Every lever a PE firm used to pull on a consumer business has got stiffer over the last three years or so. I don’t think that’s a controversial thing to say. What I find more interesting is what you’re left holding once you’ve worked through them all.
Financial engineering worked when leverage was cheap and multiples were expanding, and neither of those things is true now. Buy-and-build runs straight into a consumer who’s being careful with discretionary spend… in the UK, discretionary spending hit a three-year low in Q1 2026, with nearly half of adults planning to cut back on non-essentials. And cost optimisation has already been wrung out. Hospitality is running on margins of 3-5%, with 82% of operators having already raised prices and cut staffing hours in response to wage legislation and food inflation. There isn’t a lot left to take out that doesn’t start showing up in the customer experience.
Meanwhile the hold clock keeps getting louder. More than 63% of active North American portfolio companies have been held for over four years, and LP distributions have roughly halved from their historical average, down from around 25% of NAV during 2013-2021 to just 12% across 2022-2024. The exit that was supposed to happen hasn’t happened. All that pressure to make something of the hold period has nowhere to go except operations.
So it tracks that 72% of GPs are now citing operational improvements as their primary value creation lever. And when you dig into what operational improvement actually means inside a consumer business, a lot of it lands on the data that tells you what’s working and what’s about to go wrong.
Put more bluntly: a large part of the remaining upside is in understanding a customer base better than the people who sold it to you did, and then evidencing the headroom to the next buyer.
Meanwhile, the customer got harder
The portfolio companies are being squeezed from the other side too.
Brand loyalty is falling apart faster than it has in a decade. 38% of shoppers in 2025 said they were loyal to five or fewer brands, up from 22% in 2023. That’s a huge move in two years. On top of that, 60% said they’d switch brands if prices went up, which is awkward timing in a market where operators are being forced into price rises by wage legislation, food inflation and tariff pressure. You can see where that goes…
So: thin margins, a customer base that’s more promiscuous than it’s ever been, and a macro environment suppressing discretionary spend across every income group. Organic volume growth isn’t going to pick up the slack. And acquiring a new customer costs five to 25 times more than keeping one you’ve already got.
All of which points at the one thing most of these businesses are already sitting on and mostly ignoring.
The asset that’s already there
Every consumer business generates it, continuously and without trying: transaction records, visit frequency, basket composition, recency patterns, loyalty interactions. It just accumulates.
And in most cases it’s being analysed with tools built for a different era. Basic RFM tiers. Demographic segments. Boolean rules designed around assumptions about what customers look like rather than what they actually do. The signal is in there, it just hasn’t been mined, and in an environment where every other lever has been pulled to its limit, getting at that signal is probably the highest-ROI thing left on the table.
Why deep learning changes the shape of this
Standard segmentation treats customers as static categories. A “high-value” tier defined by spend above a threshold. A “lapsed” cohort defined by recency. Clean enough to explain in a board update, but built around assumptions that came from a human analyst deciding in advance what to go looking for.
Neural networks don’t work like that. They learn latent behavioural patterns straight from the data and surface clusters of customers nobody would have thought to look for, because those patterns don’t map to anything with a simple name. Less “high-spend and high-frequency”, more “this group of 4,000 people has a distinctive and consistent way of moving through your product that correlates strongly with lifetime value, and they’re more loyal than their spend tier suggests.” Good luck writing that as a WHERE clause.
The precision advantage shows up hardest at the edges. Every customer in a deep learning-derived cluster comes with a distance-to-centroid, which is a measure of how archetypal they are within their group. The customer at the centre is the cleanest expression of that behavioural pattern. The customer at the edge is still technically in the segment, but they’re already drifting. That drift is where churn prediction starts to earn its keep, as an early indicator that intervention is worth the money rather than a retrospective label slapped on someone who has already gone.
When you’re operating on 3-5% margins and every lost customer is expensive to replace, the gap between broad retention spend and targeted intervention is enormous. A 5% improvement in retention can lift profits by 25-95%. Treat that range as directional rather than gospel, but the direction is the whole point when the baseline margin is that thin.
