Private Equity Is Building AI One House at a Time

A partner at a New York–based industrials fund told me last month why his AI plan had stalled. His words, close to verbatim:

We’ve been looking at a lot of AI providers, but we haven’t gone ahead with any of them. Hard to say which ones can actually deliver a tiny amount of real business value.

He was not behind. He had read the memos, sat through the demos, and built a shortlist. He was stuck one step before the place everyone assumes the difficulty begins, at the question of where the first dollar goes. Every provider in his inbox promised the same outcome in the same language, and none of them could tell him which of his portfolio companies to start with, or what the work would move on his return.

That conversation was not unusual. In the diligence sessions, operating-partner calls, and dinners I have had with industrials sponsors over the past year, the pattern holds. The plans exist. The shovel is not in the ground.

AI is a tool, not a strategy

Before the question of where to start, there is a question about the artifact. Nearly every fund I speak with is working on an AI strategy, and the document is part of what is holding them up. In 2000, serious companies wrote internet strategies. The ones that came out of that decade ahead were not the ones with the better strategy document. They were the ones that rewired something specific: how an order got taken, how distribution worked, how a customer got an answer. The companies that had an internet strategy bought a brochure website.

AI is a general-purpose tool, and general-purpose tools do not keep their own strategy for long. What has changed is not the shape of a plan but the boundary of what can be automated. Work that resisted software for thirty years, because it required reading an unstructured document, interpreting an exception, or exercising judgment inside a narrow domain, is now addressable. That expansion is genuinely large, and it is the only part of this that is new. It does not change how a business gets better, which is still one problem at a time.

This is why “we are doing AI now” reliably produces chatbots. Handing every employee a chat window is the 2026 version of the brochure site: visible, cheap, easy to announce at a board meeting, and detached from every number in the operating plan. No portfolio company’s margin has ever moved because its staff can ask a model questions. Meanwhile the quote that went out four days late still goes out four days late.

The artifact worth having is duller and more useful: a ranked list of the problems in each business, what each one is worth, and which of them have become solvable in the last two years. Work that list and AI stops being a strategy and becomes what it actually is, an implementation detail inside the solution to a specific problem. The rest of this piece is about that list. Which problem to take first, how to get a portfolio company CEO to want it solved, and how to avoid paying for the same solution six times.

The same fund can produce either of these. Only one of them has a number attached to it.

Selection, not deployment, is where portfolios stall

The common assumption is that AI in a portfolio fails on execution: the integration breaks, the model underperforms, adoption never comes. In industrials, most sponsors never get far enough to find out. They stall earlier, on selection. Which company, which workflow, which provider, and in what order.

Part of this is a market problem. Every provider in the category has learned to describe itself as revolutionary AI solution, which pushes the entire burden of judgment onto the sponsor. When forty vendors promise AI disruption, the shortlist stops being a shortlist and becomes noise.

The data suggests the stall is widespread. In its analysis of 471 private equity–backed companies, McKinsey classified each company on a four-level AI maturity ladder: opportunistic adoption, operating-model enhancement, AI embedded in products and services, and business building. Of the 471, 137 sat at level one and 213 at level two. Roughly three-quarters of the sample had not moved past productivity tooling, and only 21 companies, fewer than five percent, had reached level four. McKinsey’s own read is blunter than most sponsors expect: many portfolio companies are earlier in the journey than their owners assume.

The better funds are responding by putting a name on it. One specialty-materials sponsor I work with created a fund-level VP of data analytics and AI, whose mandate is to establish what good looks like across the portfolio, absorb the vendor flow, and find structures that repeat from one company to the next. Note what that role is not. It is not the owner of an AI strategy. It is the owner of the problem inventory and the reuse mechanics: which problems recur across the portfolio, what they are worth, and what has to be built once so that no company builds it again. That is the right thing to centralize. But a mandate is not a shovel, and someone still has to choose the first cut.

Choosing it is more mechanical than it looks. Decompose value in each business into its drivers, revenue and cost, then keep going until each branch lands on a process that somebody owns. Cost the process the boring way: how often it runs, how many hours it takes, what an hour there actually costs. Do the same for the pools that never appear as a line item, the cost of errors and the revenue lost to being slow, which in industrials is usually larger than the labor sitting inside the process. Then map each driver to the applications that could move it, and discard any candidate that cannot be traced back to one. What falls out is not a strategy. It is a ranked list of problems with a number next to each, which is the only artifact that makes a vendor conversation short.

