AI Series: The Upside
“The real problem of humanity is the following: we have Palaeolithic emotions, medieval institutions, and god-like technology.”
— E.O. Wilson, Biologist, 2009Wilson’s observation was about nuclear weapons. It applies with equal force to AI. The technology is arriving faster than the institutions designed to govern it. But before examining the risks — which Part III does in full — it is worth taking seriously what AI can actually do for human civilisation when it works as intended.
The Productivity Case: Evidence and Honest Caveats
The headline productivity numbers for AI are compelling but must be read carefully. The evidence for transformative productivity gains is strong at the task level and weak at the macroeconomic level — which is exactly the pattern we should expect at this stage of diffusion. When electrification arrived in factories in the 1890s, individual electric motors were clearly more efficient than the steam systems they replaced. The economy-wide productivity boost did not appear in the statistics until the 1920s, when factory layouts had been fully redesigned around the new technology. The question is not whether AI can make individual tasks faster — it demonstrably can. The question is whether that task-level efficiency has yet propagated into organisational restructuring sufficient to show up in national output data.
The two strongest pieces of evidence for near-term impact are: first, Goldman Sachs’ Q4 2025 earnings analysis which found median productivity gains of approximately 30% in organisations that have deeply integrated AI into software development and customer service workflows[1]; and second, multiple controlled studies showing measurable output improvements for knowledge workers using AI assistants — GitHub Copilot studies showing 55% faster task completion for coding exercises, and Stanford/MIT research showing 14–15% output increases for customer service workers with AI assistance.
Goldman’s 2023 foundational estimate — a 7% global GDP uplift equivalent to nearly $7 trillion — rests on three mechanisms: labour cost savings from automation, new job creation in AI-enabled roles, and a productivity boost for non-displaced workers who are augmented by AI tools.[2] McKinsey’s higher-end estimate of $2.6–$4.4 trillion in annual value focuses on the use cases where impact is most concentrated: customer operations, software engineering, marketing, and R&D.[3]
The historical analog from electricity and computing suggests the economy-wide productivity inflection point arrives when roughly 50% of businesses have adopted the technology and — crucially — have redesigned their workflows around it. We are at maybe 15–20% on the former and far less on the latter. This implies the large majority of AI’s measurable economic impact is still ahead of us, not behind us. Investors pricing AI purely on current productivity statistics are looking at the wrong metric.
The Healthcare Revolution: From Diagnosis to Drug Discovery
Healthcare is the sector where AI’s positive case is most empirically grounded and the stakes are highest. Three distinct impact channels are worth separating: diagnostics and clinical decision support; drug discovery and development; and pandemic preparedness and population health.
AI diagnostic systems have reached or exceeded specialist-level accuracy on specific imaging tasks in controlled settings. In diabetic retinopathy screening, Google’s DeepMind system achieved 94.5% sensitivity and 98.1% specificity — both higher than the average ophthalmologist in the study. AI pathology systems for skin cancer detection, mammography screening, and chest X-ray interpretation have produced similarly strong results in peer-reviewed trials.
The key word is “specific”: these systems perform well on the narrow task they were trained for and less reliably outside it. A chest X-ray AI that excels at pneumonia detection may not handle rare presentations that an experienced radiologist would catch through pattern recognition built over years of varied exposure. The correct framing is not “AI replaces radiologists” but “AI-augmented radiologists can process more cases, at greater consistency, with fewer errors of fatigue.” The FDA had approved over 500 AI-enabled medical devices as of 2023, with approvals accelerating.
Clinical decision support — AI systems that alert physicians to drug interactions, flag deteriorating vital sign trends, or identify sepsis onset before it is clinically obvious — is perhaps the more immediately impactful application. Studies of AI-assisted sepsis detection protocols have shown material reductions in mortality rates when integrated into hospital workflows.
DeepMind’s AlphaFold represents one of the most consequential scientific breakthroughs of the past fifty years. Protein structure prediction — determining the three-dimensional shape a protein folds into from its amino acid sequence — had been an unsolved problem since Anfinsen’s 1961 Nobel-winning work established that structure determines function. AlphaFold effectively solved it, achieving accuracy comparable to experimental methods at a fraction of the cost and time. Its database of 200 million predicted protein structures, made freely available, has been accessed by researchers in 190 countries and has already accelerated work on antibiotic resistance, malaria vaccines, and cancer biology.
The broader drug discovery pipeline is being transformed. AI systems can now screen billions of molecular candidates against target proteins computationally — a process that previously required physical synthesis and testing of each candidate. Insilico Medicine brought an AI-designed drug into Phase II clinical trials. BenevolentAI has used AI to identify baricitinib as a COVID-19 treatment candidate (subsequently validated in trials). The traditional drug development timeline of 10–15 years and $2.6 billion per approved drug is the target for compression. Even moderate improvement — reducing failures in late-stage trials, where costs are highest — would represent enormous economic and human value.
