Key Takeaways
- Data from OneGuide's State of Portfolio Ops Report shows 51% of private equity firms report software engineering efficiency gains from AI, but fewer than 6% capture a 10% EBITDA lift.
- AI code assistants flood engineering pipelines with pull requests, exposing downstream operational bottlenecks like manual QA, brittle test suites, and fragmented data.
- Dave Mangot points to deployment frequency from the DORA metrics as the single operational proxy for whether an engineering organization can absorb AI velocity.
- Portfolio CTOs frequently blame enterprise customer contracts for weekly or monthly release cadences, yet DORA benchmarks confirm smaller, continuous shipments behind feature flags carry lower operational risk.
The Bottleneck Moves Downstream
Private equity sponsors bought into the promise that generative AI would cut software development costs by double digits. The reality inside portfolio companies looks very different. While OneGuide's State of Portfolio Ops Report found that 51% of private equity firms cite efficiency gains from AI inside software engineering, fewer than 6% saw a 10% EBITDA lift. Coding assistants make it cheap to generate raw code, but raw code is not finished product.
Dave Mangot pointed out that the operational basics inside most mid-market software assets are broken. When developers use AI to produce code twice as fast, they simply slam more pull requests into existing bottlenecks: manual QA teams, broken staging environments, and slow compliance sign-offs. Mangot recalled hearing similar frustrations from top technical leaders: “There's the quote the basics and I was like oh my god it's so refreshing to hear the CISO of Anthropic frontier model lab saying focus on the basics. And that's exactly what I'm seeing inside these portfolio companies: you release once a week if you're lucky and you do it in this weird thing, but you want agents to crank out 6,000 changes a week.” (play from 48:53)
Generating 6,000 changes a week into an engineering pipeline designed for manual monthly sign-offs creates inventory pileups, not free cash flow.
The Broom Factory Test
When deal teams evaluate software operations, they often get buried in vanity metrics like developer velocity, story points, or AI seat adoption. Mangot argues that the single best metric for boards to review is deployment frequency from the DORA metrics. Shipping code to production on a daily basis requires mature automation, fast test suites, and modern feature flagging.
Mangot uses a simple manufacturing analogy to describe the problem: “If I have a factory and I'm producing brooms, do I want to produce a broom once every 5 days? Like no, I want to be producing brooms every day all the time. And that means the system is just flowing.” (play from 51:38) Batching production into massive weekly or monthly releases creates inventory drag and operational volatility. Daily release cadence proves the assembly line works.
Portfolio engineering teams routinely push back on this standard, arguing that their enterprise clients demand stability and cannot tolerate daily updates. Mangot notes that this argument confuses deployment with feature release. Teams can deploy code to production continuously while hiding new functionality behind feature flags until customers are ready. As Mangot observed, “releasing more often is safer.” (play from 55:20) Small, continuous releases isolate bugs, eliminate panic during deployments, and allow engineering teams to actually capture the throughput gains promised by AI tools.
“Just as a simple first pass proxy, the thing that I would be asking if I'm the board, my CTOs, how often do we deploy? And if it's not daily, that's where you start digging in,” (play from 51:04) Mangot said. “Deployment frequency to me says if I can deploy every day, it's very likely that I have my processes set up in an automated way that allow me to take advantage of AI.” (play from 50:27)
Why It Matters
The gap between engineering efficiency and EBITDA expansion signals that software margins will not expand purely through seat licenses for AI coding tools. Sponsors that underwrite margin expansion to AI developer productivity without auditing pipeline automation will see higher defect rates rather than reduced headcount or faster time to market. Realized software value creation is concentrating in assets with mature continuous delivery infrastructure, widening the operational performance spread between modern SaaS platforms and technical debt-laden rollups.