Where the AI Value Actually Is

A Framework for Assessment

· 11 min read · ai-ml , futures

Where the AI Value Actually Is: A Framework for Assessment

We’re living through the most rapid deployment of AI in history. Between 2023 and 2026, AI capabilities have moved from research labs into millions of daily workflows. The question everyone is asking - from investors to engineers to the general public - is deceptively simple:

Where is the value?

Not the hype, not the potential, not the demos. The actual, measurable value being created right now. And equally important: where is value being destroyed or wasted?

Let me map what’s actually happening, examine the evidence from multiple perspectives, and give you frameworks to assess this question yourself.

The Current Landscape: What’s Actually Being Built

To understand where value resides, we first need to see the full picture of AI deployment in 2025-2026. Let’s look at what exists, not what’s promised.

Healthcare and Life Sciences

Deployed systems serving patients:

  • Google’s diabetic retinopathy screening operates in India and Thailand, providing specialist-level diagnosis in clinics without ophthalmologists. Over 150,000 patients screened.
  • Viz.ai identifies strokes in CT scans and alerts specialists in real-time, reducing treatment time by 52 minutes on average across 1,400+ hospitals.
  • AlphaFold has predicted structures for 200+ million proteins, accelerating drug discovery research globally.
  • AI discovered halicin, a new antibiotic, by screening 107 million molecular structures - something traditional methods couldn’t achieve at this scale.

Resource investment: Medical AI attracted $29 billion in investment (2024). Training AlphaFold cost an estimated $100+ million in compute.

Measurable outcomes: Diabetic retinopathy screening shows 90%+ sensitivity. Stroke detection systems demonstrate 15-20% improvement in treatment timing. Research acceleration is harder to quantify but documented through publication citations and experimental validation.

Accessibility and Inclusion

Deployed capabilities:

Resource investment: Major tech companies invest billions in accessibility features, often at minimal or no additional cost to users.

Measurable outcomes: User surveys show 70%+ of accessibility feature users report significant quality-of-life improvements. Employment rates for people with disabilities using AI assistive tech show measurable increases.

Scientific Research

Deployed tools:

  • AI protein design tools used by thousands of researchers across biology labs globally.
  • Computational chemistry platforms accelerate materials discovery for batteries, reducing experimental iterations by 40-60%.
  • Automated microscopy analysis processes millions of cellular images, identifying patterns humans would miss.
  • Literature review systems help researchers navigate 3+ million scientific papers published annually.
  • AI processes CERN collision data, identifying rare particle interactions in petabytes of information.

Resource investment: Research institutions spend $5-10 billion annually on AI research infrastructure.

Measurable outcomes: Materials discovery timelines compressed from years to months. Fusion reactor optimization shows 10-15% efficiency gains. New antibiotic candidates identified that traditional screening missed.

Climate and Energy

Deployed systems:

Resource investment: Climate AI companies raised $8+ billion (2024). Data center optimization required minimal additional compute - AI optimizing AI’s own infrastructure.

Measurable outcomes: Documented energy savings in TWh. Crop yield improvements measured in controlled studies. Forecast accuracy improvements verified against historical data.

Productivity and Business Tools

Deployed applications:

  • GitHub Copilot and similar coding assistants used by 10+ million developers, with studies showing 55% faster task completion.
  • Customer service automation handles 60-70% of routine inquiries, reducing wait times and costs.
  • Content creation tools produce billions of images, videos, and text pieces monthly.
  • Marketing optimization systems manage trillions of dollars in advertising spend.

Resource investment: Productivity AI represents $50+ billion in annual spending. Training large models costs $100+ million; inference costs have dropped 90% since 2023.

Measurable outcomes: Developer productivity studies show 20-50% time savings on routine tasks. Customer satisfaction scores for AI-assisted service match or exceed human-only baselines in structured scenarios. Marketing ROI improvements of 10-30% documented by adopters.

