Key Takeaways

  • According to OneGuide's State of Portfolio Ops Report, 51% of private equity firms report software engineering efficiency gains from AI, but under 6% achieve a 10% EBITDA lift.
  • Portfolio companies face their primary AI hurdle not in talent shortages or model access, but in basic data infrastructure.
  • Applying Eliyahu Goldratt's Theory of Constraints from The Goal, accelerating code generation simply shifts the traffic jam onto manual quality assurance and fragile pipelines.
  • Deployment frequency remains the most reliable operating proxy for determining whether an engineering organization can absorb AI tooling without stalling.

The Code Generation Illusion

Private equity sponsors poured capital into generative coding assistants expecting immediate margin expansion. The thesis seemed straightforward: write code faster, trim engineering headcount, and expand margins. Yet the math failed to translate to portfolio income statements. While half of sponsor-backed tech assets write code faster with AI, almost none capture meaningful operating earnings from that speed.

Dave Mangot points straight to the software delivery life cycle to explain the disconnect. Generative tools speed up typing, but typing was rarely the true constraint. “These AI systems were built inside the same software delivery life cycles that we use now,” (play from 12:47) Mangot noted. “It's just now the coding part isn't as slow and all these companies that are out there that are still doing, and I hope this isn't too technical for like your audience, but like manual QA.” (play from 12:56)

When developers produce code at triple their normal rate, manual testing teams collapse under the review burden. Pull requests pile up. Untested features languish in staging environments. The portfolio company burns cash on AI licenses while code sits idle waiting for human sign-off.

The Theory of Constraints in Portfolio Ops

Mangot traces this operational trap to Eliyahu Goldratt's classic operational text, The Goal. Increasing throughput at an unconstrained work center does not increase system throughput; it only increases work-in-progress inventory.

“If our bottleneck isn't how fast we can write code anymore and we start putting let's say 10 times as much code through the system, it doesn't mean the rest of the system is going to be like oh yeah we were totally built for something 10 times the throughput that we were doing before,” (play from 9:41) Mangot explained.

When a portfolio company feeds broken data architectures into automated workflows, it does not fix the underlying operation. It scales the failure rate. As Mangot observed, “I see portfolio companies that take broken stuff and then try to put it inside their AI factory and guess what, it is just as broken as it was.” (play from 11:39)

Plumbing Over Exotic Models

The OneGuide findings reveal that poor data infrastructure ranks as the primary barrier to AI implementation across portfolio companies. Sponsors often search for proprietary algorithms or scarce machine learning researchers, yet the actual impediment is basic plumbing.

As the Private Equity Funcast hosts defined it, “when we talk about data infrastructure, what we're going to talk about is the pipes, where your data lives, whether data is clean, whether a system can get at it without a human exporting a spreadsheet first.” (play from 7:54)

If analysts must manually export CSV files between systems or reconcile fractured databases before a model can run, automation grinds to a halt. Mangot highlighted the inevitable outcome: “if your data infrastructure has bottlenecks in it, AI is going to expose those bottlenecks pretty severely.” (play from 0:10) Until companies automate testing, clean up fragmented data stores, and achieve high deployment frequencies, AI investments remain stranded on the balance sheet.

Why It Matters

This gap signals that software underwriting models banking on rapid AI-driven margin expansion are mispriced. Deal teams modeling margin expansion through automated software delivery will miss exit hurdles if they treat code generation as an operational fix. Valuation multiples will increasingly separate platforms with automated deployment pipelines from legacy codebases choked by technical debt and manual data manipulation.