Cross-Portfolio AI Economics: How to Benchmark Across 12 Portcos

Most PE firms have 12-30 portfolio companies. Fewer than 20% can build a cross-portfolio AI economics rollup that answers the question: "Which company has the most efficient AI, and where can we transfer the playbook?"

This capability—the ability to see AI unit economics across the entire portfolio, spot variance, identify best practices, and replicate them systematically—is a high-leverage operating moat. It's also surprisingly simple to build. This article walks through the construction of a real-world 12-company rollup, shows where the operating leverage lives, and provides a template for replicating it at your firm.

Why Cross-Portfolio AI Economics Matters

A PE firm with $10B in AUM and 12 portfolio companies, where each company has $300K-$800K in annual AI spend, has a total AI spend footprint of roughly $6M-$9.6M across the portfolio. If 30% of that spend is inefficient (companies running at 1.2x cost per outcome compared to the efficiency leader), then there's $1.8M-$2.9M in portfolio-wide optimization upside. That's real money.

But the leverage is not just about cost reduction. It's about replication. If Company A has figured out how to run an AI claims-adjudication agent at $0.65 per claim and Company B is running a similar agent at $1.10 per claim, and both companies process 50,000 claims/month, then Company B is spending $22,500/month on the same capability that costs Company A $16,250/month. That's $6,250/month or $75,000/year of delta.

The operating partner's job is to: (1) spot the delta, (2) understand why it exists, (3) transfer the best practice from A to B, and (4) capture the savings.

Cross-portfolio benchmarking makes this visible, repeatable, and systematic.

Building the Rollup: Three-Month Timeline

Month 1: Collect Baseline Data from 3-5 Companies

Start by collecting baseline AI cost and outcome data from your most mature portfolio companies—the ones that have already implemented cost-per-outcome attribution.

For each company, gather:

• Company name and industry vertical
• Primary AI-enabled business process (claims processing, customer support, loan underwriting, etc.)
• AI agents in use (count, primary vendors)
• Total monthly AI spend (all-in, including infrastructure)
• Monthly work volume (claims processed, tickets resolved, applications underwritten)
• Cost per outcome (monthly spend ÷ monthly volume)
• Attribution maturity stage (1-5 on the maturity curve)
• Headcount change attributable to AI (before/after FTE count)

By the end of Month 1, you should have 3-5 companies with clean data. Your operating partner should be able to see: Company A is running at $0.65/claim, Company B is running at $0.85/claim, Company C is running at $0.72/claim. Variance exists. You don't yet know why, but you've spotted the delta.

Deliverable: A 5-company baseline rollup with cost-per-outcome for each, and a list of questions about why variance exists.

Month 2: Expand to 8-10 Companies and Deep-Dive Variance

Expand the rollup to 8-10 companies. As you add companies, you'll discover that many of them are not yet at stage 3+ on the maturity curve (they haven't done cost-per-outcome attribution). For those companies, help the CFO and CTO calculate it. It's typically a 2-week sprint per company.

Now you have enough data to spot patterns. Example real rollup for a portfolio with five claims-processing companies:

| Company | Industry | Agents | Monthly Volume | AI Cost/Month | Cost/Outcome | Maturity | Notes | | --- | --- | --- | --- | --- | --- | --- | --- | | Company A | Healthcare (RCM) | 3 | 45,000 claims | $29,250 | $0.65 | Stage 4 | Leader; optimized prompts, batching | | Company B | Healthcare (RCM) | 2 | 38,000 claims | $31,900 | $0.84 | Stage 3 | Higher human review rate; newer agent | | Company C | Healthcare (RCM) | 4 | 52,000 claims | $37,440 | $0.72 | Stage 3 | Fine-tuned model; harder to replicate | | Company D | Healthcare (RCM) | 2 | 28,000 claims | $30,800 | $1.10 | Stage 2 | No attribution yet; rough estimate | | Company E | Healthcare (RCM) | 3 | 41,000 claims | $26,240 | $0.64 | Stage 4 | Cost leader; older, simpler agent design |

The rollup immediately shows:

• Outliers: Company D is 72% more expensive per outcome than Company E. Why? It's worth investigating.
• Best practices: Company A and E are cost leaders. They use similar models and vendors but slightly different prompts. Interview both teams.
• Replication opportunity: Company B is processing similar volume to A but at 30% higher cost. Can we transfer A's prompt to B?

Deliverable: An 8-10 company rollup with cost-per-outcome variance clearly visible. Identify top 3 variance questions and interview the teams.

Month 3: Build Quarterly Reporting and Scaling Template

By month 3, you've collected data from 8-10 companies and you understand why variance exists. Now formalize quarterly reporting and build a template that can scale to 20-30 companies.

