Walk any PE conference floor this year and AI is the only conversation. Every operating partner has a thesis. Every portfolio company has an initiative. Every LP update has a slide.
Then look at what is actually running in production.
According to Grant Thornton, only about 5% of PE-backed companies have fully integrated AI. The cross-industry figure is roughly 14%, and tech-sector companies run far higher still. About 45% of PE-backed companies are still in piloting.
So the sector that talks about AI the loudest ships it the least. That gap is not an accident. It is structural, and it traces back to the same root every time.
The loudest sector and the lowest integration
Hold those numbers next to each other. PE companies are integrating AI at roughly a third of the rate of the broader economy, and a small fraction of the rate of tech portfolios. Yet nobody is generating more AI noise than private equity right now.
The volume is real. The pressure on management teams to “have an AI story” is enormous. LPs ask about it. Operating partners push it. The board wants it in the deck.
The output is not. A 5% integration rate against a 45% piloting rate tells you exactly where the work is stuck. It is stuck in the pilot. The same trap I described in The Pilot That Never Ships is now playing out across an entire asset class, on AI specifically, at scale.
This post is the sequel. The pilot trap explains why a single initiative stalls. The 5% problem explains why PE as a sector stalls more than anyone else. There are four structural reasons, and they compound.
Reason one: the committee scopes the pilot
In a portfolio company, AI rarely has a single owner with budget and authority. It has a committee.
The committee forms because nobody wants to be wrong about something this visible. Finance is in the room. IT is in the room. The relevant business unit is in the room. Sometimes the operating partner dials in.
A committee does not scope a sharp question. It scopes a safe one. The pilot that emerges is broad enough that everyone can support it and vague enough that nobody is exposed if it fails. “Explore AI for operational efficiency.” “Assess AI opportunities across the customer lifecycle.”
That is not a pilot. That is a heading. It has no binary outcome, no deadline that means anything, and no single person who funds the rollout if it works. A committee-scoped pilot is designed to survive, not to resolve. It produces status updates, not decisions.
Reason two: the vendor runs the POC on clean data
The second reason walks in with a demo.
The AI vendor arrives, plugs in, and shows the model working beautifully. The board is impressed. The committee is relieved. The POC gets funded.
The demo ran on clean sample data. It always does. The vendor curated a tidy subset because that is what makes the model look good in the room.
Then the POC meets the company’s actual data. Duplicates. Inconsistent customer hierarchies. Revenue figures in the CRM that do not match finance. Product codes changed three times with no history. The model still works. The outputs are useless, because the data underneath was never ready.
The company now faces the decision it was avoiding. Fund the data foundation, or narrow the scope, find another clean subset, and produce another round of promising results. Most narrow the scope. The vendor is happy to keep running. The POC becomes a permanent fixture that demonstrates capability and ships nothing.
Reason three: AI is a side project to the day job
The third reason is the quietest and the most common.
Nobody at the portfolio company is doing AI as their job. They are doing it on top of their job. The finance analyst running the pilot still owns the close. The ops manager testing the tool still owns the floor. The IT lead still owns everything that breaks.
AI gets the hours that are left over. In a mid-market company running lean, there are no hours left over. So the pilot advances in the gaps between real work, which means it advances slowly, inconsistently, and only when the day job permits.
This is why pilots that should resolve in 60 days drift for nine months. Not because the question is hard. Because nobody is assigned to answer it full time. A pilot run in the margins produces results in the margins. The integration rate stays at 5% because integration is real work and real work needs real ownership, not a volunteer.
Reason four: there is no funded path to production
The fourth reason is where the other three come home.
Even when a pilot produces a clean answer, there is usually no money set aside to act on it. The pilot was funded. Production was not. Nobody budgeted for the data engineering, the integration, the change management, the ongoing run cost.
So the pilot succeeds and then sits. The committee notes the success. The board sees a green slide. And the initiative quietly waits for a production budget that was never scoped, because scoping it would have surfaced the real cost up front, which is exactly what the pilot was structured to avoid.
A pilot without a funded path to production is not a step toward integration. It is a demonstration that integration is possible, filed next to all the other demonstrations. That is the gap between 45% piloting and 5% integrated. Forty points of companies have proven they could, and have not funded the work to actually do it.
The common root: a data foundation nobody resourced
Read those four reasons again. The committee, the vendor demo, the side-project staffing, the missing production budget. Underneath all four sits the same thing.
The data foundation was never resourced.
The committee scopes vaguely because a sharp question would expose what the data cannot answer. The vendor runs on clean data because the real data is not ready. The pilot stays a side project because resourcing it properly would mean confronting the data work first. And there is no production budget because the production cost is mostly data cost, and nobody wanted that number in the room.
AI does not fail in PE portfolios because the models are weak. The models are extraordinary. It fails because you cannot run an extraordinary model on data that disagrees with itself. I made this case directly in AI Readiness Starts With Data Readiness. The 5% problem is the proof at sector scale.
This is also why so many initiatives present as stalled rather than failed. A failed initiative produces a clear answer and frees the resources. A stalled one keeps reporting progress while the underlying constraint stays untouched. I walked through that specific pattern in Why Your Portfolio Company AI Initiative Is Stalling. The stall is not a failure of ambition. It is a foundation that was never built.
What the 5% did differently
The companies in the 5% did not buy better AI. They sequenced the work correctly.
They named one owner with budget and authority, not a committee. They scoped a pilot with a binary question, a deadline, and success criteria written down before launch. They ran it on real operational data from day one, not a curated sample. They staffed it as a job, not a favor. And they funded the path to production before the pilot started, so a successful result had somewhere to go.
Most of all, they treated the data foundation as the first deliverable rather than the inconvenient discovery halfway through. They established single sources and single definitions for the numbers the AI would consume, so the model had clean inputs to work with. Then the AI was the easy part.
That sequence is unglamorous. It does not demo well. It is the difference between the 5% and the 45%.
If you want to know which side of that line your portfolio company sits on, the AI Readiness assessment is built to surface it before a vendor does.
Stop talking. Resource the foundation.
The PE sector will keep being the loudest voice on AI. The conference slides will get more confident. The LP updates will get more ambitious.
None of that moves the 5%. What moves it is naming an owner, funding the data work, staffing the pilot as a real job, and scoping the production budget before the demo ever runs.
The companies doing that will integrate AI while the rest of the sector is still presenting promising results. At exit, a buyer can tell the difference in about ten minutes. One portfolio company runs on AI that touches real operations. The other has a demo that runs on a spreadsheet.
Stop talking. Resource the foundation.