AI in private equity · Part 3
Exposed, and not building
Outside software, the portfolio companies most exposed to AI rarely sell anything built on it. The ones that do are far more likely to have an owner that talks about AI. Under a sponsor that says nothing, it’s one company in ten.
In Part 1 we counted which private equity firms have an AI lead. In Part 2 we looked at who those leads are. This time we looked at what the firms own: how exposed their portfolio companies are to AI, and whether those companies have done anything about it.
Measuring exposure
We scored each portfolio company from 1 to 5 for how much of its core work AI can do today. The rubric asks two things: how much of the company’s time goes into reading, writing, deciding, scheduling and answering, rather than physical work, and whether that work is repetitive enough to follow patterns. A 5 is a business whose service is itself repetitive knowledge work, such as claims processing, revenue-cycle management or staffing. A 1 is almost all physical or asset-based, such as mining or agriculture. Every score is tied to a specific use case in that business, and “highly exposed” means a 4 or 5.
Across 6,122 PE-backed companies in our wider dataset, at least a third are highly exposed, whoever owns them.
It’s 33% at lower mid-market firms, 32% at mid-sized firms, 41% at large firms and 49% at mega-funds. The two largest tiers lean on samples weighted towards firms with an AI lead, but the point holds at every size: exposure is a property of the business, not of who owns it. A claims processor is just as exposed under a $300 million fund as under a $30 billion one.
The capacity to act on it is not spread the same way. By our October count, 39% of mega-funds have an AI lead, against 4% of lower mid-market firms. So in the lower mid-market, 84% of highly exposed portfolio companies are held by a firm with no AI lead. Not because those firms own unusually exposed companies, but because almost none of them has an AI lead.
Exposure says where AI could change a business. It doesn’t say whether the business has done anything about it. So we asked a narrower question. Of the companies whose core work AI can already do, which ones have put AI into what they sell? And does it matter who owns them?
Where exposure lands
For the rest of this post we use a narrower sample, researched in more depth: 2,022 PE-backed companies owned by 181 sponsors. For each one we also recorded whether it markets an AI product or feature.
Software is the obvious case. Almost every software company is highly exposed, and three in four of those already sell AI. Industrials, consumer and energy companies sit at the other end: their work is mostly physical, and almost none of them score highly.
What does a score look like in practice? Pick a sector to see a company at each end of the scale and one in the middle.
Software
1of 5
No software company in our data scores 1. The lowest, a 2, builds robotic telesurgery hardware.
3of 5
Video intercoms and access control for apartment buildings
The workSells and installs smartphone-connected door panels and access hardware across thousands of buildings, plus the app that runs them.
Where AI fitsSupport tickets and install scheduling. The value is in the hardware, the installs and the network of buildings.
5of 5
Freight audit and payment for global shippers
The workReads carrier invoices in dozens of formats, matches every line to contracts and shipment records, and chases the exceptions.
Where AI fitsExtraction and matching are the job itself. AI can take the first pass on almost every invoice, leaving people the disputes.
Financial services
1of 5
No financial services company in our data scores 1. The lowest, the 2s, include an armoured cash-in-transit and ATM service business.
3of 5
Wealth advisory for wealthy families
The workAdvisers manage money and plan finances for high-net-worth clients, relationship by relationship.
Where AI fitsDrafting financial plans, meeting summaries and client reports. The trust and the advice stay with the adviser.
5of 5
Chargeback disputes for merchants
The workWhen a shopper disputes a card payment, assembles the evidence and argues the merchant’s case with the bank.
Where AI fitsGathering the evidence and writing the response is the service itself, and it follows a pattern AI can learn.
Business services
1of 5
Vegetation management for electric utilities
The workCrews clear trees and brush from power lines, and restore lines after storms.
Where AI fitsRouting crews and planning work. Useful, but the work is chainsaws and bucket trucks.
