How PE firms actually use AI
Beyond the deck: where AI creates real value across a portfolio, from diligence to the back office, and where it does not.
The useful question about private equity and AI is narrow: inside a company you already own, where does this actually change a line on the P&L. Almost everything else said on the subject is a slide, a logo wall of vendors, a bar chart with no baseline, a fund that calls itself AI-native and cannot tell you which number moved. I build and run an operating holding, so I read that question the way a buyer reads a P&L, not the way a keynote reads a trend.
The four words, because they are not the same thing
Most writing about AI in private equity is unreadable because four different activities share one word. They sit at different points in the deal, they are bought by different people, and only one of them changes what a company is worth.
- Sourcing
- Finding companies worth a call. A ranking problem over public and licensed data. It changes how many targets you see, not what any of them is worth.
- Diligence
- Reading what you have been sent. Contracts, accounts, tickets, call logs. It changes how fast and how completely you read, and occasionally what you find, but the thing being valued is unchanged.
- Value creation
- Changing how the company operates after you own it. The only one of the four that moves earnings, and therefore the only one that moves the price on exit.
- Reporting
- Getting numbers out of a portfolio company and into a format the fund can act on. It removes cost at the centre and changes nothing inside the business.
Where AI plugs into the deal cycle
A deal is not one moment. It is a cycle. You source targets, you diligence the one you like, you own and operate it, and one day some of them exit. AI shows up in three of those four stages and it runs none of them. Naming where it helps, and where it does not, is most of the honesty missing from the pitch.
Sourcing is the top of the funnel and the least glamorous place AI earns its keep. A longlist of a few thousand companies is too much for a person to read and too structured to ignore. A model can sort it: filter by the shape you buy (a revenue band, an owner near retirement, a boring service business that survives a soft patch), read hundreds of listings, and push the twenty worth a human hour to the top. The failure mode is trusting the sort. The list tells you where to look, not what is true, so every company that clears the filter still gets read by a person before anyone picks up the phone.
Reading everything, not the top of the pile
Diligence is a race against a data room and a clock. You get a pile of documents, a short window, and a decision worth years. AI does not make that decision. It compresses the reading.
Point a model at three years of contracts and it surfaces the auto-renewal clauses, the change-of-control triggers, the one supplier who quietly carries a large share of cost. Feed it the customer export and it clusters the revenue: how much is a single account, how much churned last year and came back, how much is genuinely recurring versus a column in a spreadsheet labelled recurring. This is not magic. It is the analyst work you always wanted done and never had the hours for.
- Data room
- The folder of documents a seller opens during diligence: contracts, accounts, customer lists, the paperwork behind the numbers.
- Value-creation plan
- The written list of moves meant to make a company you own worth more, next to pricing, procurement, and headcount.
- Operating partner
- The person who installs those moves inside the company, rather than advising on them from outside.
Why a human still signs
The value is coverage, not brilliance. A wrong summary in diligence is worse than no summary, because it hides the thing that kills the deal behind a confident sentence. So a person confirms every flag that matters. What changes is where you start. You walk into the room having read all of it, not the top of the pile, and you ask the seller the question that came from page two hundred.
A fund and an operator-holder use the same tool differently
Two buyers can run the identical model and get opposite value from it, because the thing they are optimising is not the same. A fund is on a clock. It buys, holds for a fixed stretch, and sells, so its incentive is a story that survives the next raise and lifts the exit multiple. An operator-holder keeps the company, so the only thing that counts is whether a number actually moved and stayed moved.
| A fund | An operator-holder | |
|---|---|---|
| Time horizon | A fixed hold, then a sale | No sale date |
| What AI is for | A story for the raise and the exit | A number inside a company you keep |
| How it lands | Bought as licenses, counted as progress | Installed as a workflow, run beside the old way |
| Who does the work | An external team parachutes in | The person who owns the result |
| What survives the deal | The slide | The process, carried to the next company |
The difference is not the software. It is who has to live with the result. When you are selling the company in a few years, "we deployed AI across the portfolio" is a line in a deck that nobody audits too hard. When you are keeping it, an install that quietly degrades support and costs you retention is your problem for as long as you own it, which is meant to be forever. That is why I would rather install one change that holds than announce ten that do not. If you run a firm and want the operating layer installed rather than described, that is the work I do in advisory.
| A fund with a clock | A holder with no clock | |
|---|---|---|
| What it buys | Speed through the deal cycle | Durability inside the company |
| Where it spends | Sourcing and diligence, before ownership | Operations, after ownership |
| Payback window | Inside the hold period, or it does not count | Whenever, because the company is not being sold |
| Who is trained | The deal team | The people answering the phone |
| What is left at exit | A process the next owner does not inherit | Nothing to hand over, because nobody leaves |
The value-creation plan: three places, not everywhere
Once you own the company, AI stops being a strategy and becomes a line in the value-creation plan. The honest version of that line is boring and specific. It names three places and leaves the rest alone.
