What measures can be used for RoI of AI? The article [[Why Utah Uses More AI Than California?]] makes some attempt (unsuccessfully, I would argue) to identify some ways of measuring it:
**Enterprise ROI**
– Claim: 95% of organisations see _zero_ measurable ROI from GenAI despite $30–40B investment.
– Source: MIT report, _The GenAI Divide: State of AI in Business 2025_.
**Consumer monetisation**
– Claim: Only ~3% of consumer AI users pay; for ChatGPT specifically 2–5%.
– Source: OpenAI usage data (Sept 2025) and Menlo Ventures survey (2025).
**Work v personal use**
– Claim: ChatGPT usage has moved from ~50/50 work–personal in June 2024 to ~70% personal in Sept 2025.
– Source: OpenAI usage report (2025).
**Augmentation vs automation**
– Claim: Wealthier countries lean toward augmentation over automation.
– Source: Anthropic _AI Usage Index_ (2025).
– Interpretation (author’s): this shows “delegation failed.”
**"Productive use" vs “comfort use”**
– Claim: Growth area is “expressing” (emotional/sounding board use) vs “asking” or “doing.”
– Source: OpenAI usage report (2025).
– Interpretation (author’s): this is non-productive use.
My thoughts on this:
**Blurring consumer and enterprise evidence**
– The article collapses two distinct issues: enterprise ROI (MIT) and consumer willingness-to-pay (OpenAI, Menlo). They are different dynamics and should be analysed separately.
**Misattributed motivation**
– Author wrongly applies organisational motivation (economic output) to individuals. Employees adopt AI for ease of use, relief from effort, or job interest; organisations adopt AI for measurable output gains.
**Expressing as productivity**
– Article dismisses expressive use (e.g. "I am feeling overwhelmed with my work tasks, help me order my thoughts") as “comfort use” rather than "productive use". I would argue that augmented cognition (clarifying thinking, reducing overwhelm) is a legitimate productivity gain, though not yet captured in ROI frameworks.
**Narrative vs data**
– The “delegation failed” line is not from Anthropic, but the author’s gloss. Other explanations exist (labour costs, skills, incentives in different economies).
**Lack of a framework**
– None of the sources provide a coherent framework for measuring AI ROI. The article inherits this weakness, critiquing without offering structured alternatives.