Editor's note:Aarif Nakhooda is the Chief Financial Officer of CoreAI and Customer Success atKeystone.ai, a company headquartered in San Francisco, California, that provides global technology and consulting services to large enterprises, law firms, and governments. The views expressed in this article are solely those of the author.

In a previous article, I pointed out that traditional ROI is flawed and can even be misleading when evaluating AI investments. AI creates value not in the neat, linear way of capital projects or targeted marketing spend, but by influencing numerous micro-levers within complex systems, producing immediate and downstream effects that are difficult to isolate, quantify, and time-bound.

Since ROI falls short, what tools should CFOs adopt? My answer is simple: use AI to measure AI. If AI systems are too complex and dynamic to be evaluated with static financial models, the solution is to apply intelligent, adaptive measurement tools, such as digital twins and downstream impact analysis.

Digital Twins

Imagine trying to determine whether a single marketing touchpoint—such as an email, a promotion, or a product recommendation—led to a customer conversion. Traditional A/B testing provides a rough comparison: one group receives the touchpoint, another does not. But what if testing is infeasible, unethical, or difficult to scale? In healthcare, we cannot deny life-saving treatment to a group of patients; in marketing, turning off campaigns for half of all customers is costly and impractical.

Digital twins solve this problem precisely. A digital twin is a simulated version of a real person or real event created using AI and historical data, allowing experiments to be conducted without disrupting reality. A "customer" version receives a treatment (e.g., watching a video during a free trial), while the twin version does not. By comparing outcomes such as conversion rates, revenue, long-term engagement, and more, we can isolate the true effect of that action.

Unlike A/B testing, digital twins can scale to thousands of variables, touchpoints, and customer segments. They enable CFOs to track not only direct impacts but also long-term effects: how small interventions create ripple effects through customer behavior and business outcomes.

Many CFOs struggle with the intangible benefits of AI, such as customer engagement, content consumption, or decision speed. These behaviors seem positive, but how do you assign a dollar value to them? Digital twins help address this. Suppose a twin version streamed a piece of content during a free trial, while the original did not. Twelve months later, the twin spent more, subscribed longer, or referred others. We thus establish a causal financial link between what was once an intangible behavior and a measurable business outcome.

This transforms what CFOs might view as "soft metrics" into inputs that can be planned, predicted, and attributed.

Downstream Impact Analysis

AI behaviors often produce compounding effects over time. A customer influenced by a discount today may not just make one purchase but may completely change behavior: repurchasing sooner, upgrading later, or referring others.

Downstream impact simulation allows us to use digital twins to model these long-term evolving journeys. We can observe how a single input—such as watching a tutorial video, clicking an ad, or receiving a delivery offer—affects not just a single transaction but also the customer's lifetime value.

This type of analysis is crucial for CFOs because they need to justify not only short-term returns but also the full value creation trajectory that AI can unlock.

Entitlement Analysis

Sometimes, the right way to evaluate AI is not to compare a treatment group against a control group, but to ask: how close are we to an ideal outcome? In entitlement analysis, we define a theoretical "perfect" scenario—where every decision is optimal in terms of cost, time, or customer experience—and compare real-world performance against it. For example, in logistics, a perfect system might ship every order from the nearest warehouse at the lowest cost. This may not be achievable in practice, but it provides a benchmark. If your algorithm closes a $2 billion gap between the current state and the perfect state, then you have created tangible value.

This technique does not have to be implemented by AI, but it is an excellent way to evaluate whether AI is improving performance over time—especially in operations, fulfillment, or supply chain scenarios.

From Measurement to Action

Once CFOs can isolate and quantify the value of intangible inputs, the next step is planning. Based on these insights, finance teams can build causal models linked to specific events. For example, want to increase trial-to-paid conversion rates? Identify the drivers—video engagement, early usage, time-to-first-value—and plan to increase those behaviors. Then, forecast revenue based on the expected response rates observed in the synthetic control models.

This transforms AI from an experimental tool into a deliberate financial planning engine. You know how many touchpoints to deliver, what returns to drive, and whether you are on track—while strengthening the tight connection between marketing, product, and finance.

Editor's note:This is the second article in a two-part series. The first explained whytraditional ROI tools are not ideal for evaluating AI investments。