The PE Operating Partner's AI Cost Attribution Playbook

Operating partners often inherit a portfolio company where AI spend is buried across credit cards, cloud accounts, and vendor contracts. Nobody can answer the question: "What did that AI agent actually cost to run?" This playbook walks through a six-step operating partner workbook to establish cost attribution from first principles, move the company up the 5-Stage AI Cost Maturity Curve, and create the cost-per-outcome KPI that ties AI economics to value creation.

Step 1: Collect and Consolidate All AI Spend (Week 1-2)

Your first move is to gather every dollar the company spends on AI. This is harder than it sounds. AI spend hides in six places: API subscriptions to OpenAI and Anthropic; cloud infrastructure bills (AWS, Azure, GCP) where LLM inference is buried alongside regular compute; third-party SaaS tools with AI features baked in; in-house infrastructure costs (GPU servers, model hosting); human labor for prompt engineering and evaluation; and shadow charges (individual subscriptions on personal credit cards, team trial accounts left running after the pilot ended).

Create a simple spreadsheet with columns for: vendor/service, monthly cost, business unit served, cost classification (API, infrastructure, labor, SaaS, or shadow). Send it to the CFO, CTO, and each business unit head. You're looking for completeness, not perfection. Budget one week to gather and one week to reconcile misclassifications.

Common discoveries at this stage: a $15K/month Anthropic subscription nobody knew was running on an old API key. A $8K/month OpenAI account left over from a pilot that ended two years ago. $22K/month in AWS GPU costs for a model fine-tuning exercise that the data science team abandoned. Add these up: you've found $45K/month of waste without optimizing a single inference call.

Deliverable: A consolidated AI spend register showing total monthly spend, vendor breakdown, and business unit allocation (rough—it's okay if it's 80% correct). This register is your baseline. You're now at stage 1 or 2 on the Maturity Curve (Invisible or Tracked).

Step 2: Map AI Spend to Business Outcomes (Week 3-4)

Now that you have the spend number, ask: What is this money actually buying? Which business processes does the AI support?

For each major vendor (OpenAI, Anthropic, your GPU infrastructure), identify the agents and workloads. A contact center company might have: (1) an AI agent handling tier-1 customer support (redirecting easy questions, escalating to humans for complex issues); (2) another agent processing refund requests; (3) a third agent mining support tickets for product feedback. Each agent has a cost and an outcome.

Create a second spreadsheet with columns for: agent name, vendor used, monthly cost, business outcome (e.g., "resolving support tickets"), volume (e.g., 10,000 tickets/month), cost per outcome (monthly cost ÷ volume). Fill this in for your top 5 agents.

Real example for a mid-market claims processor:

| Agent | Vendor | Monthly Cost | Outcome | Volume/Month | Cost Per Outcome | | --- | --- | --- | --- | --- | --- | | Claims Triage | Claude (via API) | $8,500 | Claims adjudicated | 12,000 | $0.71 | | Fraud Detection | GPT-4 | $6,200 | Claims flagged for review | 8,000 | $0.775 | | RCM Coding | Internal Llama 2 | $4,100 | Claims coded | 5,500 | $0.745 | | Appeals Processing | Anthropic Batch | $3,200 | Appeals processed | 4,000 | $0.80 |

This table immediately shows you where value lives. Claims Triage is $0.71 per outcome; Appeals is $0.80. If volume grows 20% next quarter, which agent will you invest in optimizing? The highest-volume one. You're now allocating AI spend to business outcomes. You're moving to stage 3 on the Maturity Curve (Allocated).

Deliverable: A cost-per-outcome table for top 3-5 agents, showing vendor, volume, and unit economics. This is the KPI that ties AI spend to value.

Step 3: Implement the AI Workforce P&L (Week 5-6)

Here's where most operating partners get stuck: they treat AI like infrastructure. AI should be treated like payroll.

The AI Workforce P&L framework says: each AI agent needs four things. First, a timecard—what work did it do (volume, resolution rate, quality metrics)? Second, an attribution path—which P&L line does it serve (gross margin, operating expense, cost of goods sold)? Third, a 1099-vs-W-2 classification—is it a third-party API (contractor, variable cost) or internal infrastructure (employee, fixed cost)? Fourth, a clear retirement trigger—under what conditions do we turn this agent off?

Build this out for your top agent. Say your Claims Triage agent processes 12,000 claims/month and costs $8,500/month. The agent is a Claude API (third-party, variable cost). It serves the claims processing business unit (allocate to COGS, since it's directly tied to claims adjudication). Timecard: 12,000 claims processed. Retirement trigger: if claims volume drops below 8,000/month or if cost per outcome rises above $0.95, we evaluate replacement with a cheaper model or rule-based system.

This discipline forces you to treat AI like you treat headcount. A CFO would never run a contact center without knowing exactly how many people you have, which team they serve, how much they cost, and when you'd add or remove headcount. The AI Workforce P&L forces the same rigor on AI agents.