The area operating partners should care about most
This next part is hard to see from inside any single portfolio company.
A firm that has eight or ten consumer brands e.g. across restaurants, retail and hospitality is sitting on a cross-portfolio data asset that no individual management team is able to get to or analyse on their own. But run consistent deep learning across that whole portfolio and things start to surface that are invisible at the single company level:
- Which behavioural segments show up in more than one brand? If a cluster at one company looks structurally similar to a cluster at another, that’s either a cross-sell opportunity or a signal that a retention playbook which worked in one place is worth testing in the other.
- Which companies have the cleanest, most loyal customer bases ahead of exit? And which ones need work before they’ll command a premium multiple?
- Where is churn risk concentrated across the portfolio, rather than inside whichever brand happens to be reporting on it this quarter?
That second one matters a LOT but loyalty ≠ cleanliness when it comes to data. Acquirers right now are awarding premium valuations to businesses that can show strong customer retention, a coherent customer base, and the evidence to support it all. Walking into an exit process with cross-portfolio benchmarks on retention quality, churn trajectory and lifetime value is a differentiated position, because it lets you put a number on the headroom instead of asserting it. Buyers will pay for that. The businesses that can show it clearly, derived from 1P data rather than from analyst intuition or rented insights, should outperform comparable assets at exit.
None of this is new, by the way
Google, Meta and Amazon made deep learning the intelligence behind their recommendation and ad systems roughly a decade ago. You can see the result in where media budgets have gone ever since. Those platforms have taken an ever-increasing share of global ad spend precisely because their models read customer behaviour at a granularity no human analyst and no rules engine can match. That shift wasn’t luck. It was what happens when you point neural networks at first-party behavioural data at scale.
The obvious move for consumer businesses is to do the same thing on their own data. Transaction histories, loyalty interactions, visit patterns… the raw material is sitting right there. What’s been missing is the tooling and the institutional weight to do something properly sophisticated with it, which is a big part of why this stayed a big-tech privilege for so long.
The catch is that most portfolio companies can’t close that gap alone. They don’t have the engineering headcount, the infrastructure or the data science team to build it from scratch, and they can’t justify the spend on a single-company business case. That’s where the PE owner has a structural role beyond providing capital: acting as the enabling layer that gives portfolio companies access to capabilities none of them could stand up on their own.
What this looks like in practice
At Neuralift this is the work we do for consumer businesses. Transaction data, event streams, loyalty interactions and purchase sequences go in through a properly structured input layer. Six or seven deep learning models then run iteratively until the latent space converges, dropping every customer into the segment they actually belong in based on real behaviour. What comes out is a set of stable, actionable segment IDs, each with a distance-to-centroid you can use to tune precision depending on what you’re doing with the segment.
For a single portfolio company that might mean identifying the 6,000 customers whose behaviour puts them at real churn risk in the next 90 days, and separating them from the 3,000 whose apparent inactivity is seasonal rather than structural. For a portfolio operations team it’s more likely to mean running consistent segmentation across several brands and building a shared intelligence layer nobody in the portfolio has had before.
This isn’t theoretical positioning either. We’ve already had PE firms introduce us directly into their portfolio companies, usually a warm intro from a partner who’s watched the model land in one place and wants the same lens applied elsewhere. I’d expect a lot more of that over the next 12 to 18 months as the operational-improvement pressure bites harder, and the funds that move early will be working from a much better picture than the ones that wait.
I’ve written about this before in the context of lookalike seed quality, and the argument holds in the PE context too. The data is already there. The urgent bit is getting more granular with it, and that’s not something you can do with human-only approaches inside businesses that are already resource-constrained. The analyst time isn’t there, the headcount isn’t there, and the pattern complexity is beyond what rules-based segmentation was ever going to surface anyway.
For a lot of PE consumer portfolios, that’s the only lever nobody has pulled yet.