook, and the gap between them is the whole subject of this essay. FTI Consulting's 2026 survey of 200 fund and operating leaders shows the same shape from the other side: 95 percent say their AI initiatives met or beat the original business case, yet only about a third say their portfolio companies use AI in day-to-day operations. The pilots work. The companies do not change. It is not a technology gap. Every firm in those surveys has access to the same models. It is an execution gap, and execution is an operator's problem.

I have sat on both sides of it. For a decade I was a managing partner at Fortress Capital Partners, with fifteen-plus exits and a board seat's view of what a value-creation plan looks like when the operating team is not bought in. Before and after that I was the operator: founder and CEO of Interland through a Nasdaq IPO, EVP at New Relic, VP at Cox Automotive running a $250 million budget through a cloud and DevOps transformation, COO in a turnaround, and CEO of Product Genius, where we built and launched an AI engagement platform and rolled it out nationally in under six months. What follows is the playbook I use now, written for the partner who owns the AI mandate and the CEO who has to deliver it.

Start with the rule that makes AI legible to a P&L

Here is the operating rule I apply to every portfolio company. Any process you can model can be mostly automated, with the exceptions routed to humans.

That sentence does three things for a sponsor. It tells you where to look: not at the org chart, at the processes. It tells you what "done" means: most of the volume handled without a person, the rest handed to one. And it tells you what to measure: the share of a workload's volume that no longer touches a human queue.

Applied to customer service, customer success, support and claims processing, that pattern typically takes 30 to 70 percent of the work off human queues. The public record brackets the same range. Klarna's assistant handled two thirds of its customer service chats in its first month, the equivalent of 700 full-time agents, by the company's own account. Independent tests of off-the-shelf agents land much lower, near 38 percent. The spread between those numbers is not the model. It is whether someone modeled the process first and designed the exception path. That is execution, and it is where the 8 percent separate from the 73.

Run a workload sweep, not an AI strategy

The document a portfolio company does not need is another AI strategy. What it needs is a workload sweep: a list of every recurring process in the business, scored on four questions.

One, can it be modeled? If the inputs and the decision rules can be written down, it qualifies. If every case is a judgment call, it does not, yet.

Two, what is the volume, and what does a unit of it cost today? This is the EBITDA line, in advance.

Three, what happens to the exceptions? Every workload has a residue that must go to a person. If nobody owns that path, the automation fails on the first hard case and the team stops trusting it.

Four, who inside the company will own it after the consultant leaves? A workload with no internal owner is a pilot, and pilots are where the 73 percent live. MIT's NANDA initiative put a number on that in 2025: 95 percent of enterprise generative AI pilots deliver no measurable P&L impact. The ones that do are owned inside the business.

Rank the list by cost times confidence, and start with the boring wins: the high-volume, low-glamour workloads that nobody puts on a board slide. They fund the next one, and they teach the team the muscle before anything ambitious is attempted. Across a portfolio, the sweep does something a single company cannot: the same workload shows up in six companies, and the second implementation costs a fraction of the first.

One anonymized example of what the sweep becomes when it takes hold. A private equity company stood up a whole-of-company AI initiative: every function, not just engineering, charged with finding what in its own work could be modeled and automated, and bringing the idea or the working automation back. That flips AI from an IT project into an operating cadence, and it's where the durable EBITDA comes from, because the people closest to the process find the wins no consultant's sweep would.

Expect the bottleneck to move

Engineering is where sponsors see the biggest multiplier and the biggest surprise. In software teams, output runs three to seventeen times with AI, depending on the codebase and the guardrails. That range comes from my own builds and from direct conversations with engineering leaders at Fortune 500 enterprises, and the independent telemetry is now catching up to it. Faros AI's 2026 study of 22,000 developers across more than 4,000 teams found epics completed per developer up 66 percent at high AI adoption. It also found pull-request review time up 91 percent, bugs per developer up 54 percent, and 31 percent of code reaching production with no review at all.

