Private equity built its value creation playbook with financial engineers. AI is an engineering problem.
If you build value creation plans for a living, take a look at the AI section of your latest one. How much of it describes a better business, and how much describes the same business with fewer people?
Operating teams increasingly talk about an AI strategy for the portfolio. Too often, the plan underneath is a cost target with AI written next to it.
Peter Drucker put it bluntly:

"Marketing and innovation produce results; all the rest are costs."
It's a useful line to remember when an entire AI strategy begins and ends with the payroll.
The existing value creation playbook deserves its place. Pricing, procurement, sales effectiveness, working capital, add-on M&A: these are proven ways to improve a business. The cost work matters too. Shared services, fewer management layers and a more efficient back office can all create real value.
But AI can change what the company delivers, how quickly it delivers it and what a customer is willing to pay for. A plan built around doing the same work for less captures only part of that opportunity.
It can tell you how many roles you might remove. It tells you much less about whether the business will be more competitive afterward.
So here's the exercise I'd suggest. Set the savings aside long enough to ask what you would do with AI if the goal were to win more customers, keep the ones you have and give them something better. Then bring the cost line back into the discussion.
Every deal team knows where returns come from: earnings growth, the multiple at exit and the time it takes to get there. Most AI plans I see work on the first of those, and only on the cost side of it.
A cost line moves EBITDA. It does little for the multiple and nothing for the holding period. If AI is as important as everyone in the room says it is, it should be in all three.
The order matters. Start with a headcount target and every project has to justify itself against that target. Start with the business you want to own at exit and you give yourself a much wider set of choices.
A quick disclosure: at Limestone Digital, we embed AI engineers into PE-backed companies. I have a commercial interest in this argument. I've also seen enough AI plans organized around roles to know how easily the savings become the whole strategy.
Some of those cuts make sense. Others put capabilities at risk before anyone has asked what they're worth.
Here's what I'd look at before the next value creation review.

1. The AI business case is too often a headcount spreadsheet

This summer, the AI value creation lead at a tech-focused PE firm walked us through a software portfolio company with about 500 engineers. He estimated the annual AI opportunity at $10 million to $15 million.
Then he said: "Whether that's headcount reduction or improved throughput and therefore better products. We can discuss where that productivity goes."
That was the most useful part of the conversation. He had left the allocation decision open. The productivity was an opportunity, and someone still had to decide what to do with it.
In too many plans, that decision has already been made. Headcount is the default because it's easy to put into an EBITDA bridge. A product you haven't shipped yet is harder to underwrite.
Here's what the other choice can look like. For a logistics company, two Limestone engineers worked with the client's product and technical team to build a new Freight Audit & Pay service, replacing an older solution and integrating with the existing system. A core rebuild had been estimated at seven to eight months. The two engineers merged 122 pull requests in the first 90 days and had the service in production by month six, and the client's CTO called them the top-performing team in the organization.
The question on that project was what the business could ship, and how soon. The productivity went into the roadmap.
This isn't only our experience. When Box surveyed 1,640 IT decision-makers in the US, UK, France and Japan this spring, 79% of the organizations furthest along with AI expected their headcount to grow over the next three years, compared with 30% of those just starting. Ramp's economists, working from spending and workforce data on more than 21,000 US firms, found that heavy AI adopters grew headcount by about 10% in the two years after they adopted. The companies furthest along aren't using the capacity to shrink. They're using it to take on work they couldn't afford before.
That distinction matters more as the return requirements get harder. Bain's February 2026 Global Private Equity Report estimated that a typical deal now needs roughly 10% to 12% annual EBITDA growth to reach a 2.5x return over five years, compared with about 5% under the conditions of the 2010s. Separately, it reported that actual buyout holding periods at exit were around seven years.
A lasting cost reduction can improve earnings every year. But the same reduction doesn't create a new source of growth every year. Once you've taken it, you still need to explain where the next year's improvement comes from.
This is why I want operating teams to change how they assess AI projects. Ask only how many roles a project removes, and you'll tend to fund the most obvious savings. Many of your competitors will find those too, using much the same tools.
Ask what customers could get sooner, what keeps them from buying more, or which part of the roadmap has been waiting for capacity, and different projects become worth funding.
Some will take longer to prove. They can also give the business something more durable than a lower payroll: a reason for customers to choose it.
https://pbs.twimg.com/media/HUH8UV7XQAAVQPt.png

2. Four blind spots in a plan built around savings

A savings model can be perfectly accurate and still leave out the consequences that matter most. These are the four gaps I'd check.