Back office first
Because it is the least romantic and the most reliable. Invoice matching, order entry, the reconciliation someone does by hand every Friday, the report rebuilt from scratch every month. These are rules-heavy, high-volume, low-judgment tasks. Automating them does not cut the team so much as it stops good people from spending their week on data entry. Every business I run sits on a back office like this, and it is the base layer of the four-layer system everything else stands on.
Sales ops second
Not "AI closes deals." AI keeps the pipeline honest. It drafts the follow-up that otherwise never gets sent, scores which stale leads are worth a call, writes the first version of a quote so a rep spends minutes instead of an afternoon. The rep still owns the relationship. AI removes the reasons the relationship goes cold.
Support third
Most inbound is the same forty questions asked in a thousand ways. A well-built assistant handles the repeat volume and hands the genuinely hard ticket to a human with the history already summarised. Done carelessly it degrades the customer experience and you feel it in retention. Done well it shortens response time and frees the team for the calls that actually need a person.
AI is a line in the value-creation plan, not the plan.
Installation is a project, not a purchase
Here is the part the pitch deck skips. Someone has to install this without breaking the company, and that someone is the operating partner. The job is nothing like buying software.
A tool is a purchase. An installed change is a project. You map the actual workflow (not the one on the org chart), pick the one process where a win is measurable, run the new way beside the old until it is trusted, and train the people who have to live with it after you leave. The failure mode is a firm that buys licenses, declares transformation, and books nothing, because the work of changing how thirty people do their jobs never happened.
Knowing which process to touch first, and which to leave alone, is judgment you earn by running a business, not by reading about one. Before I install anything, the process has to clear a short bar:
- The work is repetitive and high volume, not a rare judgment call
- A win shows up in a number someone already tracks
- The old way can keep running beside the new one while it earns trust
- Someone on the team will still own it after I hand it back
- Getting it wrong is annoying, not dangerous to a customer
Why the same install is a bad idea once and a good idea eight times
Say installing a working follow-up layer in one company costs $60,000 of build and takes a quarter. In a single business turning $2M with a 12 percent margin, that is a quarter of the year's profit spent on one fix, and it has to work. Across eight companies the build is largely the same work, so the cost per company falls toward $15,000 to $20,000 while the return repeats eight times. Nothing about the technology changed between those two paragraphs. The only thing that changed is how many times you get to reuse the answer, which is the actual argument for owning more than one.
Where it does not help
An honest account has to name the ceiling. AI does not fix the hard parts of owning a company. It does not decide whether to move on from a founder who no longer fits the next stage. It does not sit across from a supplier and rebuild a relationship that broke over a late shipment. It does not read a room, hold a difficult manager accountable, or carry the weight of a decision that has no clean answer.
Those are the parts that actually determine whether a business improves, and they stay stubbornly human. Anyone selling AI as a substitute for judgment, trust, or accountability is selling the thing that does not exist. The framing that holds is narrow: AI removes the repetitive work so the people have more time for the work only people can do. A firm that forgets that ceiling automates the easy things and neglects the hard ones, which is exactly backwards.
Why a portfolio changes the math
The reason this matters more for a holding than for any single business is the portfolio. A standalone company builds an AI process once and uses it once. It pays full price for a single install. A holding builds it once and carries it across every company that shares the same back-office shape.
The math
The cost to build a working process is paid once. Its real value is the cost to build it once spread across the number of companies you can install it in. A fund with one company divides by one. A holding divides by many, and the same build gets cheaper with every acquisition.
Picture two companies you own that both send a few thousand invoices a month and both reconcile them by hand. Solve it properly in the first: map the workflow, build the match, run it beside the manual process until the team trusts it. In the second company the hard part is already done. The pattern arrives on day one of ownership, not month twelve, because the expensive work (figuring out what actually holds) is already paid for. The same is true for the support assistant, the lead scoring, the monthly report. This example is hypothetical, but the shape is the whole thesis: every company you own makes the next install cheaper and faster.
That is the compounding edge, and it is not a smarter model than anyone else can rent. It is a bench that installs the same proven change again and again. A senior operator who wants that leverage on a company of their own is who I work the deal with in buy.
AI does not run the company. It clears the repetition so the people who own the result can do the work only people can.
The short version
- Across the deal cycle, AI works in three stages (sourcing, diligence, ops) and runs none of them; the exit stays a human call.
- In diligence it compresses the reading so you walk in having covered everything; a person still confirms every flag that decides the deal.
- In an owned company it belongs in three places: back office, sales ops, support. Everywhere else is mostly a slide.
- The work is installation, not purchase, and it does nothing for judgment, relationships, or accountability.
- The portfolio is the edge: build the change once, carry it to the next company on day one.
None of this needs hype to be real. It is operations. You find the repetitive work, you install a proven change without breaking what works, you check whether a number moved, and you carry the pattern to the next company. That is what "private equity AI operations" should mean, and it is the work I do alongside firms that want it installed rather than announced.