Decompose value into drivers, cost the processes underneath, then map each driver to the applications that could move it.

This is a workshop, not a quarter. With the fund’s own people in the room, the ones who already know where each company loses money, a day or two is usually enough to produce a first ranked list across five to ten companies. Intuition about where the money leaks is better than most investment teams give it credit for. What the exercise adds is a number, an owner, and an order.

A portfolio is a neighborhood, not a street of custom homes

Walk into an industrials portfolio company and you find the same job done five ways under one roof. One mid-market platform I work with runs four brands on seven separate ERP systems while its sponsor adds three to four acquisitions a year. Quoting alone happens a dozen different ways depending on which legacy business takes the call. Its commercial lead described the data architecture as a car that looks fine until you open the hood and find hamsters running inside. Widen the lens to the whole fund, where that fragmentation repeats inside every company, and the count runs into the hundreds of distinct workflow variants, each one looking like it needs its own custom build.

Faced with that, funds do the reasonable thing. They pick one company, rebuild its processes end to end, prove the case in a board deck, and then move to the next one. Finish the first house down to the doorknobs before breaking ground on the second.

The math of a hold period does not forgive that sequence. Five to seven years is not enough time to build twenty houses one after another. The twentieth is still framed out when the fund already needs to start finding buyers for it. It’s also incredibly slow compared to the speed at which new AI capabilities become available.

No developer builds a subdivision that way. They pour the foundations and frame the lower floors across the entire block in one pass, because the structure under every house is the same. The finishes that make each home specific come later, and they go quickly, because the load-bearing work already stands.

A portfolio has that same shared structure. Quoting is one clear example that we see. A cutting-tool manufacturer, a robotics distributor, and a contract shop each quote differently in the particulars and identically in the bones: someone reads a request, prices it against cost and capacity, and returns an answer before the customer’s patience runs out. Order entry has the same shape almost everywhere. So do the service requests that stack up in every one of these businesses on a Monday morning.

That is the move. Standardize each company’s version of the workflow into one clean pattern, group the companies that share that pattern, and build for them together. A hundred bespoke projects collapse into a handful you build once and reuse.

Hundreds of workflow variants collapse into a handful of patterns built once and reused.

The compounding starts at close

Private equity will not rip out an ERP, and they shouldn’t. The group commercial officer of one PE-owned test-and-measurement platform, a man who has rolled up twenty companies into a single operating model, put it plainly:

ERP conversions cost a lot. They take a lot of time. I have never seen one that didn’t disrupt the business. Private equity has no tolerance for it.

AI sits at the other end of the risk scale, which is the entire reason it belongs early in a hold. A working quoting or order-entry workflow can stand up in weeks, wired into the systems the company already runs, without asking anyone to migrate their record of truth. Few investments a fund can make are that fast.

Speed matters because the clock on value creation starts the day the deal closes, and AI compounds against that clock better than almost anything else in the operating plan. Deploy in year one and the gains build on themselves for the length of the hold, while the data the workflow generates becomes an asset in its own right. Wait until year four and two things have gone wrong at once: most of the compounding is already forfeited, and the company’s habits have hardened into processes that resist standardization at every turn.

Deploying at entry compounds for the length of the hold. Waiting forfeits most of the curve.

There is a second-order benefit that rarely makes the investment committee memo. A sponsor that has deployed the same workflow across four companies knows, in advance of the next deal, what the workflow is worth and how long it takes to stand up. That becomes an underwriting input, and eventually an edge in diligence.

Where the multiple actually moves

AI touches three drivers of a fund’s return: the earnings, the multiple those earnings command, and the time it takes to reach an exit. Operational cleanup handles the first, expanding EBITDA by running one process where there used to be five. The larger prize is the second, and it does not come from the use cases most funds start with.

McKinsey’s ladder makes the pattern visible. Companies at level one traded at a median revenue multiple of 13 times and companies at level two at 14 times, a step that is effectively flat. The multiple climbs at level three, to 20 times, and again at level four, to 31 times: a 133 percent gap from the bottom of the ladder to the top. Back-office productivity, the most common way funds use AI, moved neither the multiple nor the exit.

