Abundant Agents, Scarce Organizations

01The question

The price of machine intelligence is collapsing. Its measured economic footprint is not. Meanwhile, firms are hiring more people to implement AI.

Agents perform tasks, as labor does, and are owned and bought, as capital is. They are also reproducible: an additional instance costs only the compute it consumes. The paper's answer is that the agent is not the scarce factor. Because agents can be bought in any quantity at price \(p_A\), they earn no scarcity rent. What becomes scarce, and captures rents as AI gets cheaper, is the capacity to implement them.

02Implementation as a fixed cost per task

The model is a task-based model of production with one addition. Before agents can perform a task inside a firm, the task has to be implemented: integrated with the firm's systems, the workflow redesigned, the data prepared, the output evaluated. Implementation uses a distinct input, implementation labor \(H\) at wage \(v\), and costs \(\eta\) units per task, once, whatever the scale. A firm implements the marginal task when

The left side is the cost bill of the marginal task times the log saving from automating it. Agents are non-rival in design and rival in deployment. Implementation is a fixed cost in organization, so economies of scale come from the organization, not from the code.

03Who automates

Firms automate more when agents are cheaper, when their market is larger and when implementation is cheaper. Market size and implementation costs enter only through their ratio \(D/(v\eta)\). An entrant out-automates an incumbent whenever \(D_e/\eta_e>D_i/\eta_i\). The advantage of AI-native firms is organizational, not technological: both buy the same model at the same price.

04Jevons for complements

Cheaper agents raise the demand for implementation labor. The same technology substitutes for the workers who perform automated tasks and complements the workers who deploy it. This is why firms hire more forward-deployed engineers as AI gets cheaper.

05The bottleneck and the Hulten bound

With implementation supply \(\eta I=\bar H v^{\zeta}\), the effect of cheaper agents on unit cost is

with equality when implementation capacity is fixed (\(\zeta=0\)). The productivity effect of cheaper AI is then exactly the current automation share, the Hulten term that macroeconomic estimates take as given. The extensive margin is closed, and its value goes to implementation labor as a rent. In the illustration, a 55% fall in the price of agents lowers unit cost by 19% with fixed capacity and by 36% with elastic capacity.

06The bottleneck shields labor, for a while

With capacity fixed, cheaper agents only lower costs on tasks already automated. The firm expands and hires more production workers. Displacement requires the automation set to grow, which requires implementation capacity. The sign of the employment effect depends on the elasticity of implementation supply and switches at a threshold. Small early employment effects are what the model predicts while organizations are the bottleneck. They say little about later effects.

07Results and predictions

Agents are abundant. What is scarce is the organization able to deploy them.

08Status

WP-002 shows that variable inference costs erode the economies of scale of software firms. This paper shows that they reappear in the organizations that deploy the software. WP-003 studies how AI's gains are split across workers; this paper studies how they are split across factors.