Read those together and the lesson is plain. When coding stops being the constraint, the constraint moves to product and to QA. The 2025 DORA report calls AI a mirror and a multiplier: it amplifies whatever operating discipline was already there. A portfolio company that adds AI coding tools without adding product judgment and review capacity gets more code, more bugs, and no more EBITDA. Plan for the new bottleneck before you buy the multiplier, or the multiplier eats itself.

Decide who owns the mandate

The reason most AI plans stall is that nobody owns them at the level where trade-offs get made. The role that fixes this has a name now: the AI operating partner. Heidrick and Struggles defines it as an operating partner who drives AI-enabled value creation across a portfolio, and the definition is right. The search firms describe the talent pool as fund-level PE operators on one side and machine-learning consultants on the other. What they cannot tell you is how to test a candidate, because the profile that passes is rare. Here is the test I would apply, whether the seat is full time, fractional, or a consultant on a scoped engagement.

Has this person shipped an AI product, not advised on one? Building and launching is the only way to know what breaks in production.

Has this person run a P&L of real size? Otherwise the AI plan will be technically elegant and commercially beside the point.

Has this person sat on the sponsor's side of the table? The plan has to survive an investment committee and an exit narrative, not just a demo.

Can this person get an operating team to adopt something they did not ask for? That is the actual job, and it is a leadership job, not a data science job.

Fewer people pass all four than you would think. Most AI consultants pass the first. Most operating partners pass the second and third. The plan needs all of them in one seat, or in a small team that reports to one.

Run it across the ownership cycle

A sponsor's AI work does not start at the first board meeting after close. It starts in diligence and ends in the exit story.

In diligence: run the workload sweep on the target before you own it. It tells you what the EBITDA bridge is worth and whether the management team can execute it.

At close: the AI plan belongs in the first hundred days, and it survives contact with the operating team only if the team helped write it and owns the first workload.

During the hold: sweep the portfolio, find the same win in several companies, build it once, adopt it everywhere, and measure adoption, not licenses. Licenses are a cost; adoption is the return.

At exit: the AI story a buyer will pay for is not "we deployed AI." It is a list of workloads, the share of volume each one handles without a human, and the margin it moved. That is a number a buyer can diligence.

What to do this quarter

If you are the partner with the mandate, pick two portfolio companies, not ten. Run the sweep in each. Put one boring workload from each into production with an internal owner and a measured share of volume off the human queue. Report the two numbers to the investment committee next quarter. Then do the next two companies with what you learned.

If you are the CEO who has to deliver it, ask your sponsor for one thing: the authority to say no to AI projects that fail the four sweep questions. The fastest way to become one of the 8 percent is to stop funding the pilots that keep you in the 73.

Where I fit

I do this work three ways: an AI roadmap, which is the sweep with numbers behind every workload; an adoption sprint, which takes one or two workloads to production with your team; and fractional AI leadership, where I hold the mandate across a company or a portfolio without the full-time hire. I am based in Alpharetta, Georgia, work remotely with sponsors and CEOs everywhere, and come on site when the work calls for it. If your firm is trying to get from the 73 percent to the 8, the first conversation is forty-five minutes on video, and if I am not the right person for it I will say so.

Common questions

What does an AI operating partner do? Drives AI-enabled value creation across a portfolio: the workload sweeps, the EBITDA bridge, the internal owners, and the adoption. The seat can be full time, fractional, or a consultant on a scoped engagement; what matters is that someone owns the mandate at the level where trade-offs get made.

How do you choose an AI operating partner or consultant? Apply four tests. Has this person shipped an AI product, not advised on one? Has this person run a P&L of real size? Has this person sat on the sponsor's side of the table? And can this person get an operating team to adopt something they did not ask for? Fewer people pass all four than you would think.

How much of a process can AI actually automate? Any process you can model can be mostly automated, with the exceptions routed to humans. In customer service, customer success, support and claims processing that typically takes 30 to 70 percent of the work off human queues. The spread is execution, not the model.

How do you measure AI in EBITDA? Per workload: the share of volume that no longer touches a human queue, times what a unit of that volume cost before. Measure adoption, not licenses, and report a baseline and a measured result the CFO can audit. That is the number a buyer can diligence at exit.