If AI disrupts work, can welfare systems respond?

As an associate fellowship at the Windfall Trust, I worked with Avinash Kothuri and Deric Cheng to develop the Welfare Resilience Assessment Framework. The framework asks a question that receives less attention than forecasts of AI’s impact on jobs: if work and incomes change, are welfare systems equipped to respond?

We also applied to framework to Kenya and Brazil. These two countries face different pressures but the comparison reveals a common lesson that whilst a policy can exist on paper, the systems needed to identify affected people, reach them and support their next move are equally important.

You can read the framework and country studies here.

Key takeaways

• Exposure is not preparedness. Knowing which jobs may be affected by AI does not tell us whether workers can access income support or move into new work.

• Different risks test different capacities. A sudden loss of earnings, a prolonged change in occupations and a shift in the distribution of economic gains require different responses.

• Delivery is often the binding constraint. Eligibility rules, earnings data, social registries, payment systems, training capacity and appeal routes determine whether a policy reaches the people it is meant to help.

• Sequencing matters. Governments should distinguish measures they can implement through existing systems from those that depend on legal, fiscal or administrative reform.

From job exposure to welfare preparedness

A great deal of AI research asks which tasks and occupations could be automated or transformed. That is useful, but it leaves out the institutions that determine what happens to people after a shock. Two countries with similar exposure could have very different outcomes if one can identify affected workers, stabilise their income and connect them to new opportunities while the other cannot.

The timing and scale of AI-driven disruption remain uncertain. It may arrive as abrupt job loss in some sectors, a slower erosion of earnings or hours in others, or a gradual reorganisation of tasks. It could also change where economic activity takes place and how welfare systems are financed. A single measure of “AI exposure” cannot capture these differences.

Our framework starts with three linked questions:

1. What is the risk?

2. Which policy could address it?

3. What conditions would allow that policy to work?

Consider an income shock among workers whose earnings are irregular or poorly recorded. A government might propose an earnings responsive benefit. But that response depends on being able to observe income changes, identify eligible people, authorise data use and make timely payments. Without those foundations, a promising policy design will struggle to function.

Four capacities

We organise the assessment around four capacities:

1. Coping capacity: can the system stabilise incomes and access to essential services during a sudden shock?

2. Adaptive capacity: can it help people move between jobs, occupations or sectors as tasks change?

3. Transformative capacity: can it respond to longer-term changes in the distribution of income, economic geography and the fiscal base?

4. Systemic capacity: can its data, rules and delivery institutions make those responses work in practice?

The fourth runs through the other three. An unemployment benefit may be well designed but exclude non-standard workers. A retraining programme may have places available but lack a way to identify the people most exposed. A digital payment system may transfer cash quickly while an out-of-date registry misses the household that needs it.

The aim is a diagnosis, not a league table. It should show which risks matter most in a particular country, what can be done now and which options require stronger foundations first.

What Kenya and Brazil reveal

In Kenya, more than 80% of workers are informally employed. Mobile money and existing programmes offer a route to rapid payments once recipients are known. The harder problem is identifying changes in need when earnings sit outside formal payrolls and registries are unevenly updated. AI-related changes in digitally traded services could compound income volatility that is already difficult for the state to observe.

In Brazil, large-scale registries and payment systems give the state substantial capacity to reach households. The gap lies more in the transition between losing work and finding new work. The case study finds no standing system that consistently connects identification of displaced workers, temporary income support and mid-career retraining. It also examines whether oversight and appeals have kept pace with automated welfare decisions.

These are different diagnoses, so they imply different priorities. Kenya’s immediate challenge includes better visibility of informal earnings and more reliable enrolment. Brazil can build on stronger delivery infrastructure, but needs a clearer pathway for workers whose jobs change or disappear. Neither case supports lifting a fashionable policy from another country and assuming it will work without the institutions beneath it.

Why the framework matters

AI policy often moves quickly from a forecast of exposed jobs to a list of proposed interventions. The missing step is testing whether those interventions can be delivered, to whom, and under what conditions.

The Welfare Resilience Assessment Framework makes that step explicit. It links risks to policy levers and then tests the data, legal authority, fiscal space and delivery capacity on which those levers rely. That allows policymakers to separate practical near-term measures from reforms that need years of institution-building.

We do not know exactly how AI will reshape labour markets. We can, however, examine whether the systems intended to protect people are ready for several plausible forms of disruption. That is a more useful starting point than treating exposure estimates as a policy plan.

The full framework and the Kenya and Brazil case studies are available from Windfall Trust.

The work was co-authored with Avinash Kothuri and Deric Cheng (framework), Emma Kimani and Deric Cheng (Kenya), and Taiye Chen and Deric Cheng (Brazil).