How AI affects EBITDA in portfolio companies, and how a CFO should model it

Where AI shows up in a portfolio company's P&L

AI moves EBITDA through two channels: a lower cost to deliver the same work, and more or faster revenue. Both are offset by new recurring costs, and both are easy to overstate in a board deck. A third item, cash from licensing the company's historical records to AI developers, is real money but normally sits outside run-rate EBITDA.

How to build an AI EBITDA bridge

Build the bridge from workflows, not from vendor benchmarks. The guide to revenue acceleration levers covers the top-line side in more depth.

1. Baseline each workflow before launch: volume, hours, cost per unit and error rate.
2. Measure adoption by named users at 60 and 120 days; seats bought are not seats used.
3. Separate capacity from savings. Freed hours reduce cost only if overtime, contractors or open roles go away, or if the same team absorbs growth that would otherwise have required hiring.
4. Net off recurring AI costs, including usage-based fees that rise as adoption succeeds.
5. Tag one-time costs such as implementation consultants, and agree with the sponsor and lenders whether they will be treated as adjustments.
6. Count new revenue only once booked, with attribution the sales team accepts.
7. Revisit every quarter and restate the bridge when an assumption fails.

Common mistakes in AI EBITDA claims

Where a data licensing payment fits

A company with years of operational records in email, chat, CRM, support, finance and engineering systems can license them to AI labs and data buyers. The company keeps ownership. Its proceeds arrive as one payment at a single all-in price, with SourceX's fee built in and no separate charges, typically within about 60 days of invoicing once a buyer has selected the data. Exclusivity for AI training usually runs for an agreed term.

For a CFO, three points matter.

• Presentation. A single license payment is typically presented as non-recurring. Keep it out of run-rate EBITDA, the management case and any covenant calculation unless the credit agreement's definitions say otherwise.
• Timing. When license revenue is recognized depends on the terms. Deloitte's ASC 606 roadmap on licenses explains the difference between a right to use intellectual property as it exists when the license is granted, recognized at a point in time, and a right to access it throughout the license period, recognized over time. Ask the company's auditors how a specific agreement should be treated before you forecast the quarter.
• Contract value is not annual revenue. Public disclosures show why the distinction matters. Reddit's February 2024 registration statement reported data licensing arrangements with an aggregate contract value of $203.0 million over two- to three-year terms, of which it expected to recognize at least $66.4 million in 2024, delivered through continuing API access and quarterly data transfers. That is an ongoing-access arrangement at a large public platform, a different shape from a one-time license of a mid-market company's historical records.

This is general information, not legal, tax or financial advice. Confirm the treatment of any license with the company's auditors and tax adviser before acting.

How to present AI effects to the board and lenders

What this means for a portfolio CFO

CFOs see the systems list, retention policies and customer contracts that decide whether a company could license its records, so they are well placed to recognize a fit. The baseline is a US company with 50+ full-time employees at peak (contractors excluded), several years of documented operations, rights to what it recorded and an executive able to authorize a deal; the who qualifies page lists the rest. The portfolio CFO referral guide explains how to raise it with your CEO and board, and the AI readiness assessment shows how one systems inventory serves both AI adoption and licensing.

If you introduce a company as a referral partner: partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company. The reward is paid only after the buyer pays and SourceX receives its fee; an introduction, meeting or signed agreement alone does not trigger payment, and no reward is guaranteed. It never reduces what the company itself receives. If you are an officer of the company you would refer, disclose that to the board and the sponsor and check your employment terms before you register.

Limits and open questions

• Vendor and survey estimates of AI savings rarely match a specific company's workflows; rely on your own baseline.
• Usage-based AI pricing changes, so model a range for software costs rather than a single number.
• Not every company can license data. Records that mainly belong to clients, data that is mostly consumer or patient information, or deleted archives rule it out.
• No license payment is certain until a buyer selects the data and pays. The explainer on whether AI data licensing is a real revenue line sets realistic expectations, and the AI value creation playbook puts both AI tracks in context.

Next step

Rebuild your AI bridge with realized and planned savings in separate columns, and add a non-recurring line for any license. If a portfolio company may qualify, register as a partner and make the introduction, or have the CEO apply at sourcex.si/apply using your referral link.