AI employment effects: a primer

All your questions on what AI means for your job answered

Will AI automate your job? Will your income increase? What happens to the labour share? Is AI complementary to labour inputs or a substitute? What are the theories that can answer these questions? This post will hopefully answer at least some of these questions by focusing on two prominent papers. I may update this as new theoretical insights are introduced. Hence one should see this as an indefinitely ongoing project.

Jobs as task bundles

I will start by introducing the Coasian notion of jobs as a continuous bundle of distinct and discrete tasks: inspired by Garicano et al (2026). The idea is that exposure measures can be misleading as jobs aggregate multiple tasks rather than being uniform in composition. Some of these tasks are complementary, and some substitutes. Different inputs can complement some tasks and substitute others. Bottlenecks in enacting one task can delay the entire bundle operation¹. Therefore exposure measures assume a linearity that simply does not exist.

Crucially, the effects of AI in this model depend on how tight the bundle is. If the bundle is weaker, so it’s cheaper to break the bundling, AI automation of some tasks are more likely to narrow the job boundary and pose greater disemployment effects, and vice versa. In fact, if the bundling is sufficiently strong and bottlenecks tight, then AI automation improving productivity in some tasks may enhance the overall occupational productivity - leading to positive employment and wage effects.

Consider a job as a production function of two tasks:

y = (q1^α)(q2^(1-α)).

Task one can be automated, whilst two cannot. There exists a time constraint, t1+t2<=1 for nonnegative times devoted to tasks one and two respectively.

Two scenarios are now introduced. If bottlenecks are sufficiently tight such that task one can only be partially automated, then output is given by:

1. y = ([{(μ(m))^(1-η)}{t1s1}^η]^α)[t2s2]^(1-α)

where μ(m) is an increasing function of frontier AI capability m. Si denotes the share of the job represented by task i. The shares are equivalent to the comparative advantage of the worker with respect to different tasks; in other words his abilities (relevant later). η gives the “human share” of output within a job. Notice that we have simply represented qi as a function of ti, si, and μ(m). Output remains Cobb-Douglas, with AI featuring as labour-augmenting technology.

If bundling is weak, then:

2. y = (1-c)[{μ(m)}^α]s2^(1-α).

Our employee focuses exclusively on task two. Task one is fully automated at cost c, which incorporates the bottlenecks and Coasian frictions discussed. Here, AI is not labour-augmenting but a distinct input in itself.

As Garicano has noted, when m is sufficiently high and c sufficiently low, more jobs converge to the weak bundling case. Eventually, this is how AI could replace most human labour. For simplicity, let’s call this AGI. Here we treat both m and c as exogenous, yet in reality, as m rises, c falls². Nonetheless, social and political-economy factors alone (as well as Hayek's famous coordination problem) will ensure that c remains strictly positive, as well as Moravec's paradox (although robotics capabilities are advancing too, albeit at a slower rate). Therefore even with AGI, human employment remains positive. Indeed, this is a recurring theme across all the models covered here. In other words, to maintain a future career, look to sectors and invest in skills where humans still maintain a comparative advantage³.

If jobs are task bundles, what does this mean for wages?

In scenario 1, our worker aims to maximise output⁴ subject to his binding time constraint⁵:

t*1 = αη/(αη+1-α), t*2 = (1-α)/(αη+1-α).

Both the output and human shares determine task allocation within a job. Raising either raises the time devoted to the corresponding tasks. This is attenuated by the cross-task implications: both tasks are essential for the job, yet allocating more time to one task reduces the time available on the other. By substituting into our production function, you get equilibrium output.

Wages are a function of revenue, Py. Let P=p1+p2, with pi simply being the share of value that task i contributes to production. p1q1 = αPy and p2q2 = (1-α)Py. In scenario 1, the worker gets the human share of revenue for task one.

