Services is the one sector where AI maps straight to the P&L
Professional services firms sell hours and documents. There’s no factory, no inventory — the cost base is people, and the profit lever is how much billable output each person produces. That makes the thesis unusually clean: anything that lets the same headcount produce more billed work flows directly to EBITDA.
The independent evidence is a 2023 Harvard/MIT/Warwick randomized trial inside BCG: 758 consultants randomized to GPT-4 or not. The AI group completed 12.2% more tasks, 25.1% faster, with 40% higher quality on tasks inside the tool’s competence (study summary). But the same study carries the warning: on tasks outside the model’s competence, AI users were 19% more likely to get the wrong answer (Harvard Crimson).
The billable-hour conflict — the silent killer most articles miss
Read this before the plays, because it determines whether any of them produce EBITDA. Under hourly billing, cutting hours cuts revenue. A firm that bills by the hour has a direct financial disincentive to deploy efficiency tools — and an AI-disclosure paradox: disclose it and the client won’t pay for the saved hours; hide it and the invoice gets indefensible (Geek Law Blog). Fixed-fee or value-based billing is the prerequisite, not a nice-to-have. If the portco bills hourly, the AI ROI case is structurally broken before you start.
"What share of our revenue is billed hourly vs. fixed-fee? For the hourly share, who captures the efficiency gain when AI cuts the work — us or the client?" If the answer is "the client," the AI EBITDA case doesn't exist until billing changes.
Document drafting & review: the biggest, most-proven lever (legal, tax)
Document-heavy disciplines are where gains are most concrete. At A&O Shearman, an enterprise Harvey deployment across 4,000 staff cuts contract review time roughly 30%, saving attorneys up to ~7 hours per review (Harvey); DLA Piper expanded to 5,000 licenses (LegalTechnology). Tax shows the same shape: Thomson Reuters’ CoCounsel Tax helped accounting firm Copeland Buhl complete research 2–3x faster (legal.io). The honest caveat: these are document-production gains. Whether they convert to EBITDA depends entirely on the billing model above.
Staffing & recruiting: more placements per recruiter, same headcount
Staffing margin is placements per recruiter and time-to-fill. Bullhorn’s survey found 78% of firms growing revenue 25%+ use AI in their ATS, and 46% say AI cut screening time in half or better (Bullhorn). The cleanest data point: staffing firm Employment Enterprises reported 23% higher weekly gross profit with the same headcount (StaffingHub). Treat the survey as correlational, but the same-headcount-more-margin figure is exactly the operating-partner KPI.
The PE roll-up angle: AI as the shared platform Main Street can’t afford
The sharpest PE-native thesis is in the accounting consolidation wave. Blackstone bought into Citrin Cooperman at a ~$2B valuation; Charlesbank-backed Aprio and Alpine-backed Ascend are aggressively rolling up regional CPA firms, hiring “Head of AI” roles to standardize advisory offerings (CPA Trendlines). The real insight: in a roll-up, AI’s value isn’t per-firm productivity — it’s amortizing one good AI stack across 30 acquired firms.
Marketing agencies are the cautionary mirror: Publicis invested in its CoreAI stack and grew headcount ~5% while WPP cut from 108,044 to 98,655 (VideoWeek). Same technology, opposite strategies — proof that AI is an input, not an outcome.
"What is our current hours-per-document (or placements-per-recruiter), measured today? What's the target after AI, and who owns hitting it?" No pre-deployment baseline = no provable ROI = the project gets cut.
What kills these initiatives
- The billable-hour conflict. Under hourly billing, efficiency cuts revenue. Fixed-fee is the prerequisite.
- Pilot purgatory. MIT found 95% of corporate GenAI pilots fail to deliver measurable P&L return — the cause is organizational integration, not model quality (Fortune/MIT).
- Spending on the wrong function. Over half of GenAI budgets go to sales/marketing while the biggest ROI sits in back-office document work.
- No baseline. If you didn’t measure hours-per-document before deployment, you can’t prove the lift.
- The jagged frontier. Deploying AI on tasks it’s bad at increases error rates. Scope it tightly.
- Client-data confidentiality. Privileged legal and tax data can’t go into consumer tools — enterprise, contained deployments only.
Sources
| Source | What it told us | Confidence |
|---|---|---|
| BCG/Harvard RCT | +12.2% tasks, 25.1% faster, 40% quality | STRONG |
| Harvard Crimson | Same study: +19% errors on out-of-frontier tasks | STRONG |
| Fortune/MIT | 95% of GenAI pilots no P&L return; back-office best ROI | STRONG |
| Harvey — A&O Shearman | 30% faster review, ~7 hrs/contract, 4,000 staff | MEDIUM |
| LegalTechnology | 100k+ lawyers; DLA Piper 5,000 licenses | MEDIUM |
| legal.io — CoCounsel Tax | Copeland Buhl: 2–3x faster research | MEDIUM |
| Bullhorn GRID | 2,300-firm survey: 78%/55%/46% AI-KPI stats | MEDIUM |
| StaffingHub | Employment Enterprises +23% GP, same headcount | MEDIUM |
| VideoWeek | WPP headcount 108k to 98.6k; Publicis +5% with CoreAI | MEDIUM |
| CPA Trendlines | Blackstone/Citrin $2B; Aprio/Ascend roll-ups; Head-of-AI hires | MEDIUM |
| Geek Law Blog | Billable-hour disincentive + AI disclosure paradox | MEDIUM |