Median revenue multiple by depth of AI adoption, across 471 PE-backed companies. The step from level one to level two is flat.

Read the ladder as a description of where companies ended up. No business reached level three by announcing a transformation. They got there by solving one specific problem after another, in an order that made sense to the people running the place, until enough of the work had changed that the product was different.

The obvious objection from an industrials sponsor is that this is a software phenomenon. It is not. Splitting the sample, software companies trade at 24 times at level three and 33 times at level four, while non-software companies trade at 15 times and 22 times. The absolute multiples are lower, as anyone underwriting a machine builder would expect, but the step change is nearly 50 percent either way. Being a manufacturer is not an exemption from the pattern; it is a discount on the starting point.

Revenue efficiency moves in the same direction and shows where the earnings expansion sits. Median revenue per employee rises about 19 percent from level one to level two, the range most operating-model work lives in. From level three to level four it rises from $118,000 to $180,000, a 52 percent jump. And over 2018 to 2022, digital and AI leaders among PE-backed companies delivered total shareholder return CAGR 2.3 times higher than their less advanced peers in energy and materials.

Revenue per employee rises 52 percent between the top two levels of the ladder.

The mechanism is in who shows up to buy. A cost-minded buyer pays for cleaner earnings at the same multiple: a bigger number against the same figure. A growth-minded buyer pays a higher multiple, because a standardized operation is one where growth is easy to underwrite. New volume flows through without the handoffs breaking, and each added dollar of revenue keeps more of itself. Because the transformation was built to repeat, the fund also reaches that exit sooner, which turns the same value into a higher return by arriving earlier. A cleaner business does not sell for less. It sells to a better buyer.

Leadership prices effort before it prices return

None of this happens unless the operating partner and the portfolio company CEO both want it. That sentence is easy to write and is where most portfolio AI programs actually die. The usual failure is not resistance. It is a sponsor arriving with the wrong argument.

The instinct is to bring an ROI case. Every CEO of a $200 million industrial business has seen ROI cases, and has learned to discount them, because the number in the deck was never the thing they were worried about. What they are pricing is effort and feasibility: how much of their team’s week this consumes, what happens to the quarter if it goes wrong, and who they will be standing next to when it does. The commercial officer I quoted earlier explained why he insisted on having a technical partner in the room:

I know enough to be a little dangerous. What I don’t want is for it to explode upon implementation. When you have done enough with technology, you have had bad implementation partners, you have had things that just don’t work, you have had budget slips and timeline slips.

That is the real objection, and it is a reasonable one. Four things move it.

1. Speak in margin, not in AI. Most industrial executives have no interest in an AI initiative and considerable interest in gross margin. At SUPPLYCO we state the case as margin lift: basis points of gross margin on quoted volume, plus the win-rate delta on quotes that currently go out too late to win. Both are numbers the CEO already reports to the board, which means the initiative is measured on the existing scoreboard rather than a new one.
2. Bound the effort explicitly. Name the systems the deployment touches, name the single owner inside the company, name the hours per week their team owes it, and name the kill point. An initiative that needs twenty hours a week from the CEO’s direct reports will not survive contact with the operating calendar. One that needs two hours a week from one named owner, with a decision gate at ninety days, will.
3. Respect the attention budget. On one platform I worked with, the CFO had arrived in December, the CEO on January 1, and the commercial lead four months earlier. Every hour of executive attention was already committed to standing the leadership team up. Capital was not the constraint. Attention was. Sponsors who read that correctly scope the first build to fit inside the attention that actually exists, rather than the attention the value creation plan assumed.

There is one more thing that makes the conversation easier. Start with one data model and one workflow. A single clean model of customers, products, and transactions, feeding a single workflow that the company already knows is broken. Most industrial platforms need that cleanup regardless: on the platform above, one product family had spawned a new part number every time a fixture changed, so the ERP showed a sprawl of SKUs where the business had a handful of products. Do that work once and it carries every workflow that follows. That is the foundation of the house, and it is also the smallest thing a CEO can say yes to.

The first build has to be engineered for reuse

Every fund wants one success before it commits the portfolio. Having spent enough years at McKinsey to know that rule is not going away. The danger is in how the first case gets built.

