PE-backed industrial services company, ~$20M revenue, ~150 employees · three-month field study, then implementation · Client details anonymized. Study and implementation figures are actual; projected revenue is a projection.
Situation
An insulation and scaffolding contractor in a middle-market private equity portfolio: field crews, quoting, scheduling, and a back office running on spreadsheets and institutional memory. The sponsor's thesis was margin expansion through applied AI, and the sponsor's question was the one every middle-market fund now asks: where does AI actually create enterprise value here, and how much of it is real?
The instinct in most portfolios is to start from the technology and look for places to apply it. We started from the work: what happens on a job, who touches it, where the hours and the margin go.
The study
The engagement began as a three-month field study, conducted in parallel with primary research at Wharton on AI value creation in private equity: seven interviews with PE professionals and a survey of investment and operating partners. On site, we mapped the workflows end to end and specified, costed and valued five use cases, individually and in combination.
Standalone, four of five cleared the investment bar. ROI from 50% to 342%, NPV from $39k to $474k per case. The fifth was marginal.
Modelled together, all five cleared. They share the same captured knowledge and the same data foundation. The marginal case, the one most people would have cut first, was the enabler for the other four.
Four findings shaped what came next:
- The pricing logic lived in one head. How a job gets priced existed as a spreadsheet, a set of codes and an estimator's judgment built over years. Nothing was written down in a form another person, let alone a system, could execute. The largest automation opportunity and the largest key-person risk were the same thing.
- Valuing use cases in isolation understates the portfolio. Funds that triage AI opportunities case by case systematically under-invest, and often cut the enabling case first.
- The constraint was adoption, not capability. Every case was feasible with tools available at the time. In a long-tenure workforce with no software habit, an interface change is a bigger project than a model choice.
- Automating a broken process locks in the inefficiency. Two candidate workflows existed only because systems did not talk to each other. Automating them as-is would have industrialised the workaround.
The build
The recommendation was to capture before automating: extract the pricing logic into a structured, auditable knowledge base, with the senior estimator reviewing outputs before anything ran unattended. The capture is the project. The automation is the easy part afterwards.
The first build was AI bid generation and outreach: bids drafted from the captured pricing logic and the job history, reviewed by the estimator, and a CRM-driven outreach loop so business developers could carry more bids without more hours.
Results
- Payback in 1.45 months.
- 836+ hours a year saved on bid preparation.
- ~$800K projected incremental revenue from CRM plus AI outreach, assuming two additional bids per year per business developer.
- Zero additional hires to scale reporting.
What this means for a sponsor
Three lessons for any middle-market fund.
Fund the enabling case even when it scores lowest. Sequence the portfolio so the shared foundation is built first, then value the remaining cases on top of it.
Price the key-person risk, not just the hours. The estimator's knowledge was the asset. Capturing it protected the business before it accelerated it.
Adoption is the project. With a workforce that had never used software for this, the interface and the review loop decided the value, not the model.