Six months ago, AI exposure in a portfolio review could be treated as an emerging technology question. The category was new, definitions varied, and a high-level inventory helped start the conversation. That is no longer enough for companies selling AI-enabled products. Their ongoing model cost affects what it costs to serve a customer and how gross margin behaves as the product grows.

For 2027, private-equity operating teams should bring AI COGS into the operating model before an exit process forces the company to reconstruct the history. The objective is not a universal benchmark. It is a defensible company-level baseline and a credible account of what changed.

Comparable does not mean identical

Portfolio companies differ in architecture, usage patterns, pricing, and product maturity. A single cost-per-token metric will not make them comparable. Start with consistent questions instead: what customer-facing work includes AI, what unit does the customer buy, which provider and infrastructure costs are included, and how does the resulting cost reconcile to invoices?

Then preserve the company-specific unit that makes the answer operational. For one business, that could be cost per customer. For another, cost per completed task or a priced feature. The common standard is transparency about definitions, coverage, and variance, not forced uniformity in product economics.

Operating review prompt: What changed in cost per unit since the prior review, what drove the change, and which action is now owned by the company?

Keep product COGS apart from experimentation

Research and development spend supports learning. Production workloads support a product promise. Both matter, but mixing them blurs the questions an operating partner needs to ask. Is experimentation generating a product capability? Has the capability reached customers? Is the production cost visible in gross margin, pricing, and forecasts?

Six months ago, a combined number might have been an acceptable starting point. It should not be the ending point. Establish a transition rule from experimentation to production and a named owner for classifying changes. The rule need not be perfect on day one. It needs to be explicit, reviewable, and applied consistently.

Make value creation testable

Cost optimization is not a value-creation plan unless it names the outcome to protect. A lower provider bill can be a false win if it increases manual review, reduces completion quality, or changes customer retention. A higher model cost can be rational if it supports a better-priced product or a material improvement customers value.

For each initiative, document the baseline, the expected operating change, the cost boundary, and the evidence that will confirm or reject the hypothesis. Assign a company owner. Put the result into the operating review, not just the technical backlog.

Do not wait for diligence to discover the gap

Exit readiness is an outcome of an operating record, not a last-minute formatting exercise. A buyer may challenge how a company defined AI COGS, whether the number ties to actual invoices, and whether reported margin reflects customer-facing usage. If the company cannot reproduce the answer, a confident summary will not repair the record.

Ask portfolio teams to replace the AI slide they would have shown in April with a narrower, evidenced view: defined unit, reconciled cost, clear split between product and research and development, accountable owner, and next decision. That is how AI COGS becomes an operating metric instead of a diligence surprise.