We now arrive at one of the key results. The wage effects differ according to whether AI is complementary (the labour-augmenting scenario 1) or a substitute (scenario 2). ω = (αη+1-α)Py or ω = (1-α)Py for scenarios 1 and 2 respectively. The relevant shares determine the allocation of the wage across tasks and inputs, as is standard in perfect competition. Crucially, in scenario 1, as long as η is positive, then αη+1-α>1-α, so wages are higher when AI is labour-augmenting relative to when AI is a substitute.

There exists heterogeneity however in how workers will allocate across occupations depending on their abilities. Let s*1 divide workers into low task one ability and high task one ability. Workers of low and high task one abilities hold s1<s*1 and s1>s*1 respectively. s*1 is:

s*1 = ([(1-α)(1-c)/{(αη+1-α)(t*1^αη)(t*2^(1-α))}][{μ2(m)^α}/{μ1(m)^(α(1-η))}])^(1/αη)

for μi being frontier AI's contribution to output in scenario i. AI substitutes weakly bundled tasks faster than it augments strongly bundled tasks given the coordination costs in AI adoption within a task. s*1 is increasing in frontier capability and the output share of task two. s1 is decreasing in coordination costs. Increases in these are weighted by the human and output shares of task one however. Time constraints feature in the attenuation. If and only if s1<=s*1, their wages are weakly higher in weakly bundled than strongly bundled occupations as they hold a comparative advantage in the former. Therefore as capabilities advance and adoption accelerates, more workers will be exposed to automation. By solving for the equilibrium, whether overall wages increase for workers of low task one ability depends positively on the output share of task two. In general, there is high ambiguity and uncertainty on the long-run earnings implications here, at least for those not holding capital.

So how does AI affect employment in task bundling?

Let U’/P denote the relative price of participation in labour markets, with U’ the utility from the outside option. Households maximise a value function incorporating U’ and wages in the strongly and weakly bundled cases, which depends on s, m, P, c, and U’ (all exogenous). Equilibrium sorts workers into strongly and weakly bundled tasks, and inactivity, depending on their abilities and the value of their outside option.

We can also model P as an inverse demand function, P(Y) = Y^(-1/β) for β>0, so demand in the goods market is downward sloping and convex. Substitution effects are also now endogenous by substituting in P(Y) for P in U’/P. Solve for the equilibrium, and employment falls only if:

(1/β)(dlnY/dm) > αdln(μ2(m))/dm.

Employment falls when the output gains from frontier capability advancements exceed the productivity gains from automation, weighted by their elasticities. If goods demand is more inelastic, we will observe greater disemployment effects (and AI reducing employment is more likely), and if the output share of task one is greater, we will see lower disemployment effects (and employment falls less likely). The intuition is that more inelastic demand means that sales are less responsive to prices, so all other things equal profits adjust more rapidly. The AI productivity shock reduces prices and profits to the extent that firms cut back on hiring. If the output share of task one is greater, then less workers are in weakly bundled occupations.

Our overall conclusion is that the extent to which AI is labour-augmenting or a direct substitute to labour inputs will determine the employment effects. In all competitive equilibria (unique but changes according to the parameter changes), the number of both weakly and strongly bundled occupations is nonzero however.

Nonetheless, (rapidly) increasing frontier capability is sufficient for more roles to be at risk of automation. This does NOT require a singularity nor superintelligence. As m rises, s*1 rises too. If this model broadly holds, then the risk of large disemployment effects within the near future is high, especially given current forecasts on the advent of AGI. Organised rent-seeking and regulatory capture can delay this via raising c, yet output growth will also be lower. I would rather redistribute via fiscal policy (automatic stabilisers) and taxes/transfers than via predistribution, perhaps via UBI or NIT, yet some policy response will likely be required to cushion the shock.⁶

O-ring technology

An alternative yet related approach by Joshua Gans (2026) generalises this task-based framework, yet departs from Garicano et al in abandoning Cobb-Douglas assumptions. There are no diminishing returns to tasks as inputs - following the logic of learning-by-doing endogenous growth models. Spillovers between tasks generate complementarities, which roughly offset diminishing marginal returns. Hence the implications for wages and employment are more optimistic.

