Most AI investment reviews produce a number. Dr. Jonah Tebaa's argument is that the number is usually the wrong one — not because executives are measuring carelessly, but because they are applying an industrial-era measurement instrument to a technology that doesn't behave like the machines that instrument was designed for.
In his work with enterprise teams across the MENA region and beyond, Tebaa has observed a consistent pattern: organizations deploy AI, calculate time saved, multiply by labor cost, and present the result as their AI return. It is a defensible calculation. It is also, in his view, a systematically incomplete one — and the incompleteness has strategic consequences that compound over time.
Substitution Value vs. Creation Value
Tebaa draws a sharp distinction between two types of value that AI can generate. The first is substitution value: AI performs work that a human previously performed, faster or at lower cost. Standard ROI formulas capture this well. The second is creation value: AI makes possible something that simply could not be done before — not due to budget or headcount constraints, but due to structural impossibility at the required scale or speed.
Labor-market research points the same direction: because most occupations consist of tasks that require human input, transformation of jobs, not outright replacement, is the most likely impact of generative AI, according to the International Labour Organization's 2025 index on generative AI and jobs.
His examples are concrete. A legal team that previously reviewed vendor contract compliance once per year, manually, can now monitor 300 active contracts in real time. A research function that spent weeks generating hypotheses can now walk in each morning with 200 tested alternatives ranked by AI overnight. A commercial team that sent templated outreach can now deliver genuinely adaptive communication calibrated to 50,000 individual account profiles.
None of these have a "hours saved" figure. They didn't replace existing work. They created new work that had no prior existence. Applying a substitution formula to them produces either a distortion or a zero — and a zero, in most budget cycles, means defunding.
The Two-Bucket Framework
To address this measurement gap, Tebaa has developed what he calls the Two-Bucket Framework. The premise is that before selecting any KPI, a program must be classified into one of two categories.
Bucket A — Efficiency: The program substitutes for or accelerates existing work. Appropriate metrics are time reduction, error rate, and cost per output unit. The realistic improvement expectation is 10 to 40 percent. These programs are important, fundable, and should be measured with standard rigor.
Bucket B — Capability Expansion: The program creates a capability that previously did not exist. Appropriate metrics are adoption rate of the new capability, output quality relative to the business decision it informs, and downstream outcome — revenue influenced, risk avoided, relationship depth sustained. The improvement expectation is not incremental. When the baseline is zero, the relevant question is not "how much better?" but "what is the value of this being possible at all?"
Tebaa is emphatic on one point: the two buckets require different measurement systems, and mixing them produces meaningless results. A Bucket B program evaluated with Bucket A logic will always appear to underperform, because it was never designed to replace anything. The instrument gives a wrong reading — not a low reading.
Decision Quality as a Standalone Metric
Beyond the two buckets, Tebaa advocates for tracking a third dimension that most AI program reviews omit entirely: decision quality. He argues that if AI is embedded in any workflow that involves human judgment — pricing, hiring, vendor selection, resource allocation — then the calibration, speed, and consistency of those decisions should be measured before and after deployment.
Decision quality is not easy to operationalize, but Tebaa contends it is the highest-leverage metric in an AI portfolio. Better decisions compound across every function. The absence of a baseline before deployment, he notes, is one of the most common and most costly implementation errors he sees.
The Strategic Consequence of the Wrong Instrument
For Tebaa, the measurement problem is not an analytics question. It is a strategic one. Organizations that measure only Bucket A will fund only Bucket A. Their AI portfolios will drift entirely toward cost reduction, which is a legitimate goal but a limited one. Meanwhile, competitors who measured correctly will be building capabilities — at scale, at speed — that the efficiency-only organization cannot replicate, because it never knew those capabilities were what AI was for.
His view is direct: in three years, the gap between companies that measured AI correctly and those that didn't will not appear on a cost-savings dashboard. It will appear in what each company is structurally capable of doing.
