Best of: The future of the innovation economy

Summary

The discussion frames AI as a general-purpose technology that could reshape products, services, jobs, and entire industries. The panel’s central business lesson is that access to AI alone is not enough to generate value. Organizations must also address adoption costs, redesign workflows, train workers, and decide where human judgment remains essential.

Business and organizational takeaways

• Expect productivity gains to take time. Past technologies, such as electricity and personal computers, often had profound effects locally before their impact appeared in aggregate GDP growth. Firms may need to restructure production and industries to capture the benefits.
• Treat adoption as an operational challenge, not just a technology purchase. Barriers include organizing data, hiring skilled people, integrating new tools, choosing solutions that may be difficult to change later, and gaining employee support.
• Use AI to augment work where possible. Fei-Fei Li argues that AI is more likely to change tasks within many jobs than to replace whole occupations. For example, AI could help with some nursing-home tasks without eliminating the nurse’s role.
• Design for human complementarity. Neale Mahoney suggests shifting the innovation question from “How do we maximize innovation?” to “How do we shape innovation to complement human skills and meaningful work?”
• Plan for last-mile implementation. Universities and other research institutions can help fund early-stage, high-fixed-cost innovation. Entrepreneurs can then adapt products to real-world settings and solve adoption problems.
• Look for opportunities in underserved sectors. The panel names childcare, nursing, healthcare, and elder care as areas where demand for human work remains substantial and AI could help workers move into more productive roles. Governments can support demand through procurement and investment.
• Make AI accessible to smaller firms. Natural-language interfaces could let small businesses operate software through chat rather than relying on coding, spreadsheets, or expensive enterprise systems, potentially helping them become more efficient and scale.

Frameworks and playbooks

Bottleneck-based technology adoption

1. Identify the constraint AI removes, such as the cost or availability of software development.
2. Reassess what now prevents adoption, including data readiness, workflow redesign, training, employee acceptance, or implementation capacity.
3. Invest in those complementary changes rather than assuming cheaper software will lead to immediate deployment.

Human-centered innovation

• Evaluate whether a product complements people’s capabilities and supports meaningful work.
• Involve users and workers in shaping how tools fit into workflows.
• Measure success by more than automation or cost reduction.

AI policy principles proposed by Fei-Fei Li

• Science, not science fiction: Base regulation on evidence and measurement.
• Pragmatic, not ideological: Work with practical, existing regulatory approaches where appropriate.
• Invest in innovation capacity: Support universities and the public sector as part of the innovation system.

Transition planning under uncertainty

• Prepare for both gradual, geographically dispersed change and rapid, concentrated disruption.
• Invest in education and social safety nets before it is clear exactly which workers will be affected.
• Consider protections for workers whose jobs or human capital lose labor-market value.

Competition, markets, and risk

• Scale economies can produce concentrated markets, but even limited competition may benefit customers and innovation.
• Dominant firms can act as “toll booths,” extracting fees across many transactions and potentially slowing innovation around them.
• Open models may lower AI prices and help businesses build on the technology. The panel emphasizes monitoring both price and quality, especially for smaller businesses and countries that depend on outside AI providers.
• If an economy pays high prices for AI while automating labor, workers’ wages could fall without corresponding reductions in the prices of goods. The panel presents this as a risk to watch, not a forecast.

Examples and evidence cited

• Small-business digitization: Many smaller firms remain less digitized because implementation involves fixed costs and difficult choices. A restaurant operating with paper, WhatsApp, or WeChat may face a significant leap to more integrated software.
• Bank adoption: Banks may struggle to use machine learning because of hiring difficulties and the work required to organize data and implement systems.
• Creative work: Filmmakers, storytellers, visual creators, and game developers describe AI tools as ways to increase creativity and productivity.
• New occupations: Neale Mahoney cites economist David Autor’s observation that roughly 70% of occupations in the economy emerged over the prior 70 years, illustrating how work changes over time.
• Disruption risk: The panel contrasts gradual adaptation with concentrated shocks, citing factory-town impacts associated with automation and China’s entry into the WTO as examples of how localized disruption can be severe.
• Labor-market uncertainty: Mahoney raises a hypothetical scenario in which 5–10% of people lose their primary occupation due to AI, underscoring the need to consider safety nets.
• Public sentiment: Mahoney cites a reported decline in belief in the American Dream from 70% to 25% over a generation; he notes that the causes are uncertain.

Metrics and timelines

• No company revenue, margin, CAC, LTV, churn, or adoption targets are provided.
• The panel notes that economy-wide effects of general-purpose technologies can take a long time to appear; no specific forecast timeline is given.
• The discussion identifies education investment, workforce transition support, and AI pricing and quality as areas to monitor, but does not set quantified targets.

Presenters and sources: Russ Altman (host); Fei-Fei Li, Susan Athey, and Neale Mahoney (panelists). Other sources mentioned include ChatGPT, Claude, Perplexity, David Autor, and John Rawls.