David Mkhitaryan profile photo

David Mkhitaryan

Venture capital analyst at Innovis VC specializing in AI/ML startup evaluation. Engineering background enables deep technical due diligence.

AI adoption is rising fast - but ROI is not evenly distributed

4 minute read

Published:

AI adoption is rising quickly

But adoption alone is a weak proxy for economic value. The more relevant question is whether AI changes measurable operating performance - not whether a team has a tool in hand, but whether that tool moves a number that matters.

This short research note pulls together evidence from the Stanford AI Index, McKinsey, NBER / Quarterly Journal of Economics, the Economic Journal, the IFR World Robotics report, and the European Commission. The pattern across all of them is consistent.

A six-slide visual version of this note - the same one shared on LinkedIn:

Slide 1 of 6: AI ROI is not evenly distributed - value appears when AI meets four conditions (specific workflow, proprietary data, clear metrics, process change).
Slide 2 of 6: Where AI shows measurable operating impact - a 2x2 map of measured operating impact versus a firm's control over workflow, data, and assets.
Slide 3 of 6: The strongest evidence for AI ROI is task-level and workflow-level.
Slide 4 of 6: Most companies use AI, but most cannot yet measure the EBIT impact.
Slide 5 of 6: The AI economy is splitting by strength.
Slide 6 of 6: Does AI change a measurable production loop? The takeaway question, with the research base.

Swipe or scroll sideways through all six slides

AI ROI is not evenly distributed

The strongest, most defensible cases for return on AI show up under a specific set of conditions. AI delivers measurable value when it is embedded into operational reality, not bolted on as a feature:

  1. A bounded workflow - a defined task with clear inputs and outputs, not an open-ended “assistant.”
  2. Firm-specific data - proprietary context the model cannot get anywhere else.
  3. Measurable outputs - cost, throughput, quality, or cycle time that can be tracked before and after.
  4. Process change around the tool - the operating procedure is rebuilt around the AI, not just supplemented by it.

Where these four conditions hold, the productivity and quality gains in the literature are real and repeatable. Where they don’t, “AI adoption” tends to stay a line item rather than a performance driver.

Why this matters for venture capital

This distinction is where a lot of AI investing gets decided.

Many AI companies can show product usage. Far fewer can show that their product changes a production loop - that it reduces cost, increases throughput, improves quality, or shortens a measurable process. Usage tells you people opened the tool. A changed production loop tells you the business runs differently because of it.

That gap is the whole game. The key diligence question is not:

“Does this company use AI?”

It is:

“Does AI change a measurable production loop?”

That is the line where AI stops being a software feature and becomes economic infrastructure - and it’s where durable margins, defensibility, and pricing power tend to come from.

When evaluating AI companies, what comes first?

The same four conditions map onto what I look at first in an AI company:

  • Model quality - necessary, rarely sufficient, and increasingly commoditized.
  • Workflow control - does the product own a bounded, repeatable process?
  • Proprietary data - is there a data position competitors can’t replicate?
  • Measurable ROI - can the customer point to a number that moved?

My weighting leans toward the last three. Model quality is the price of entry; workflow control, proprietary data, and measurable ROI are what compound.

I’d be interested to hear how investors, founders, and operators think about this distinction. When you evaluate AI companies, what do you look at first?

Full research note

A formatted version of this note, with the underlying evidence, is available as a PDF:

Sources

  1. Stanford HAI, AI Index Report 2026, Economy chapter — PDF
  2. Stanford HAI, AI Index Report 2026, Medicine chapter — link
  3. Stanford HAI, AI Index Report 2025, Technical Performance chapter — link
  4. Brynjolfsson, Li & Raymond, “Generative AI at Work,” NBER Working Paper 31161 — link
  5. Brynjolfsson, Li & Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025 — DOI
  6. McKinsey, “The State of AI: How organizations are rewiring to capture value,” 2025 — link
  7. Koch, Manuylov & Smolka, “Robots and Firms,” The Economic Journal — DOI
  8. International Federation of Robotics, World Robotics 2025 / China robotics strategy — link
  9. European Commission, AI Continent Action Plan — link

#AI #VentureCapital #AIAdoption #DueDiligence #Productivity #Robotics

Original LinkedIn post: AI ROI - what can be measured (with the full carousel and discussion)

About the Author: David Mkhitaryan is a Venture Capital Analyst at Innovis VC specializing in AI/ML startup evaluation and technical due diligence, with an engineering background and a focus on the Berlin startup ecosystem.