Every PE firm has an AI thesis. Every operating partner has been asked to find AI use cases across the portfolio. Every conference panel this year has a slide on it.
Then you read what GPs actually report when they are surveyed, and a different picture shows up.
S&P Global asked general partners how far AI has made it into their own work. Only about 31% say AI is somewhat or fully integrated into their diligence process. That is not the back office. That is the core of what a PE firm does, the part where edge is supposed to come from, and roughly two thirds of firms have not put AI into it in any meaningful way.
The ambition is real. The integration is not there yet. That gap is the most useful thing in the data, because it tells you what is actually hard, and it is not the part most people assume.
The hype says models. The data says effectiveness.
The story the market tells is about access. Better models, cheaper tokens, easier tools. The implication is that adoption is a procurement problem. Buy the tool, point it at the work, capture the edge.
PE’s own numbers do not support that. The firms that have tried AI in the places it should help most are not impressed with the results.
In the same S&P Global research, large majorities rate AI as ineffective where it would matter most. Around 64% call it ineffective for deal sourcing. Around 75% call it ineffective for portfolio monitoring. These are not the opinions of people who have never touched the tools. These are the people closest to the work telling you the output is not good enough to rely on.
Sit with the portfolio monitoring number for a second. Monitoring is, on paper, the perfect AI use case. It is repetitive. It runs across many companies. It is pattern recognition over financial and operational data, month after month. If AI cannot earn its keep there, the problem is not the model’s reasoning ability.
The problem is what the model has to read.
Why monitoring is the tell
Deal sourcing being hard is easy to explain away. It is a noisy, relationship-driven game, and a lot of the signal lives outside any database. You can argue AI was always going to struggle there.
Portfolio monitoring does not have that excuse. The data exists. It sits in the portfolio companies. A monitoring model should ingest each company’s financials and operating metrics and surface what is drifting, what is at risk, and what needs a call this week.
When three quarters of GPs say that does not work, you are looking at a data problem wearing an AI costume. The model is fine. It is being asked to reason over numbers that do not reconcile across companies, definitions that differ from one portco to the next, metrics that arrive late, and reporting that was built for a quarterly board deck rather than a machine.
Feed that into the best model on the market and you still get unreliable output. Not because the model is weak, but because the input is. This is the same pattern I wrote about in why your portfolio company’s AI initiative is stalling. The tool works on clean demo data and falls over on real company data. PE is now seeing that pattern in its own operation, one level up.
The same logic explains the diligence number. Only about 31% have integrated AI into the diligence process, and diligence runs on the target’s data, which is the least controlled data a firm ever touches. A target hands you a data room, and that room reflects whatever shape the company’s data is actually in. If the model has to reason over a sell-side data room full of inconsistent definitions, gaps, and numbers that do not tie out, it produces work a deal team cannot defend in an investment committee. So the deal team does the work by hand, the way it always has, and the AI stays a side experiment. The integration rate is low because the data the model would read is unreliable by default.
What GPs say the barrier actually is
This is the part that should change the conversation.
When S&P Global asked what is holding AI adoption back, the most cited barrier was not cost. It was not model quality. It was not security or regulation. The single most cited barrier, named by roughly half of respondents, was lack of in-house expertise.
Read that against the effectiveness numbers and it lines up. Firms are not failing to adopt AI because the models are inaccessible. They are failing because they do not have the people who can connect a model to data that is fit to use, judge whether the output can be trusted, and build the plumbing that keeps it fed.
That is not a model gap. It is a data and capability gap. The expertise that is missing is the expertise that makes data usable in the first place. Cleaning it, reconciling it, defining it once so two people asking the same question get the same number, and wiring it into something a model can actually consume.
There is a version of this barrier that firms misread. They hear “lack of in-house expertise” and they hire a head of AI, or they buy a platform with a vendor team attached. Then nine months later the effectiveness numbers have not moved, because the new expertise was pointed at the model and the constraint was always the data feeding it. The person who closes this gap spends most of their time on things that do not look like AI at all. Reconciling sources, agreeing definitions across a portfolio, fixing the pipeline that arrives three weeks late. Unglamorous, and the precondition for everything downstream.
PE has run straight into the wall it keeps telling its portfolio companies to climb.
The mirror
Here is the uncomfortable symmetry.
Operating partners ask portfolio company CEOs to deploy AI. The CEOs hit the same problem the GPs just reported. The customer data is fragmented across four systems. Revenue does not reconcile. The historical data has a hole from a migration two years ago. The pilot works on clean data and dies on production data.
The barrier inside the portfolio company is the same barrier the GP named for its own firm. Not access to models. The data underneath, and the people who can make it ready.
If PE cannot make AI effective on its own diligence and monitoring, where the firm controls the process and the talent is as concentrated as it gets, the expectation that a mid-market portfolio company will do it on a thinner team and messier data needs a reset.
AI cannot work on a portfolio company’s data if that data is not ready. PE is discovering this on itself first. The numbers in the S&P Global survey are the early read on a lesson the whole asset class is about to learn at the portfolio level.
What the numbers tell you to do
The takeaway is not that AI is overhyped and you should wait. The firms that figure out the data layer will pull ahead precisely because most have not. The takeaway is about sequence and where you put the money.
Stop treating AI adoption as a tooling decision. The 31% integration figure and the 64% and 75% ineffectiveness figures point to one conclusion. The model is rarely the constraint, so a better model rarely moves the needle.
Spend on the layer underneath. For your own firm, that means data your monitoring can trust. Consistent definitions across the portfolio, metrics that arrive on time, financials that reconcile. For each portfolio company, it means the same discipline one level down, before any AI use case gets funded.
Fix the expertise gap honestly. Half of GPs named it as the barrier. The first hire that unlocks AI is usually not an AI hire. It is the person who makes data usable, the data engineering and data quality capability that has to exist before a model has anything dependable to read. I have made that case at length in why AI readiness starts with data readiness.
Use the same bar on diligence. If AI is going to help you assess a target, it has to read that target’s data, which means the target’s data has to be readable. The firms raising the standard here are turning data readiness into a diligence question rather than a post-close cleanup, which I covered in how AI-powered due diligence is raising the bar.
If you want a fast read on whether a company is actually ready for any of this, the AI readiness assessment walks through the data conditions that have to be true before a model can deliver anything you would stake a number on.
The bottom line
PE’s own survey data undercuts the AI hype, and it does so in PE’s voice, not a vendor’s. Most firms have not integrated AI into diligence. Most call it ineffective for sourcing and monitoring. The barrier they name is expertise, not model access.
Strip that down and it is a data readiness problem at the heart of the industry. The model is ready. The data, and the people who make data usable, are not. The firms that act on that order, data first, model second, are the ones that will make AI earn its place, on their own desk and across the portfolio.