If a blog post can define an age, Marc Andreessen’s “Why Software Is Eating the World” was the prophecy of the tech company era. The fifteen years between the global financial crisis and ChatGPT saw trillion-dollar firms upend industries, fuse people to devices, and turn attention into an asset class. Social and business life changed in a million small ways that added up to big changes, while public and social sectors largely proceeded as before. This will change in the AI era.
Andreessen’s argument was about business. Facebook and Twitter were paragons of internet-native companies “poised to take over large swathes of the economy.” Disintermediation, crowd-sourced content, and network effects hollowed out newspapers, travel agencies, retailers, and taxi services. These changes had little traction in health ministries, schools or extension services, where most production is nonmarket and the main inputs are human time and skill. AI is different because it diffuses into organizations to supplement human tasks. Don’t expect AI firms to run primary healthcare centres or farmer field schools: AI will eat policy in portions and from within.
That distinction matters because it changes who captures the upside. AI labs such as Anthropic, OpenAI, Mistral and DeepSeek are the latest hypergrowth, capital-burning firms sprinting from idea to multi-billion-dollar valuation. Market jitters about AI eating software-as-a-service (SaaS) remind us that “software eating the world” happens in cycles. This one may create greater value as adoption diffuses more widely across the economy, including nonmarket sectors. Some of these new giants are valued at hundreds of billions because they can reach spheres of activity that eluded the internet giants.
Palantir is the one major tech firm built specifically for the public sector. Instead of displacement, it aims for infiltration, bringing data-integration and workflow capabilities to existing systems. Its “data ontology” makes it easier to build machine learning and AI applications on messy public databases. Sales are exploding not just because of convenience, but because AI’s General Purpose Technology aspect—that other reading of GPT—makes adoption valuable in activities less touched by the prior technology cycle, strengthening the case for investment in technical capacity.
The public and voluntary sectors did evolve during the internet era, but mostly at the edges: citizen portals, online enrolment, digital case files. The production processes of services themselves—lessons taught, patients consulted, farmers advised—were left largely untouched, because the platform economy had little leverage on them. AI is different in that it can contribute directly in the lesson, the consultation, the advice. The outer boundaries of the state and the industrial organization of public services may largely stay where they are, but AI is likely to reshape their core.
The most relevant uses appear, perhaps surprisingly, in low- and middle-income countries (LMICs), where the marginal value of an extra hour of teacher attention or clinician judgement is highest, and appetite for new delivery models is greatest. Consider three large public-service sectors:
- Education: Three recent studies in LMICs show substantial gains from AI-assisted teaching. In Nigeria’s Edo State, 13 hours of AI tutoring raised English scores by 0.23 standard deviations (SD); in Ghana, 30 hours of AI-supported maths tutoring produced a 0.37 SD gain; in Sierra Leone, 12 hours of in-class maths tutoring raised scores by 0.28 SD. Large effects from modest dosages in resource-constrained classrooms suggest scope for AI assistance in teaching.
- Health: Two recent studies show AI helping frontline clinicians in African settings. In Kenya, clinicians using a large-language model (LLM)-based safety net for several months gave 14 percent fewer recommendations posing critical care risks. In clinical simulation in Rwanda, leading LLMs significantly outperformed local clinicians on answering clinical questions, with one model scoring 0.8 points higher than local general practitioners on every five-point metric. The signal points to substantive support where specialist capacity is thin.
- Agricultural extension: LLM-based agriculture impact evaluations remain incomplete, but the underlying channel has a strong randomized controlled trial (RCT) base. PxD‘s SMS- and interactive voice response-based extension trials found yield gains of around 8 percent for Kenyan sugarcane farmers and Indian cotton growers, and around 26 percent for cumin farmers in Gujarat. Digital Green’s video-based extension model lifted yields by 12–18 percent and profits by 9–24 percent under some specifications. Its LLM successor, FarmerChat, has handled over 16 million queries from 1.3 million users across six countries; three on-going RCTs will show whether personalized interactions costing below $1 enhance results.
These are selective examples, and the caveats are real. Some education studies measure after-school tutoring rather than reformed classroom practice, and more instructional time tends to improve outcomes with or without AI. In health, Jason Abaluck and colleagues show limited benefit in a early real-world deployment. And agriculture is not the only field where pilots have failed to translate into national programs. Can AI solve previously intractable problems? Surely not: no virtual agent can fix crumbling health facilities or restore degraded topsoil. Predictions of the internet revolutionizing education and other sectors proved tenuous. Similarly, AI availability solves neither political economy nor resource constraints, and the real-world web of market, institutional, and governance failures will prove as sticky as ever.
But AI need not be a miracle solution. Realizing part of its tentative potential is prize enough, use case by use case. Barriers to turning policy potential into reality include:
- Access: AI adoption is already diverging along productivity lines both within and across countries. Public LMIC access risks a similar gap unless financing, infrastructure and implementation capacity can grow quickly.
- Effectiveness: Learning to build strong solutions in-house (through evaluations), creating competitive markets for AI solutions, and choosing the right suppliers will be necessary for effective implementations on limited budgets.
- Transformation: The expansion of tech into public and social services cannot be one-way. Transformation cuts both ways, calling for reimagining the arm’s length relationship between tech and nonmarket sectors.
Progress on these topics will accelerate adoption, but either way: software in the guise of AI is bound to eat the policy world. It will happen in bite-sized pieces, and from the inside as AI diffuses across activities that the previous technology cycle left untouched. Tech firms will not take over, because these sectors are public and voluntary for reasons rooted in their economics and politics, not their technology stack. But AI will take up an increasing share of the work, augmenting and transforming tasks that used to rely on human effort alone. Channelling that integration towards the public interest is the policy mission of our age. As AI eats policy—think small-plate tasting menu over fast-food delivery—policymakers need to make sure that it doesn’t make a meal of it.
With thanks to Han Sheng Chia and Charles Kenny for insightful discussion.
