AI Value Assessment — Lesson 5 of 10
If cost reduction is the straightforward chapter of AI ROI, revenue growth is the contested one. When an AI recommendation engine increases average order value by 12%, the data science team claims credit. So does the product team that designed the user interface. So does the marketing team whose campaigns brought the customers to the site in the first place. And so does the operations team whose fulfilment speed enabled the five-star reviews that drive repeat purchases.
Revenue attribution for AI is inherently imperfect. The goal is not to achieve scientific precision — that is impossible in a complex business system — but to establish directionally accurate measurements that enable informed investment decisions. This lesson provides the practical frameworks for doing so across the three primary AI revenue mechanisms: personalisation, pricing optimisation, and new product discovery.
Revenue attribution for AI should follow the "preponderance of evidence" standard used in civil law, not the "beyond reasonable doubt" standard used in criminal law. You need to demonstrate that AI more likely than not contributed a quantifiable amount to revenue growth — not prove it to absolute certainty. A/B testing is the gold standard, but matched cohort analysis and incremental lift modelling provide credible alternatives when controlled experiments are impractical.
The Three Revenue Mechanisms
AI drives revenue growth through three distinct mechanisms, each requiring a different attribution methodology.
Mechanism 1: Personalisation and Customer Experience
AI personalisation encompasses product recommendations, dynamic content, personalised search results, targeted offers, and adaptive user interfaces. The revenue impact flows through three channels: higher conversion rates, increased average order values, and improved customer lifetime value.
A/B Testing: The Gold Standard
The most rigorous attribution method is a randomised controlled experiment. Randomly assign customers to two groups: one receives the AI-personalised experience (treatment), the other receives the default experience (control). Measure the revenue difference between groups over a statistically significant period.
A/B Test Design for AI Revenue Attribution
| Design Element | Best Practice | Common Mistake |
|---|---|---|
| Sample size | Minimum 10,000 per group for revenue metrics | Using conversion-optimised sample sizes for revenue tests (different distributions) |
| Duration | At least 2 full purchase cycles (varies by industry) | Ending tests too early on significance |
| Segmentation | Stratify by customer value tier | Treating all customers as homogeneous |
| Metric | Revenue per visitor, not conversion rate alone | Reporting conversion without revenue impact |
| Holdout | Maintain a permanent 5% holdout group | Eliminating the control group after initial test |
The permanent holdout group is critical. Without it, there is no ongoing measurement of the AI system's contribution. Over time, as the AI model is updated and retrained, the holdout group provides continuous evidence of incremental value.