Mental health is both a large unmet need and a domain where AI can expand access materially. The WHO estimates a global shortage of 1.18 million mental health professionals. AI-based triage, psychoeducation, and CBT-adjacent interventions (apps like Woebot, clinical tools being developed by healthcare systems) cannot replace human therapy but can serve the large proportion of people with mild to moderate symptoms who currently receive no support — not because treatment is unavailable in principle, but because access, cost, and stigma create barriers that a private AI interaction lowers.
Pandemic preparedness is an underappreciated AI application. AI genomic surveillance systems can detect novel pathogen variants weeks earlier than traditional epidemiological reporting. During COVID-19, AI systems at BlueDot and HealthMap flagged unusual pneumonia cases in Wuhan before WHO official notification. AI-accelerated vaccine development — exemplified by Moderna’s mRNA design tools — could compress the response window from years to months for future pandemics.
Scientific Acceleration: Compressing the Innovation Cycle
Beyond healthcare, AI is beginning to function as an accelerant of the broader scientific enterprise — not by replacing human scientists but by dramatically expanding the hypothesis space they can explore. The implications for climate technology, materials science, and fundamental physics are material.
The common thread across these domains is that AI excels at pattern recognition in high-dimensional data — precisely the challenge that creates bottlenecks in scientific research. The bottleneck in drug discovery is not lack of scientific ideas but the inability to evaluate millions of molecular candidates experimentally. The bottleneck in materials science is synthesis time. The bottleneck in climate modelling is compute time. AI attacks all three.
Democratisation of Expertise: The Access Dividend
Perhaps the most underappreciated positive externality of AI is what it does to access to expertise. For the 8 billion people who are not in the top decile of income in their country, access to high-quality legal advice, medical second opinions, financial planning, and specialised educational support has always been rationed by cost and geography. AI changes that equation materially.
Legal & Financial Access
Access to justice is among the most persistent inequalities in developed democracies — not merely in the Global South. In the United States, roughly 80% of low-income people with civil legal needs receive inadequate or no legal help (Legal Services Corporation, 2022). AI legal tools capable of drafting documents, explaining rights, and identifying relevant case law do not replace litigation representation, but they substantially close the information gap between a layperson and a trained attorney for the large proportion of legal situations that do not require courtroom advocacy.
Financial planning has a similar structure. The advice provided by a fee-based wealth manager — goal-setting, tax optimisation, asset allocation, estate planning — is genuinely valuable but only accessible above a net worth threshold that excludes the majority. AI financial planning assistants can deliver a credible version of that service at near-zero marginal cost. The implications for wealth accumulation across income distribution are potentially significant, though regulatory barriers to AI financial advice (particularly around personalised recommendations) remain real.
Education & the Personalised Learning Dividend
The global teacher shortage — estimated at 44 million by UNESCO — creates a structural access constraint in education that will not be solved by training more teachers in time for current student cohorts. AI tutoring systems offer a partial solution: adaptive, patient, available at 2am, capable of explaining the same concept twelve different ways until it lands. Khan Academy’s Khanmigo has demonstrated measurable improvements in student outcomes in controlled settings. The democratisation of Socratic tutoring — historically available only to the wealthy through private instruction — is a genuine positive externality of AI deployment.
| Service Category | Pre-AI Access Model | AI-Enabled Access | Primary Beneficiary |
|---|---|---|---|
| Legal Advice | $250–500/hr attorney; income-rationed | AI legal tools: free or low cost for document drafting, rights explanation | Low-to-middle income individuals |
| Financial Planning | Fee-based advisors: ~$5,000/yr; minimum asset thresholds | AI financial planning: goal-setting, tax optimisation, portfolio guidance | Mass market / underbanked |
| Medical Second Opinion | Specialist referrals: weeks of wait, co-pays, geographic barriers | AI diagnostic support: instant symptom assessment, triage guidance | Rural / underserved communities |
| Private Tutoring | $60–150/hr tutors; available to affluent households | AI tutors: adaptive, 24/7, curriculum-aligned, near-free | Students in under-resourced schools |
| Translation / Language | Professional translators: $0.10–0.25/word; legal/medical require certified | High-quality real-time translation across 100+ languages | Immigrants, non-English speakers, SMEs |
| Mental Health Support | Therapists: $150–300/session; 4–6 week wait times; geographic scarcity | AI triage, psychoeducation, CBT tools: accessible, low-stigma | Mild-to-moderate symptoms; unserved populations |
The access dividend is real but should not be overstated. AI legal tools do not replace courtroom representation. AI medical assistants generate liability and accuracy concerns in clinical settings. AI financial advice has regulatory constraints that limit personalised recommendations. The democratisation of expertise is genuine, but it operates at the margin — reducing information asymmetry and improving access to standardised guidance — rather than fully substituting for credentialed professionals in high-stakes situations.