A pilot that lives inside one company’s cloud account, wired to one company’s IT stack, tuned by one company’s admins, proves only that the thing can work once, but it doesn’t translate. When the second company asks for the same result, the team starts over, and the fund is back to building houses one at a time with extra paperwork. If every portfolio company has to stand up its own AWS account and its own integration layer before anything runs, the rollout is already too slow for the hold period.

McKinsey frames the underlying choice as rent versus buy: centralize platforms and expertise at the fund level, or embed dedicated teams inside each portfolio company. Their finding is that leading firms do both, centralizing infrastructure for the use cases that recur and reserving in-house teams for genuine product work. In practice, for a workflow like quoting, that means the infrastructure hangs off the systems each company already runs, and no portfolio company has to build a stack to find out whether the thing works.

Then measure the result in a number the investment committee already trusts, and keep the pattern portable enough that the second and third deployments are configuration work instead of construction. Build the first house so its blueprint pours the next ten.

Reusable agents, clustered by workflow

The unit of a portfolio rollout is not the company. It is the workflow cluster: the set of companies whose version of a workflow shares the same shape. Cluster on the shape, not on the vertical. A hydraulics maker and a test-instrument brand can sit in the same quoting cluster while two businesses in the same end market do not, because what determines reuse is how the work moves, not what the company sells.

Each cluster gets one agent, built once. What differs by company is configuration: which ERP holds the transaction history, which pricing rules and margin floors apply, which motion the sales team runs, who approves an exception. Six companies can be covered by three agents, and the build is paid for once per cluster rather than once per company.

Three reusable agents cover the workflow needs of six portfolio companies.

The sequence that gets a fund there is the same one McKinsey prescribes for reshaping an industrial go-to-market model: diagnose, design, then execute with discipline.

1. Diagnose. Run the value-driver exercise from Exhibit 2 across the portfolio. For each company, size the problem in margin terms and note the shape of the process rather than the software behind it. Then rank the companies and group the ones whose shape matches. Five to ten companies is enough to reveal where the clusters are.
2. Design. Build the agent once, on the highest-value cluster, and prove it on a single company inside a quarter. Instrument the metric before anything ships, so the result is not a matter of opinion. This is also where you decide what stays central: the data model, the integration layer, the evaluation harness.
3. Execute. Stamp the pattern across the rest of the cluster, where each additional deployment is configuration. Track the same metric in every company so the second board conversation is a comparison rather than a pitch, and route each add-on acquisition into an existing cluster on arrival.

That last point matters more in industrials than anywhere else. A platform absorbing three or four add-ons a year does not need a new AI project each time. It needs a cluster the new company can be configured into during integration, in the window when a new owner is already changing how the business runs.

The workflow to start with is usually quoting. In McKinsey’s work on industrial channels, one OEM traced a low quote-to-sale conversion rate directly to a slow quoting process; fixing it improved conversion by more than 10 percent. Quoting sits at the point where margin is decided, it is measured everywhere, and it is one of the few workflows whose shape survives translation from one industrial business to the next.

What this looks like in practice

At SUPPLYCO, we find the workflow that repeats across a portfolio with only cosmetic differences, usually quoting, and standardize it into one pattern. The infrastructure attaches to the systems each company already runs, so no portfolio company stands up its own cloud to find out whether the thing works. We prove it on the first company inside a quarter, state the result as margin lift on volume the CFO already reports, and then configure the same pattern into every company that shares the shape, including the ones acquired after the first build.

The test of whether a rollout is designed correctly is simple. If the second deployment costs what the first one did, the fund is pouring a fresh foundation under every house, and it is back to one at a time.

Returns are compressing, the exit backlog is real, and the valuations of the last cycle ran too hot. The next cycle belongs to the funds that go back to underwriting the fundamentals of a business, name the specific problems that hold those fundamentals down, and then use whatever tool solves them fastest across a whole portfolio at once rather than one house at a time. Today that tool happens to be AI. In five years it will be something else, and the funds that built the habit of working the problem list will barely notice the difference.

If you run an industrials portfolio above $500M in revenue and you are tired of demos that cannot name the value they create, that is the conversation I want to have. Book time with us.