Deliverable: An AI Workforce P&L roster showing agent name, business unit served, cost classification, timecard metrics (volume, resolution rate), and retirement trigger. This becomes your board-ready AI governance document.

Step 4: Build the Cost-Per-Outcome Dashboard (Week 7-8)

Once you have cost-per-outcome for your top agents, build a simple dashboard. This is the operating partner's primary tool for monitoring and optimization.

The dashboard shows three things:

1. Monthly cost per outcome for each agent, trended over three months. Is cost per outcome trending down (good—efficiency improving) or up (bad—efficiency degrading)?
2. Volume and headcount leverage. As volume grows, cost per outcome should decline. If volume is flat and cost per outcome is rising, the agent is drifting toward inefficiency.
3. Anomalies. If an agent's cost per outcome spikes 20% month-over-month without a corresponding change in volume or model, flag it. Something broke (a prompt optimization failed, a fallback mechanism kicked in, or the upstream API raised pricing).

This dashboard lives in your CFO's hands. It's the conversation every month: "Cost per outcome is improving on the Claims Triage agent, flat on Fraud Detection, and degrading on Appeals. Where should we focus engineering effort?"

You're now at stage 4 on the Maturity Curve: Optimized. AI spend is tied to specific work items with cost-per-outcome KPIs.

Deliverable: A monthly cost-per-outcome dashboard for all top agents. Publish it monthly to the executive team. Tie optimization work to specific cost-per-outcome targets.

Step 5: Benchmark Across the Portfolio (Month 3+)

Once you have attribution dialed at one portfolio company, replicate it across 3-5 more. The power of PE is the ability to benchmark.

A healthcare company processing 50,000 claims/month with a cost per claim of $0.71 is efficient. A sister company processing 40,000 claims/month with cost per claim of $1.10 is not. Why? Is the second company using a different model? Are they reprocessing claims more often (more retries)? Do they have higher human-review rates? The operating partner's job is to find the delta and transfer the playbook.

Create a cross-portfolio rollup. Take your top outcome metric (claims processed, tickets resolved, deals underwritten) and collect cost per outcome across the portfolio. Real example:

| Company | Business Outcome | Monthly Volume | AI Cost | Cost Per Outcome | | --- | --- | --- | --- | --- | | Company A (Healthcare) | Claims processed | 50,000 | $35,500 | $0.71 | | Company B (Healthcare) | Claims processed | 40,000 | $44,000 | $1.10 | | Company C (Contact Center) | Tickets resolved | 25,000 | $8,500 | $0.34 | | Company D (Contact Center) | Tickets resolved | 18,000 | $9,200 | $0.51 |

Now you can see: Company C is 34% more efficient than Company D on cost per outcome. Why? Interview the teams. Transfer the practice. You've now built a repeatable operating playbook that scales across the portfolio.

Deliverable: A quarterly cross-portfolio rollup table. Identify outliers. Document why. Transfer best practices from high-efficiency to low-efficiency companies.

Step 6: Govern and Report (Ongoing, Monthly)

At stage 5 (Governed), you've built SLOs, anomaly detection, and board-grade reporting around AI economics.

Set targets: by year 2 of the hold, what cost per outcome are you aiming for? A contact center aiming to reach Intercom Fin efficiency might target $0.75 per ticket (Intercom Fin is ~$0.99; Klarna is $0.19 but that's a higher-volume, older playbook). Set a cost threshold: if monthly cost per outcome exceeds $0.85, trigger a review. Assign ownership: the CFO owns cost anomaly detection, the CTO owns prompt optimization and vendor evaluation, the business unit leader owns volume and quality metrics.

Build monthly board reporting: (1) cost per outcome for each agent, trended; (2) cross-portfolio outliers; (3) variance in cost per outcome across business units and portfolio companies; (4) anomalies and investigations; (5) optimization work in progress and expected savings.

This is the discipline that moves a company from "we have AI" to "we have proven AI economics with governance rigor."

Deliverable: Monthly board-grade AI economics report tied to cost per outcome, anomaly alerts, and cross-portfolio benchmarking.

The Operating Partner's Playbook at a Glance

Week 1-2: Consolidate spend → move to stage 1-2 (Invisible/Tracked). Week 3-4: Map to business outcomes → move to stage 3 (Allocated). Week 5-6: Build AI Workforce P&L → move toward stage 4 (Optimized). Week 7-8: Cost-per-outcome dashboard → stage 4 confirmed. Month 3+: Cross-portfolio rollup → unlocking portfolio-wide leverage. Ongoing: Board reporting and governance → stage 5 (Governed).

This six-step playbook typically takes 2-3 months to implement at a single company and another 1-2 quarters to replicate across the portfolio. The payoff is consistent: operating partners who implement this playbook capture $50K-$500K in cost visibility and optimization per portfolio company, unlock repeatable cross-portfolio practices, and create a defensible exit narrative around proven AI economics.

For the full operating partner workbook with templates and rollup scaffolding, request the PE Field Guide.