Gap 1: the multiple

More EBITDA helps. The price a buyer will pay for that EBITDA matters just as much.
Take a business with $5 million of EBITDA valued at 10x. That's $50 million of enterprise value. A sustainable $1 million cost reduction lifts EBITDA to $6 million and, at the same multiple, value to $60 million.
Now suppose the buyer pays 9x because they're concerned that AI will erode demand for the product. The business is worth $54 million. Most of the expected increase in value has disappeared, even though the savings were delivered.
Look at it the way the fund does, through the equity. Say the company carries $30 million of debt. At 10x, the cost reduction takes the equity from $20 million to $30 million, a 50% gain. At 9x, the equity is worth $24 million, a 20% gain. One turn of multiple has erased 60% of what the savings did for the equity.
That's an illustration, but the concern behind it is already visible. Bain's midyear report described a fall of nearly 30% in public software valuations in February and an approximately 8% decline in PE software portfolio marks through the end of March. PitchBook's Q2 report showed US private equity software deal value down 65.7% year on year, as investors worked out how far and how fast AI would disrupt legacy firms.
This isn't happening uniformly. In the IBBA and M&A Source survey reported in June, 67% of advisors working on deals below $50 million said AI had made no material difference to valuations. The concern is more immediate in software and other businesses whose products AI could substitute for.
Even there, an AI story doesn't automatically command a premium. In Accordion's survey of 150 AI, data and technology operating partners, 44% said buyers were asking about AI-enabled finance in diligence without a premium showing up in price. Only 9% had seen a demonstrable premium for that finance capability in a completed transaction.
If AI disruption is part of the valuation debate, expect it in diligence too. The question in the data room shifts from "how much did AI save?" to "how much of this product could a competitor rebuild with AI, and how fast?"
A savings program has no answer to that. Increasingly, neither does proprietary data on its own. For the last ten years, a lot of companies competed on having more data than everyone else. That advantage gets weaker when AI makes analysis cheap for everyone.
What a competitor can't prompt its way into is what the business has noticed and built around: the exceptions, the workarounds and the workflows customers depend on that never made it into anyone's dataset. Proprietary observation is going to beat proprietary data.
The question is what AI does to the company's ability to earn money. A list of savings answers only part of it.

Gap 2: the people who carry the process

Two people in finance operations can look like a straightforward savings opportunity. Spend a week beside them and you may discover that they handle every exception the process documentation leaves out.
They know which customer needs a different approval, why a reconciliation sometimes fails and when the system's answer should be challenged. That knowledge rarely appears beside their salaries in the workforce plan.
It's also where the best AI opportunities are hiding. They're rarely in the dashboard. They're in the workaround: the spreadsheet nobody wants to admit the process runs on, the manual reconciliation at month end, the sticky note on someone's monitor, the phone call to the one customer who always needs a different approval. None of it has become data yet, which is exactly why a competitor can't prompt its way into it.
That's why I think fieldwork comes back. The operating teams that get the most out of AI won't be the ones asking a model to summarize the portfolio. They'll be the ones who put people beside the work long enough to see where it quietly breaks. Box's 2026 survey points the same way: the most common new AI role companies are hiring for is the agent operator, which Box describes as someone embedded in a business process to make agents work inside it.
Cut the two people in finance operations first and you lose more than their salaries. You lose the only people who could have shown you where to point the AI.
Gartner expects many of these cuts to be reversed. It has forecast that, by 2027, half of companies cutting customer service staff because of AI will rehire people for similar functions under different titles.
Before removing the people, find out whether the work has actually gone away.
If the exceptions live in two people's heads, their departure is a knowledge transfer problem you need to solve before it becomes a service problem.

Gap 3: release quality

More code, written faster, is only an improvement if it works. AI tools let a team produce far more change than it could before. The review, testing and release processes around that team usually weren't built for the volume.
If the plan counts output and ignores what breaks, the productivity gain shows up in the bridge while the cost shows up somewhere else: in incidents, in rework, in a customer who quietly stops trusting the product.
That's why the number I'd point to on the logistics project isn't the 122 pull requests. It's that 104 of them went through with fewer than five reviewer comments. Speed is easy to claim. Speed that holds up in review is what the business can build on.
Before you take credit for throughput, ask whether quality held.

Gap 4: time

Bain puts buyout holding periods at exit at around seven years. Its analysis of 15 years of buyout vintages found that IRR starts to stagnate around year seven and declines after that.
Most AI programs are built one company at a time: a new vendor, a new pilot and a new learning curve in each portco. Nothing is reused, so the fifth company takes as long as the first.
The funds that get this right treat a working AI pattern as a portfolio asset. Prove it in one company, then reuse it in the next three. A headcount reduction can't be redeployed across the portfolio. A proven way of working can, and that is how AI starts to shorten a hold instead of only padding the bridge.

Three questions for your next value creation review

1. For every AI initiative, where does the productivity go: margin, capacity for the roadmap or a new source of revenue? Who owns that decision, and what number proves it?
2. Which of the three return levers does it move: earnings, the exit multiple or the holding period?
3. If a buyer asked at exit what AI built in this business that a competitor can't copy, what would we show them?
   Set the savings aside long enough to answer those. Then bring them back. They'll still be there, and they'll finally be part of a strategy instead of the whole of one.
   P. S. If you want to implement AI into your portco at startup speed, book a discovery call with me.