3of 5
Elevator and escalator inspections
The workInspectors visit buildings to test elevators and escalators for safety, then report what needs fixing.
Where AI fitsDrafting inspection reports and ranking repairs. The inspection itself still needs someone in the machine room.
5of 5
Captioning and transcription
The workCaptions, transcribes and audio-describes video for universities, companies and broadcasters.
Where AI fitsSpeech-to-text and translation now do the first pass, leaving people to review and correct. That pass was most of the job.
Healthcare
1of 5
Medical office buildings
The workBuys and runs medical office buildings and leases the space to physician practices.
Where AI fitsSummarising leases and screening acquisitions. The business is property. Only two healthcare companies in our data score 1, and both are landlords.
3of 5
Hospice and palliative care
The workNurses, aides and chaplains caring for patients at home, from dozens of local offices.
Where AI fitsVisit notes, recertification paperwork and billing. The care itself happens in person and stays that way.
5of 5
Denied-claim appeals for hospitals
The workClinicians and coders read patient charts, pull out the evidence for a denied claim and write the appeal letter.
Where AI fitsReading a chart and drafting a structured argument from it is what current models do well. A first draft of every appeal is within reach.
Energy
1of 5
Natural gas gathering and processing
The workPipelines and processing plants that move gas from the wellhead to market.
Where AI fitsPredictive maintenance and spotting flow anomalies. Useful, but the business is steel in the ground.
3of 5
Gas compression parts distributor
The workStocks tens of thousands of compressor parts and sells them to field operators.
Where AI fitsQuoting, cross-referencing part numbers and order entry. The warehouse, the trucks and the field service stay.
5of 5
Engineering staffing for energy companies
The workFinds engineers and technology contractors and places them in roles at energy companies.
Where AI fitsSourcing, screening and matching candidates to roles, which is most of the service. It is the only energy company in our data to score 5.
The interesting part is the middle. Financial services companies are nearly as exposed as software ones: 88% score 4 or 5. Yet only 21% of those sell an AI product. In IT and business services it’s 42%. In healthcare, 42% of companies are highly exposed and 17% of those have an AI product. These are businesses built on reading, checking, deciding and answering, which is the work AI is best at, and most of them haven’t turned that into something they sell.
The silent-owner gap
Then we split those highly exposed non-software companies by what their owner says about AI. In Part 1 we sorted every sponsor into one of four stances: a named AI lead, a described AI initiative, a mention of AI in its thesis or materials, or nothing at all. Of the 181 sponsors here, 73 say nothing.
Under a sponsor that talks about AI, 34% of these companies sell an AI product. Under a silent sponsor, it’s 10%. The gap shows up in every sector with enough companies to check: 28% against 10% in financial services, 19% against 9% in healthcare, and 50% against 19% in IT and business services.
Silent sponsors don’t simply own different companies. They rarely focus on software, and they skew smaller. So we held sector, exposure score, company size and age, and the sponsor’s size and software focus constant. The gap narrows but stays wide: 32% against 14%. A highly exposed company under a silent owner has about a quarter of the odds of selling AI. The result holds in each of the three batches we researched, and dropping any single sponsor doesn’t move it.
Software companies show the same lean, more weakly: 76% sell AI under a sponsor that talks about it, against 58% under a silent one. That gap isn’t statistically reliable once size and age are held constant.
Saying something isn’t a strategy
It would be neat if the gap grew step by step with commitment, from a mention to an initiative to a named lead. It doesn’t. Highly exposed companies owned by firms that only mention AI sell it most often, at 45%. Firms with a named AI lead sit at 25%, and firms with a described initiative at 16%. CHECK: the mention group is concentrated in a handful of tech-leaning sponsors (Symphony Technology Group, Achieve, Recognize); decide whether to name or footnote
So the honest reading isn’t that hiring an AI lead makes portfolio companies build AI products. It’s that silence is a signal. A sponsor that has nothing to say about AI in public is unlikely to be pushing its most exposed companies to do anything with it either.