Hence production is a simple multiplicative function of all the tasks required. Jobs are not represented as bundles, but rather final output itself is decomposed into n tasks. A representative household supplies L hours given a fixed time constraint, so for tasks performed via human labour, q=aL where a is labour productivity. Each task can also be automated by technology θ at cost r. If a task is automated, q=θ.

As output is simply multiplicative and there exists constant returns to scale, changes to labour supply scale the entire production function. Therefore wages are determined by Nash bargaining rather than marginal product (in line with DMP models). Moreover, if not all tasks can be automated, then labour is still required as an input. Under these conditions, total wage costs W(a) are set to maximise the following surplus, with a being the number of tasks able to be automated:

[{W(a)-u0}^β][{Y(a)-r-W(a)}^(1-β)]

where u0 is the outside option, the payoff from disagreement. This gives the following result:

W*(a) = u0+β(Y(a)-r-u0).

In other words, a solution very similar to a DMP model⁷.

Profits are given by Π(a) = (1-β)(Y(a)-r-u0) by substituting W*(a) into our definition. An agreement is not reached only if all tasks can be automated. However this more optimistic result ignores heterogeneity and perhaps make unrealistic assumptions regarding scale effects and complementarities. Therefore one should see this model as a description of jobs and production where technology and labour complement each other, which produces similar results to the strongly bundled case.

1. There also exists organisational constraints within and across firms and industries that delay adoption and diffusion; another Coasian point. Conversely, some industries face more rapid and instantaneous spillovers and agglomeration effects than others. For now, let’s abstract from this and focus on the within-job dynamics.

2. When c falls, the fixed costs to starting a new firm will also fall, albeit this is outside the scope of the model. Firms will shrink as Coasian coordination costs fall. More households will operate their own businesses, so the capital share of output surges.

3. For me, this involves blogging, and providing a unique commentary distinct from what you might get from an LLM. Unique is the key word here, which is why I think the concerns in academia are unwarranted: AI is augmenting innovation thus far, albeit some fields such as mathematics may be more exposed to substitution. I’m also currently developing a career in manufacturing sectors where human labour will still always be an essential task input. The best jobs in these industries involve a range of data analysis (to model KPI and inventory dynamics etc) and social coordination (i.e. ‘bundled’ management roles), or oversee automated inputs (a prime example being CNC and other roles in precision manufacturing). In both these cases, AI is complementary and labour-augmenting. In my view, the traditional prestigious professions (coding, law, IB, consulting, etc) will suffer the greatest relative decline.

4. We're abstracting from information frictions so there’s no moral hazard. Wages depend on output, as is standard in perfect competition. Hence utility maximisation is output maximisation. If you insist on incorporating these dynamics, then my priors are that raising c is sufficient here, so the overall dynamics are the same. Departing from perfect competition largely introduces a markdown, yet otherwise the optimisation problem is the same. In a minute we will model this explictly. A more interesting question is how AI affects our tradeoffs between consumption and leisure, hence the income and substitution effects for labour supply. The paper introduces outside options later. I’m blogging as I'm reading, so we'll find out later. Intertemporal decisions between consumption vs saving also feature though, especially as AI will affect r. I also wonder how workers will reallocate into bundled roles, which will require modelling the tradeoff between consumption, leisure, and human capital investment.

5. If you're an undergraduate new to economics, or seek to refresh your memory, the proof in the paper is easy to follow. Only basic multivariate calculus and a cursory familiarity with Lagrangians is necessary. On a general note, my post covers the intuition in my own reasoning, so readers should go to the paper for the proofs.

6. In some cases the HANK literature implies that redistribution can actually be efficient, so a social planner doesn't necessarily choose between efficiency vs equity according to their social welfare function. A subsequent post will cover this in more detail.

7. Note that my conclusions in this back of the envelope exercise also mirror the Garicano et al result, in a case where AI is labour-augmenting for many roles (this was written before Claude Code, Moltbook, and Mythos).