The rollup table should include:

• Company name and vertical
• Business outcome metric (claims, tickets, applications, etc.)
• Monthly volume
• Total monthly AI spend
• Cost per outcome
• Maturity stage (1-5)
• Trend (cost per outcome improving, flat, or degrading month-over-month)
• Key drivers of efficiency (if known)
• Optimization opportunities (if identified)

Real example—a 12-company rollup:

| Company | Vertical | Outcome | Vol/Mo | AI Cost/Mo | Cost/Outcome | Stage | Trend | Notes | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | A | Healthcare | Claims | 45,000 | $29,250 | $0.65 | 4 | Down 5% | Best practice in prompting | | B | Healthcare | Claims | 38,000 | $31,900 | $0.84 | 3 | Flat | Newer agent; room to optimize | | C | Healthcare | Claims | 52,000 | $37,440 | $0.72 | 3 | Up 2% | Fine-tuned model drift | | D | Healthcare | Claims | 28,000 | $30,800 | $1.10 | 2 | Down* | Under-optimized; can transfer A's playbook | | E | Healthcare | Claims | 41,000 | $26,240 | $0.64 | 4 | Flat | Cost leader; older, simpler design | | F | Insurance | Claims | 35,000 | $22,050 | $0.63 | 4 | Down 3% | Similar efficiency to A; different domain | | G | Insurance | Underwriting | 18,000 | $14,400 | $0.80 | 3 | Flat | New agent; 6mo into deployment | | H | Insurance | Underwriting | 22,000 | $18,700 | $0.85 | 3 | Up 4% | Higher human review; model underperforming | | I | Finance | Loan Apps | 12,000 | $9,600 | $0.80 | 3 | Down 2% | Reasonable efficiency for lower volume | | J | BPO | Support | 65,000 | $32,500 | $0.50 | 4 | Down 8% | Highest efficiency; high volume helps | | K | BPO | Support | 48,000 | $38,400 | $0.80 | 3 | Up 3% | Overdue for optimization | | L | BPO | Support | 52,000 | $41,600 | $0.80 | 2 | Flat | No attribution yet; estimate only |

What the Rollup Reveals

This 12-company rollup immediately shows:

1. Variance: Cost per outcome ranges from $0.50 (Company J) to $1.10 (Company D). That's a 2.2x spread. Company J is doing something right that Company D is missing.

2. Vertical patterns: The two BPO (support) companies are more efficient than the healthcare and finance companies, likely because support is a higher-volume, more repetitive task that responds well to AI. Finance loan origination is lower volume and higher variance.

3. Maturity correlation: Companies at stage 4 (optimized/governed) are generally more efficient than stage 2-3 companies. This supports the hypothesis that measurement drives improvement.

4. Specific optimization opportunities:

• Company D opportunity: Company D is 72% more expensive than Company E on the same business outcome (claims processing). The hypothesis is that Company D's agent is newer and hasn't been optimized for efficiency. If we transfer Company E's prompt and prompt-optimization framework to Company D, we can likely move D from $1.10 to $0.70, saving $11,200/month or $134,400/year.

• Company K opportunity: Company K is processing 48,000 support tickets/month at $0.80/ticket, while Company J is processing 65,000/month at $0.50/ticket. Both use similar models (Claude + GPT-4). The difference is likely in prompt engineering, caching strategy, or fallback logic. If we can transfer J's practices to K and move it to $0.60/ticket, we save $9,600/month or $115,200/year.

• Company C drift: Company C is a fine-tuned model that started at $0.68/claim efficiency six months ago and is now at $0.72. Model drift is real. We need to either retrain the model or transition to an API-based agent (higher cost but lower operational overhead).

5. Portfolio-wide optimization target: If we execute the three biggest opportunities (D, K, C), we capture $134K + $115K + $24K (partial on C) = $273K/year in annual savings on a $360K/month ($4.32M/year) total portfolio AI spend. That's 6.3% portfolio-level margin expansion, with zero new investment.

How to Use the Rollup Quarterly

1. Update every quarter. Collect cost-per-outcome data from all portfolio companies. Sort by efficiency. Identify movers (cost improving or degrading month-over-month).

2. Spot and escalate outliers. If a company's cost per outcome rises 15%+ quarter-over-quarter, trigger an investigation. Root cause: model drift, volume cliff, price increase from vendor, or infrastructure misconfiguration.

3. Identify and execute transfers. If Company A has a best practice that Company B is missing, pair them. Allocate one week of engineering time to transfer the prompt, caching strategy, or fallback logic. Measure the impact.

4. Report to LPs. The rollup becomes a portfolio-level KPI. Report median cost per outcome, variance, and trend. This is a leading indicator of AI operational discipline across the portfolio.

5. Integrate with M&A diligence. When evaluating acquisition targets, ask: "What is their cost per outcome for their primary business process, and how does it compare to our portfolio benchmark?" If they're 30% more expensive, is it worth paying to acquire operational inefficiency? Or is it an opportunity—can we apply our best practices to their business and create value at close?

The Operating Leverage

The real operating leverage in cross-portfolio AI benchmarking is not cost reduction (though that's real). It's systematic knowledge transfer. Company A figured out how to run a $0.65/claim agent. That knowledge is now a portfolio asset. Company B, C, and D benefit from it. The next company you acquire with a similar business process inherits the playbook on day one.

This is the moat. It's not sophisticated AI—it's disciplined measurement and relentless replication.

For help building and maintaining a cross-portfolio AI economics rollup, Runrate provides a template and quarterly benchmarking dashboard as part of the PE Operating Partner program. Schedule a demo to see how the rollup integrates with your portfolio management workflow.