Every PE portfolio company is being told they need AI. The pilots run. The demos land. The board is happy for a quarter. Then production hits, and the whole thing quietly collapses.
The problem is almost never the AI. The problem is what sits underneath it.
Where the theatre starts
An AI project in a portfolio company almost always begins the same way. A workshop. A slide with fifteen use cases. Two or three make the cut. A model gets trained on a slice of historical data pulled from a spreadsheet or a one-off export. The pilot looks good. Everyone signs off on the roadmap.
Then someone tries to put it in production. And it becomes clear that the operational data the model needs — clean, current, coming out of the systems that actually run the business — doesn't exist in a form anyone can plug into. The ERP is a version behind. Half the machines don't stream anything. The finance system and the ops system disagree about basic counts. Nobody owns the pipeline that would make this feed anything real-time.
The AI didn't fail. It never had a foundation to stand on.
Why this pattern is so consistent in PE portfolios
Two structural reasons.
First, PE portfolio companies are almost always the product of a carve-out, a roll-up, or aggressive geographic expansion. Each of those events fragments the data landscape by design. The carve-out leaves the portco with a read-only export of the parent's ERP and no access to the parent's data warehouse. The roll-up bolts three companies together whose systems have never spoken to each other. Cross-border expansion means three different countries running three different stacks with three different tax and accounting layers.
Second, the pressure to show value creation early in the hold pushes teams toward visible, high-status projects. AI is one of them — before the boring, unglamorous plumbing is done. Nobody wants to spend the first six months of a 100-day plan on ETL. So they don't. And then they try to run the interesting layer on top of what isn't there.
What actually needs to happen first
Not more AI use case workshops. Not another consulting engagement.
Someone needs to unify the operational data. Bridge the ERP if it's a carve-out. Stand up the pipelines that pull real numbers off the machines and into a single place. Rebuild the reporting layer so the finance system and the ops system show the same truth. Instrument the parts of the business the value creation plan actually depends on.
That work is unglamorous. It looks nothing like AI. But it is the entire precondition. Once it's done, AI stops being theatre and starts being an EBITDA lever.
Two questions before your portfolio company greenlights an AI project
- What operational data does this AI need to work in production, and does the portfolio company currently produce it — reliably, in the form the model requires?
- If not, what has to be built first, who owns it, and how does that sequence into the value creation plan?
If the answers to either are unclear, the AI project isn't ready. It's a pilot that will look good until it doesn't.
If you're about to greenlight an AI project in a portfolio company and want a two-week read on whether the data can actually support it, get in touch.