Private equity stopped debating whether AI matters somewhere around 2024. The question this year is why so few firms can point to a positive impact from deployment.

Over the past couple of years, interest in AI’s ability to drive business change has taken off across private equity. This year alone, more than 50% of our private equity deal, operating, and management team conversations touch on AI, typically in the context of solving a specific business problem tied to the value creation plan. And these aren’t just conversations about software development or customer support anymore. They span the full spectrum of business functions. Implementation inside portfolio companies is now the most frequently cited AI priority we hear, ahead of firm-level strategy or diligence. Nobody needs convincing anymore that AI can transform how a business runs. The open question is how to actually make that transformation happen.

Participate in enough of these conversations and engagements, and a pattern emerges: pilots are everywhere, ambition is high, and leadership is pushing hard to see impact. But the honest answer to “what’s the ROI so far” usually falls somewhere between “hard to say” and “give it another two quarters.” Ask how they plan to measure these initiatives, and you’ll hear something closer to “we’ll take any measurable outcome, really.” Across all of these conversations, one thing is consistent: executives are hungrier than ever for use cases that will scale and move the company’s numbers.

We’ve looked at the pilots that never graduate to production, and a specific set of issues show up more than any other: resourcing, focus, and urgency. What we’ve found is that AI that scales tends to come from putting both technical and functional experts in the same room, not sequencing the effort and handing a strategy document from one group to the other after the fact.

There’s no single clean explanation for the gap between AI ambition, execution, and impact in this industry. But the funds and portfolio companies (portcos) that actually bridge it tend to share three disciplines:

Tag team: Put the person who knows the function in the room with the person building the solution

This isn’t the only reason AI initiatives stall in PE portcos, but it’s the most common one we see. An AI strategy built without deep functional context and input tends to look sharp in a deck but falls apart the moment it meets an actual workflow. Assuming a smart engineer will figure it out with help from the portco is overly optimistic. GPs and portcos generally run lean, so the people on the ground don’t have the extra bandwidth to commit to these transformational initiatives.

The instinct is to solve this with a hire, someone who has run the function and can also build. Those people exist, but there aren’t many of them, and the good ones aren’t available. Mentions of AI talent and workforce needs at portfolio companies are up more than twelvefold year over year in our data. That’s what a market looks like when everyone is hunting the same unicorn.

The firms getting real traction with AI aren’t choosing between a domain expert and a technical builder. They’re making sure the two actually build the thing together.

That doesn’t mean both are on the ground from day one. More often the functional operator goes first, diagnoses where the value sits, and the engineer comes in once there’s something real to build. What matters is that when the build starts, it’s a genuine pairing and not a relay. The operator doesn’t hand over a specification and walk away, and the engineer doesn’t take a document and disappear for a quarter.

It’s part of why we built Catalant’s Forward Deployed Expert model the way we did: an operator who’s actually run the function before, working alongside an engineer who understands the technology’s capabilities and limitations. This keeps the build grounded in how the function really works, and it’s an encouraging sign that this pairing is getting easier to find and deploy. The funds seeing AI actually move a number are, more often than not, the ones who plan for that pairing from the start, rather than treating strategy and build as two separate handoffs or betting the whole engagement on finding that extremely rare hybrid hire.

Be specific: Point every AI investment at a lever that’s already in the value creation plan

The prevailing advice from legacy consulting right now is to go big, go systematic, and be patient: build a playbook at one portfolio company, standardize it, and roll it out across the rest of the portfolio. That’s not a strategy built to survive a five-year hold. We’d argue the opposite is the superior strategy: go narrow, tie the bet to a lever that was underwritten when the deal closed, and hold a named owner accountable for driving change through AI.

The AI investments that pay off are rarely part of a generic firm-wide program. They’re specific bets, sized to a lever that was already sitting in that company’s value creation plan before anyone mentioned AI: pricing discipline at one portco, customer satisfaction at another, commercial productivity and effectiveness at a third. A small, well-aimed bet against a lever someone already owns will beat a portfolio-wide platform license without a specific objective every day. Some of the best-capitalized funds in the market have the least to show for their AI spend because of this. The initiative got blessed and the money got approved at the GP level, but none of it ever translated into a specific bet on a business problem that was underwritten when the deal was originally closed.

The related advice worth pushing back on is patience: i.e., the argument that AI needs multi-year committed capital and shouldn’t face a return test inside the year. That may be right for a public company with an indefinite horizon. But private equity doesn’t have that. A fund has a defined hold, an investment committee that underwrote specific levers at a specific price, and LPs that want their money back with interest. Value creation that can’t show progress inside that window isn’t strategy; it’s deferral with better vocabulary.

The mistake here is treating a technology decision (standardizing a tech stack, rolling out an enterprise license across the portfolio) as if it were a value creation decision on its own. It isn’t. AI is a tool, not a strategy, and waving a hammer around a job site doesn’t build a house. Blessing a generic AI strategy at the fund level across a myriad of industrials, services, and healthcare businesses won’t turn a 6x exit into an 8x one. That kind of multiple expansion only shows up once someone points AI at a specific business process and holds a team accountable for measurably improving it. That’s a narrower, better ask than “become an AI-enabled company; we trust you to figure it out.”

Build for the long term: Treat what you build now as part of the exit story, not just the hold period

Every AI investment a fund makes during the hold gets evaluated twice: once by the operating team deciding whether it’s worth the spend, and again, years later, by whoever’s running diligence on the other side of your sale process. Buyers and their advisors are getting better at distinguishing between a robust solution and what’s still an ambition. A portfolio company that can point to a working AI capability tied to a specific value driver (pricing, churn, cost-to-serve, headcount) tells a cleaner growth story at exit than one that can only point to a pilot and financial projections.

And a critical input for that success story is both unglamorous and time-consuming to get right at scale: Data. It takes time to overcome fragmented systems, inconsistent inputs, low-quality data, and no single source of truth. Plan your timeline accordingly to get this right. This is not new advice with technology projects, but it bears repeating. It’s a big piece of the puzzle most funds gloss over on the way to a pilot, and it’s usually the first thing a buyer’s advisors ask about when trying to understand how a number or recommendation was produced.

This stuff takes time to get right, especially once you factor in add-ons that could benefit from the capability but run on different tech and data stacks than the platform. The same discipline that gets an AI initiative to scale during the hold is exactly what makes the addbacks and the forecasts believable later. Done right, and done early enough, nobody has to do mental gymnastics to make the “AI-enabled” story in the Confidential Information Memorandum (CIM) and management presentations hold up. The proof is in the results.

Where this leaves us

AI is complex. The approach to using it effectively doesn’t have to be, and none of it happens on a single quarter’s timeline. If it were my money, I’d bet on the firms deploying AI in consistent, targeted use cases consistently over the firms with the biggest AI budget every time. The gap between AI ambition and AI impact has nothing to do with who wants it more. It comes down to which funds have built the discipline and specificity to convert conviction into something that shows up in the numbers, deal by deal, portfolio company by portfolio company.

That’s a harder thing to build than a mandate. It’s also the only version of AI strategy that actually survives contact with a portfolio company.