The arrow can also run the other way. A firm that already owns companies selling AI has a reason to talk about it. Our data can’t tell those two stories apart. What it can say is where to look: highly exposed companies at sponsors that say nothing. In our sample, that’s 126 companies, and 113 of them have no AI product.
What this means
For a company in financial services, healthcare or business services, high exposure cuts both ways. Competitors that put AI into the service will be faster and cheaper. A company that does it first can sell that. Either way the clock is the same; what differs is whether anyone is watching it.
For a sponsor, the useful question isn’t what it says about AI. It’s which of its companies are highly exposed and have no answer yet, and who is working on that list. CHECK: tie to a Talas CTA
Get the next post by email.
New research on AI in private equity and the companies it owns, when we publish it.
How we did this
The companies. We mapped the current portfolios of 181 PE sponsors from our dataset, built from SEC Form ADV filings, excluding mega-funds. The first set of sponsors was chosen to compare firms with and without AI leads; later batches were drawn at random across size tiers. We kept PE-backed operating companies the sponsor still holds: 2,022 in all, 1,613 of them outside software.
Exposure. A research agent scored each company from 1 to 5 for how much of its core work current AI can do or change, using the rubric described above: the share of time spent reading, writing, deciding, scheduling and answering, and how repetitive that work is. When a company fell between two scores, we took the lower one. “Highly exposed” means a score of 4 or 5.
AI product. Research agents checked each company’s website, LinkedIn, job postings and press coverage, and recorded whether it markets AI as a feature of what it sells. An independent second pass checked every batch of five companies before anything was recorded. We deliberately don’t use the other signal they collected, AI in a company’s own operations: from the outside, it’s too weak to separate from noise.
Sponsor size and AI leads. From Parts 1 and 2, recounted in October 2026. “Lower mid-market” means under $500 million in PE assets. The firm-size figure uses our full set of 6,122 scored PE-backed companies; the rest of this post uses the 2,022-company sample.
Sponsor stance. From Part 1: a named AI lead, a described AI initiative, a mention of AI in the firm’s thesis, content or portfolio materials, or nothing.
Analysis. Adjusted figures come from a logistic regression on the 434 highly exposed non-software companies, with standard errors clustered by sponsor. We checked it separately in each research batch and by leaving out each sponsor in turn.
This measures public evidence. A company building AI quietly reads as having none, so these are floors, and smaller companies publish less. These are associations, not proof of cause.
Notes
- Talas, “The AI gap in private equity is a size gap,” September 29, 2026. https://talas.co/blog/the-ai-gap-in-private-equity-is-a-size-gap/
- Talas, “Who leads AI at private equity firms,” October 12, 2026. https://talas.co/blog/who-leads-ai-at-private-equity-firms/
- We mapped mega-fund portfolios mostly for firms with an AI lead, and our large-firm sample over-represents them (42 of the 43 large firms with an AI lead, against 54 of the 185 without).
- This sample is narrower than the one behind the firm-size figure: 2,022 PE-backed companies at 181 sponsors, researched in more depth. Companies with no sector tag are left out of the sector figures but included in the totals. CHECK: rerun on the Clarify dataset, or keep and say why
- Logistic regression of “sells an AI product” on whether the sponsor is silent on AI, controlling for the company’s sector, exposure score, headcount band and age band, and for whether the sponsor focuses on software and has $2 billion or more in PE assets. Standard errors are clustered by sponsor. Silent sponsor: odds ratio 0.28 (95% CI 0.14–0.56), n = 434. Shares quoted in the text are the model’s average predictions with every company set to each stance in turn.
- Examples are drawn from our wider dataset of scored PE-backed companies and anonymised. Descriptions are paraphrased from each company’s public materials.
- U.S. Securities and Exchange Commission, Form ADV filings, Part 1 and Schedule D (private fund reporting).