Education

Deployed platforms:

  • Khan Academy’s Khanmigo provides AI tutoring to millions of students.
  • Duolingo’s AI adapts language learning to individual patterns, serving 50+ million active users.
  • AI tools assist teachers in underserved areas with lesson planning and student assessment.

Resource investment: EdTech AI attracted $4+ billion in investment (2024).

Measurable outcomes: Learning outcome studies show 10-25% improvement in certain contexts, though effects vary significantly by implementation quality and student demographics.

Different Perspectives on Value

The same AI application can look completely different depending on who’s evaluating it and what they value. Let’s examine multiple viewpoints honestly.

The Economic Perspective

Value creation indicators:

  • AI sector represents $200+ billion in annual economic activity (2025).
  • Productivity gains estimated at $2-4 trillion in potential economic value by 2030 (McKinsey).
  • Companies adopting AI report significant cost reductionsin optimized processes.
  • New businesses enabled by accessible AI tools create jobs and tax revenue.

Value destruction indicators:

  • Job displacement in customer service, content creation, and routine coding - estimated 2-5 million roles at risk by 2027.
  • Resource consumption: data centers use 1-2% of global electricity , growing 20-30% annually.
  • Market concentration: 90%+ of AI capability controlled by handful of companies - anticompetitive effects.
  • Opportunity cost: $100 billion invested in AI could have funded alternative research or development.

The Environmental Perspective

Costs:

  • Training GPT scale models: 1,000-5,000 MWh per training run.
  • Global AI inference: estimated 100+ TWh annually by 2026.
  • Water consumption for data center cooling: billions of liters annually.
  • Hardware manufacturing: rare earth minerals, electronic waste.

Benefits:

  • AI optimization reducing energy consumption in other sectors: estimated 100-200 TWh potential savings.
  • Climate modeling and prediction improvements enabling better resource allocation.
  • Materials discovery accelerating clean energy technology development.
  • Agricultural optimization reducing water and fertilizer waste.

Net assessment: Contested. Some analyses suggest AI’s optimization benefits exceed its consumption; others suggest net negative when accounting for full lifecycle.

The Scientific Perspective

Research acceleration:

  • Protein structure prediction enabling drug discovery that was previously impossible - paradigm shift.
  • Materials discovery speed increasing 10-100x for certain applications - genuine acceleration.
  • Literature synthesis across disciplines humans couldn’t integrate - new connections.
  • Experiment optimization reducing trial-and-error - resource efficiency.

Research concerns:

  • Reproducibility: AI systems are often proprietary black boxes - scientific method violation.
  • Hallucination risk: AI confidently generates false information - epistemic pollution.
  • Homogenization: Over-reliance on similar AI systems may narrow research directions - innovation risk.
  • Funding distortion: AI hype attracts resources from other valuable research areas - opportunity cost.

Frameworks for Assessing Value

How do we make sense of these conflicting perspectives? Here are three frameworks that might help.

Framework 1: Value Multiplication vs. Value Extraction

Value Multiplication occurs when AI enables capabilities that were previously impossible or inaccessible:

  • Blind person accessing visual information through AI description
  • Researcher discovering protein structures no human could compute
  • Farmer in rural area getting specialist agricultural advice
  • Language barrier breaking enabling cross-cultural collaboration

Value Extraction occurs when AI optimizes existing processes without creating new capability:

  • Advertising optimization making marketing slightly more effective
  • Content generation flooding platforms with material of questionable utility
  • Automated calls competing with automated call screening
  • AI systems optimizing against other AI systems in zero-sum games

The multiplication/extraction distinction isn’t absolute - a tool can do both - but it’s a useful lens.

Framework 2: Resource Proportionality

Does the value created justify the resources consumed?

