The value creation plan has a data problem

The return model for private equity investing in technology has changed. The value creation plan now has to carry more of the return. Bain’s 2026 Global Private Equity Report puts a number to it. In the 2010s, a typical deal could generate a 2.5x return over five years with about 5% annual EBITDA growth, helped by multiple expansion. Today, with borrowing costs around 8% to 9% and leverage at 30% to 40%, generating that same return requires annual EBITDA growth of roughly 10% to 12%. As Bain puts it, “12 is the new 5.”

We’re all watching the same movie. Private Equity Info reports that the median hold has increased from five to six years over the last four years. In technology and business services, it has stretched from about 4.5 years to roughly six. Evercore counted about $121 billion in secondaries volume during the first half of 2026. GP-led deals accounted for roughly $65 billion, up 35%, while single-asset continuation vehicles reached $34 billion, up 88%.

None of this is irrational. Continuation vehicles are the topic du jour in PE because firms need options in a market that does not support exits at the values embedded in the original underwriting. But financial engineering can solve a timing problem. It cannot solve the return problem. Time isn’t a return. A longer hold increases the value that must be created to preserve the target IRR.

So the operating plan now has to carry more weight. That is no surprise to anyone who has rebuilt a VCP in the last two years. What has changed is how much of the underwriting now rests on customer economics: pricing, retention, expansion and cost to serve. Those are the levers that determine whether value compounds or the return case begins to unravel.

AI pays off in diligence, not yet in the portfolio

AI has real potential to improve those economics, and firms are funding it accordingly. The early results still leave operators asking when the value will show up in portfolio performance. In the Bain and StepStone GP Outlook, GPs reported their strongest generative AI returns in deal sourcing and due diligence. Inside portfolio companies, the impact remains harder to measure.

This is not a model problem. It is a context problem: the quality of what the model can see and reason over. Diligence is bounded. Someone curated the data room. The files are clean and structured, the questions are defined, and there is a deadline. A portfolio company is not bounded. It runs on a CRM nobody can fully trust, a support desk nobody migrated, call recordings in one tool, Slack in another, and a few hundred inboxes where the actual state of each customer relationship lives. The buy-and-build thesis compounds the problem. Every add-on brings another set of systems and another version of the same customer.

The board deck and the customer tell different stories

I’ve seen versions of the same pattern across PE-backed software companies. The board deck showed the largest accounts as green and supported an aggressive expansion target. Customer calls and support threads told a different story: two executive sponsors had left, one implementation had stalled, and a major customer planned to consolidate vendors. Management cut the expansion forecast, reforecasted the quarter yet again, and shifted strategy and resources from growth to retention. Thrash. Rinse. Repeat.

At the account level, it looks like this. The renewal forecast is green because the rep marked it green, while the last three email threads show that the champion left in August. Two add-ons sell into the same logo, and neither knows about the other. An account-summary agent treats CRM notes as ground truth, even though those notes are filtered through the judgment and incentives of whoever entered them.

The model answers anyway. It summarizes, scores and forecasts with the same confidence whether the inputs agree or not. Connecting it to more fragmented sources does not fix that. More retrieval against unresolved systems does not increase truth density. It increases ambiguity. The model simply gets more conflicting versions of the same account.

We have watched this pattern across enterprise AI. Most enterprises increased inference capacity before solving inference integrity. Put plainly, they gave the model more to process before establishing whether the underlying data could be trusted. For a portfolio company, that mistake is more expensive because the hold clock is running.

The view has to exist before AI scales

This boils down to a data problem. The revenue lines in the VCP assume the company can see customer behavior clearly enough to improve performance. After close, most can’t. The constraint is not access to AI. It is whether the company has enough reliable operating context for AI, or anyone else, to make decisions based on what is happening in this business rather than applying a playbook that worked in the past.

Closing that gap starts with execution, not technology. An operating team cannot deliver against the retention, pricing or expansion assumptions in the VCP without a reliable view of each account: who the customer is across every system, where the relationship actually stands, and what the customer said rather than what someone chose to enter in the CRM. That view has to exist before the AI program scales, because every agent, forecast and board dashboard downstream inherits the gaps and contradictions beneath it.

The same work pays twice

Get the operating context right, and the same work pays twice: once during the hold and again at exit.

During the hold, it grounds revenue decisions in observed customer language and behavior rather than internal assumptions. Churn risk can surface two quarters out instead of two weeks out. Expansion signals come from customer activity rather than a rep’s optimism. Pricing decisions reflect evidence of value, and cost to serve can be measured by account rather than estimated by segment. Each is a VCP line that moves from an assumption to an operating metric supported by evidence.

At exit, the same evidence becomes proof. For software and services assets, the next buyer’s hardest question is whether revenue holds as seat counts compress, services are automated and customers consolidate vendors. Retention curves and NRR show what happened, but they do not show why it happened or what is likely to happen next. A seller who can connect the retention number to the support, usage and commercial history behind it can demonstrate revenue durability rather than ask the buyer to underwrite it. Diligence AI works because the data room is clean. So make the company look like the data room. The context built to run the business becomes the evidence used to sell it.

Multiple expansion used to reduce the operating improvement required to hit the target return. It can no longer be assumed to make up the difference. The return case now depends on operating performance that can be supported by evidence, and the next buyer will test that evidence in diligence. The firms that close this gap will be selling a company that knows what it knows.

If you’re working through this in a portfolio company, during the hold or ahead of an exit, let’s talk. A useful place to start is simple: the 25 largest accounts by ARR, what the board deck says about each one, and what those customers have actually said in the last quarter.