AI for private equity portfolio companies: a value-creation playbook

Because the old levers are weaker. Bain's Private Equity Midyear Report 2026 estimates that a deal which needed 5 percent growth in EBITDA (earnings before interest, taxes, depreciation and amortization) a decade ago would now need 12 percent to reach a 2.5 times return over a five-year hold. The same report calls AI one of the most important value-creation opportunities across the portfolio, and notes that cost reduction is often the early focus.

The evidence of results is thinner than the enthusiasm. In the Bain and StepStone general partner outlook for 2026, 39 percent of general partners said they do not expect AI to have any material financial impact on their portfolio companies this year. Across all industries, McKinsey's State of AI survey found that 37 percent of respondents attribute at least some earnings impact to AI, and only about 6 percent attribute 5 percent or more of earnings before interest and taxes, as reported in a September 2026 summary of the survey.

That gap is the opportunity. Most companies have tools and pilots; few have rebuilt a workflow so that the profit and loss statement shows it. An operating partner who gets two or three workflows into production per company is ahead of most of the market.

In the workflows that repeat hundreds of times a week and are currently done by people reading, typing or calling. Services businesses, from home services and insurance agencies to lenders and industrial contractors, share the same handful of candidates, and the table below lists them with the metric that proves the gain.

Start with the companies where a workflow is large, measurable and owned by a willing chief executive, not with the company that needs help most. The first build has to succeed visibly, because every later company will ask what happened at the first one.

Score each company on five questions, then work down the list. Time remaining in the hold matters: a build that pays back after exit is worth less to the fund, although a buyer's diligence team will still look for it.

• Volume: how many times a week does the target workflow run?
• Measurability: is there a baseline today, or will you need a quarter to establish one?
• Data access: can the systems involved be read through an application programming interface (API) or export?
• Sponsorship: will the chief executive assign an owner and protect their time?
• Similarity: how many other companies in the fund run the same workflow?

Separate what is common from what is company-specific. The common core includes the connectors to widely used systems, the agent logic, the review screens, the test sets that check quality, and the reporting. Each company gets its own configuration, scripts, rules and data store, so nothing crosses between businesses.

Bain's field notes on generative AI in portfolios describe how large sponsors enforce this discipline: Vista requires each company to submit generative AI goals and quantified benefits in its annual plan, and Apollo asks companies to choose three to five use cases tied to near-term priorities. The same piece reports that nearly 20 percent of portfolio companies surveyed had operationalized use cases and were seeing concrete results. Our AI implementation approach follows the same pattern, and the 90-day plan shows what the first company's rollout looks like week by week.

Because without them, every company reports success in its own currency and the fund cannot compare or add anything up. One company counts a call as handled when the agent picks up; another counts it only when a job is booked. Both numbers are true and neither can be compared.

Write a short metric dictionary before the first build: name, formula, source system, owner and exclusions for each measure. Then report every company against it in a single portfolio dashboard on shared definitions. The dictionary also protects the savings case at exit, when a buyer will recompute every claim.

• Count hours saved only where they turned into lower cost or more output, never as a standalone figure.
• Measure against a baseline taken before go-live, using the same formula.
• Report revenue influenced separately from revenue caused, and say which is which.

Ask about the data and the workflows, not about the AI strategy slide. A target with clean records of calls, jobs, invoices and documents can be improved quickly; one whose operations live in spreadsheets and inboxes needs data infrastructure before AI, and the plan and price should say so.

• Which three workflows consume the most staff hours, and how is each one measured today?
• Where does operational data live, and can it be extracted without the vendor's permission?
• What AI is already in use, who pays for it, and does any of it touch customer data without a written agreement?
• If the company claims AI in its product, what is proprietary and what is a third-party model behind a screen?
• Who on the team could own an AI workflow after the builder steps back?

The main risks are overstated savings, data leaking between portfolio companies, dependence on one vendor, and teams quietly working around a tool they do not trust. Each one is contained by design rather than policy: separate data stores per company, connectors kept apart from the logic so a model provider can be swapped, a review step on anything customer-facing, and usage tracking that shows whether people actually rely on the system. Software margins are a warning here, too: in a Bain analysis of 33 software buyouts, 31 projected a median 560 basis points of margin improvement over five years, and actual margin growth badly trailed those models. Underwrite AI savings the same skeptical way.