Error Reduction, Decision Quality & Scenario Analysis
A category of AI benefit that receives less attention than productivity is error reduction in domains where mistakes are costly — or fatal. Cognitive fatigue, attention bias, and pattern-matching limitations are well-documented sources of human error in medicine, aviation, financial risk management, and infrastructure maintenance. AI systems that do not tire, do not anchor on prior diagnoses, and can process far more variables simultaneously are structurally better suited to these detection tasks.
Finance: Risk, Fraud & Scenario Analysis
In financial services, AI fraud detection systems now operate at a scale and speed that human analysts cannot approach — processing millions of transactions in real time, identifying anomalous patterns that would be invisible in manual review. Major card networks report fraud detection improvements of 30–50% following AI system deployment, with lower false positive rates that reduce friction for legitimate customers.
Scenario analysis and stress testing is a directly relevant application for Fenrir’s readership. AI can evaluate a portfolio against thousands of macroeconomic scenarios simultaneously — far beyond the three-to-five scenarios that traditional risk management frameworks use. This is not a theoretical capability: institutional asset managers and risk teams at major banks have been deploying AI-assisted scenario analysis tools since 2023, with material improvements in tail-risk identification. The ability to construct non-linear, correlated stress scenarios (rather than simple parallel shifts) is particularly valuable in environments like 2025–26, where macro variables are moving in novel combinations.
Infrastructure & Predictive Maintenance
Predictive maintenance — using sensor data and ML to anticipate equipment failure before it occurs — is one of the cleanest AI ROI stories in the industrial economy. McKinsey estimates that AI-enabled predictive maintenance can reduce machine downtime by 30–50%, extend equipment life by 20–40%, and reduce maintenance costs by 10–25%. The power generation, oil and gas, and aviation sectors have been early adopters; the railroad and utilities sectors are scaling deployments now. For utility analysts in Fenrir’s coverage universe, this is a directly material investment thesis: asset owners using AI-predictive maintenance have structurally lower capex replacement cycles.
AI as Creative Collaborator, Not Just Replacement
The narrative that AI inevitably destroys creative work misses an important part of the picture. In architecture, AI generative design tools can produce thousands of structural variants optimised simultaneously for cost, energy performance, and aesthetic criteria — expanding the design space that architects can explore in a given project timeline. In film production, AI pre-visualisation tools allow directors to iterate on visual storytelling at a fraction of the cost of physical shooting tests. In music production, AI tools handle technical tasks (mixing, mastering, sample clearance matching) that previously consumed studio time at the expense of creative work.
The distinction that matters analytically is between AI as a creative collaborator — expanding what individual practitioners can produce and explore — and AI as a replacement that commoditises creative output entirely. The former is already happening and is broadly positive. The latter is a real risk at specific nodes of the creative economy (stock photography, entry-level copywriting, generic illustration) but is not the complete picture. The artists and practitioners who integrate AI effectively into their workflow are, at present, more productive and more competitive than those who do not — which is the same dynamic that characterised earlier waves of digital tools in creative industries.
Bottom Line
The positive case for AI is not theoretical, speculative, or dependent on AGI. It is grounded in measurable productivity improvements in live deployments, one of the most consequential scientific breakthroughs in fifty years (AlphaFold), material improvements in diagnostic accuracy in healthcare, and a genuine democratisation of access to expertise that has been rationed by cost and geography for generations.
The honest qualifier is timing. The economy-wide productivity dividend requires diffusion and workflow redesign at a scale that is still years away from completion. The access dividend requires regulatory frameworks that are still being written. The scientific acceleration requires translation from research breakthrough to clinical or commercial deployment — a process that takes time even when the underlying science moves fast. The upside is real; the timeline is longer than AI optimists typically advertise.
Part III examines the costs that accompany these benefits — the disruptions to labour markets, the pressure on energy and resources, the concentration of capital, and the misuse risks that regulators are struggling to get ahead of.
Sources & Citations
- Goldman Sachs Research. (Mar 2026). Q4 Earnings Analysis — AI and Productivity (via Fortune).
- Goldman Sachs Research. (2023). Generative AI Could Raise Global GDP by 7%.
- McKinsey Global Institute. (2023). The Economic Potential of Generative AI: The Next Productivity Frontier.
- Jumper, J. et al. / DeepMind. (2021). Highly Accurate Protein Structure Prediction with AlphaFold. Nature.
- OECD. (2025). Macroeconomic Productivity Gains from AI in G7 Economies.
- MIT / Stanford. GitHub Copilot productivity study (2023). Peng, S. et al. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.
- DeepMind. (2023). GraphCast: AI model for faster and more accurate global weather forecasting.
- Legal Services Corporation. (2022). The Justice Gap: The Unmet Civil Legal Needs of Low-Income Americans.
- McKinsey Global Institute. (2023). Predictive Maintenance ROI estimates. Via The State of AI in 2023.
- UNESCO. (2023). Global Education Monitoring Report — Teacher Shortage Statistics.
Leave a Reply