Key Takeaways
- According to OneGuide's State of Portfolio Ops Report, 51% of private equity firms report engineering efficiency gains from AI, but fewer than 6% see a 10% EBITDA lift.
- High-cost frontier models crush unit economics when deployed on routine automated tasks, preventing efficiency gains from reaching the bottom line.
- Portfolio companies are replacing human security champions programs with cheap open-weight models running continuous 24/7 PCI-DSS compliance audits.
- Dave Mangot points to OpenAI's August software factory paper as proof that architecture and system design dictate software productivity, not model sophistication.
The EBITDA Disconnect in Portfolio Software Engineering
Private equity sponsors bought into generative AI on the promise of software developer productivity. The initial numbers looked promising on paper, yet the financial return remains elusive. OneGuide's State of Portfolio Ops Report revealed that 51% of sponsors report engineering efficiency gains, but fewer than 6% achieved a 10% EBITDA lift. Faster code generation does not produce margin expansion if that code hits downstream operational bottlenecks or runs up massive API bills.
Dave Mangot explains the gap through operational choices. When engineering teams deploy top-tier proprietary models for continuous internal tasks, computing costs cannibalize the savings. As Mangot observed: “Sure we could go see whatever the latest thing that OpenAI has to offer whatever the most expensive thing is because it'll probably work, but that's going to destroy your margins and we want a 10% EBITDA lift.” (play from 34:40)
Automating the Chores That Engineers Avoid
Real margin expansion comes from aiming automation at structural friction rather than writing raw code. Engineering organizations have historically relied on informal security champions programs, assigning regular developers to police team compliance and review code. These initiatives routinely fail because developers dislike compliance chores and prioritize product delivery.
OpenAI highlighted a structural solution in its August software factory paper by deploying an automated security agent directly into the code review pipeline. Mangot saw the pattern immediately: “If you look at the OpenAI software factory that got published in August, they have a security agent that reviews the code. And I was like, 'Oh my god, that's the security champion.'” (play from 29:45)
Rather than burning capital on frontier models, smart operators apply focused open-weight models to these repetitive operational tasks. Mangot highlighted portfolio companies running continuous 24/7 PCI-DSS compliance audits using inexpensive local models. “They weren't even using any kind of fancy, super frontier model, super expensive, whatever,” (play from 32:00) Mangot noted. “It was a very cheap open-weight model that they could just run 24/7 because it was cheap and it was effective. And I was like, this is good engineering.” (play from 32:08)
Architecture Over Raw Model Power
The mistake many operating partners make is evaluating AI tools purely by model benchmarks rather than pipeline architecture. Mangot reflected on lessons from earlier infrastructure shifts: “We've said in the DevOps movement forever, architecture matters, technology doesn't, because you put great technology into a lousy architecture, you're not going to get a great response.” (play from 33:26) Running code generators inside broken QA workflows or running $20-per-million-token models on simple validation sweeps guarantees poor returns. Margins expand when teams build automated architectures that use the cheapest possible model to take unpleasant tasks off human plates completely.
Why It Matters
This dynamic signals a shift in how private equity sponsors must underwrite AI-driven value creation plans. The easy thesis of cutting developer headcount through generic code completion has stalled against gross margin realities and downstream technical debt. Real operating leverage requires re-architecting operational pipelines to offload compliance and QA onto low-cost, self-hosted open-weight infrastructure.