High proportionality examples:

  • AlphaFold: $100M compute investment enabling $billions in drug discovery acceleration and saving researcher-decades of time.
  • Accessibility features: Minimal marginal cost serving hundreds of millions of users with life-changing impact.
  • Medical diagnosis in underserved areas: High compute cost enabling healthcare access worth multiples in DALY (Disability-Adjusted Life Years).

Low proportionality examples:

  • Training massive models for marginal chatbot improvements: $100M for capabilities users marginally prefer.
  • Generating images for social media that get viewed for 3 seconds: Compute cost exceeding value to viewer.
  • Automated email marketing at scale: Resources consumed to generate messages most recipients ignore.

Contested examples:

  • Coding assistants: High productivity gains but displacing junior developers - net value depends on what those developers do next.
  • Content generation: Enables creators but floods platforms with noise - depends on quality filtering.

Framework 3: Time Horizon and Compounding

Some value is immediate; some compounds over decades.

Immediate value:

  • Accessibility features provide value the moment they’re used.
  • Medical diagnosis saves a life today.
  • Translation enables a transaction now.

Compounding value:

  • A child whose life is saved grows up, contributes to community, helps others - decades of impact.
  • A scientific discovery enables follow-on research - cascading breakthroughs.
  • A student educated with AI tutoring develops capabilities used throughout career - lifetime value.
  • Infrastructure built today enables applications we haven’t imagined - option value.

Decaying value:

  • Optimized marketing email has value only until read, then zero.
  • Trending AI-generated content has value measured in hours or days.
  • Process optimization valuable until next optimization cycle replaces it.

What We Know and What We Don’t

After examining the evidence, here’s what seems clear and what remains uncertain.

What We Know

AI is creating measurable value in specific domains:

  • Healthcare: Lives saved, diseases detected, research accelerated - documented outcomes.
  • Accessibility: Hundreds of millions of people accessing previously unavailable capabilities - user-reported impact.
  • Scientific research: Discovery timelines shortened, new compounds identified - verified results.
  • Productivity: Time saved on routine tasks - study-measured gains.

AI is consuming significant resources:

  • Energy: 100+ TWh annually and growing - measured consumption.
  • Capital: $200+ billion in annual economic activity - tracked investment.
  • Human attention: Billions of hours engaging with AI systems - usage data.
  • Environmental: Water, rare earths, electronic waste - physical footprint.

Distribution is highly uneven:

  • Geographic: 80%+ of AI capability deployed in wealthy nations - deployment data.
  • Economic: Premium capabilities accessible primarily to those who can pay - pricing analysis.
  • Linguistic: Bias toward English and major languages - training data composition.

Both benefits and harms are real:

  • Job displacement is happening in measurable sectors - employment data.
  • New opportunities are also being created - startup formation, new roles.
  • Quality of life improvements for some populations - accessibility studies.
  • Information quality degradation in some contexts - misinformation research.

A Personal Synthesis (Not a Conclusion)

After examining the evidence, here’s where I land - acknowledging this is one interpretation among many valid ones.

The value is real in specific applications. The blind student accessing visual content, the stroke patient getting faster treatment, the researcher discovering new protein structures - these aren’t hypothetical. They’re happening, and they’re valuable by almost any reasonable definition.

The waste is also real. We’re running arms races where AI systems compete with each other, consuming resources for zero-sum or negative-sum outcomes. The Red Queen’s race is not a metaphor - it’s actually happening in advertising, content generation, and automated communication.

The distribution matters as much as the aggregate. Even if AI creates net positive value globally, if that value concentrates among the already-privileged while costs are borne broadly, we should care about that.

We’re early in understanding trade-offs. The technology is moving faster than our ability to measure impacts. We’re making decisions with incomplete information, and humility about what we don’t know seems appropriate.

Questions Worth Asking

Rather than claiming to know where value is, here are questions that might help us find it:

  • What do we actually value, and is AI helping us achieve it?
  • Are we using AI to do things better, or just to do more things?
  • What are we optimizing for, and is that what we want to optimize for?