Bain's 2026 Global Private Equity Report puts the arithmetic plainly. Through the 2010s a typical buyout needed about 5 percent annual EBITDA growth to return 2.5 times invested capital; the deals being done now need 10 to 12 percent, because cheap debt and expanding multiples are gone. Bain's phrase for it is "12 is the new 5". That growth has to come out of operations, and AI in private equity is the lever every operating partner is now asked about at every portfolio review. The question worth answering is narrower than whether AI matters. It is where, inside a mid-market portfolio company, it pays during a hold period, and the evidence on that has become specific.
What the evidence says about where AI pays
McKinsey's analysis, published in June 2026, is the most useful dataset so far. Its authors sorted 471 private-equity-backed companies across 31 industries into four levels of AI use: opportunistic pilots, AI embedded in the operating model, AI embedded in the product, and AI used to build new businesses. The valuation results point in one direction. Companies at the second level traded at a median revenue multiple of 14x against 13x for the first, a difference McKinsey calls negligible. The jump comes at the third level, 20x, and the fourth, 31x. Markets pay for AI that changes what a company sells, and pay little for AI that changes how it runs.
Read that as a fund would, though, and the conclusion inverts. Most portfolio companies are not software businesses. A distributor, a contract manufacturer or an insurance brokerage cannot embed a model in its product within a five-year hold, and the multiple expansion at levels three and four is mostly unavailable to it. What is available is the operating-model level, where McKinsey found revenue per employee roughly 20 percent higher than at level one after inflation. That is the number that feeds EBITDA, and EBITDA growth is where the 12 percent has to come from. The multiple is set by the buyer at exit. The earnings are made every month in between.
The portfolio is further behind than the fund assumes
The same McKinsey work reports that many portfolio companies are earlier in their AI adoption than their owners believe, and that fund-level activity has run ahead of company-level results. EY's AI Pulse survey of US executives in late 2025, with 50 private equity respondents among its 500, found that 84 percent of private equity firms had appointed a chief AI officer and that nearly half were putting between a quarter and half of their budgets into AI projects. The same respondents said, 62 percent of them, that their organizations struggle to link specific productivity gains to AI adoption. Spend at the fund, uncertain results in the companies: that is the shape of the gap, and it is the gap an operating partner is paid to close.
The reason it persists is structural rather than technical. The fund has a chief AI officer; the portfolio company has a controller who closes the books in a spreadsheet, a two-person AP team keying invoices, and no one whose job is to change any of that. We describe that company in the customer all four delivery models leave behind. It cannot hire an AI team and would not know how to run one. The capability has to arrive from outside, as a delivered result, or it does not arrive.
Why the back office is where an operating partner starts
The activities that models handle best are the ones that fill a mid-market back office. The McKinsey Global Institute's 2017 study of automation, still the reference, found that fewer than 5 percent of occupations could be fully automated but about half of all paid activities could be, and that data collection and data processing sit alongside predictable physical work at the top of the list. Accounts payable is data processing. So are cash application, order entry, vendor onboarding, payroll checks and the reconciliations that make up a month-end close. Ardent Partners, in its 2023 State of ePayables report, put the average invoice exception rate at 20.7 percent, which means a fifth of the volume in a typical AP function is handled by a person deciding what to do with a mismatch.
Those processes have three properties an operating partner wants. They are high-volume, so the saving is measurable in hours and headcount within a quarter. They are the same in every company in the portfolio, so a fix built once travels. And they do not require changing the system of record. A delivered agent works on top of QuickBooks or a legacy ERP rather than replacing it, which keeps the program out of the two-year migration that most portfolio companies have already decided not to do. We walked through a month-end close rebuilt this way in the month-end close, automated and a matching process in three-way match, rebuilt.
Rent or buy, and the third option
McKinsey frames the fund's choice as rent versus buy: centralize platforms and expertise at the fund and redeploy them across companies, or embed dedicated AI teams in the portfolio companies where proprietary data justifies the cost, with many leading firms settling on a hybrid. The framing is right about the constraint, which is talent, and we have set out the cost of that talent in build, buy, or hire an AI capability. It is incomplete about the third route, which is to contract a firm that builds the agent, runs it in production, and is paid on the savings it produces. For the operating-model level, where the work is a known process rather than a novel product, that route has two advantages over both a central team and an embedded one. The portfolio company gets a working result without a hiring plan, and the fund gets a playbook that repeats: the AP agent built for the distributor in Ohio is most of the AP agent the contractor in Texas needs. We describe how such an engagement is structured, and what keeps it aligned with the buyer, in what an aligned AI engagement looks like.
What a portfolio-wide program looks like
A program that survives its first board review starts with a baseline, not a tool. For each company, one function is measured as it runs today: hours per month, cost per transaction, exception rate, days to close. We set out the method in baselining a process for outcomes. The first deployment goes into the highest-volume manual process in that function, runs for a month against the baseline, and reports the same numbers again. Only then does the second company start, with the first company's build as the template. Sequenced this way, a fund with fifteen portfolio companies is running a repeatable operation by the fourth one, and the value creation plan carries a number per company that an exit buyer can diligence. The EY finding about gains that cannot be linked to AI is what happens when this order is reversed and the tool arrives before the baseline.
A reasonable counter, answered
A reasonable counter is that the multiples say to aim higher: the market pays 20x for AI in the product and 14x for AI in operations, so an operating partner should spend the hold period pushing companies up the ladder rather than automating invoices. There is something to it for the software companies in a portfolio, and for them McKinsey's third and fourth levels are the right ambition. For the rest, the arithmetic runs the other way. A distributor is not going to sell an AI product, and the 12 percent annual EBITDA growth its deal was underwritten on is due every year regardless. The operating-model gains are the ones a manufacturer, a broker or a services firm can bank inside the hold, and they compound across a portfolio in a way product bets do not. Start there, measure it, and let the multiple take care of itself.
Common questions
- Where does AI create the most value in private equity portfolio companies?
- McKinsey's 2026 analysis of 471 PE-backed companies sorted them into four levels of AI use and found the valuation premium concentrated at the top: companies embedding AI in their products traded at a median revenue multiple of 20x against 14x for companies using it inside their operating model. For the manufacturers, distributors and service businesses that make up most mid-market portfolios, the operating-model level is the one within reach during a hold period, and it showed roughly 20 percent higher revenue per employee than companies still running pilots. The practical answer is the back office: accounts payable, the month-end close, onboarding and order handling, where the volume is high, the work is manual, and the savings land in EBITDA.
- Should a private equity firm build a central AI team or put one in each portfolio company?
- McKinsey frames it as rent versus buy: centralize platforms and experts at the fund, or embed dedicated teams where proprietary data justifies the cost, with many leading firms running a hybrid. The constraint is talent. A $200 million distributor cannot hire an AI engineer and keep them, so the fund either supplies the capability or contracts a firm that delivers a working agent and stays on the result. The repeatable part is the playbook, because the same five or six processes recur in every company in the portfolio.
- How do operating partners measure whether AI is working in a portfolio company?
- By baselining the process before anything is deployed and reporting the same numbers after: hours per month on the task, cost per transaction, the exception rate routed to a person, and days to close. EY's late-2025 survey found that 62 percent of private equity respondents struggle to link productivity gains to AI adoption, which is what happens when a deployment starts without a baseline. A measured process makes the saving a line in the value creation plan rather than an assertion in a board deck.