A software engineer using AI coding assistants can produce more in an hour. But what happens next?
The company might deliver the same amount of software using fewer engineering hours—the same with less. It might use its existing engineering capacity to build more products—more with the same. Or it might redirect that capacity from feature implementation towards product management, architecture and testing.
All three outcomes begin with the same productivity improvement. What differs is how the company responds to it.
Making an input more productive does not tell us how much of that input the company will ultimately use. That depends on what the technology can do, how production is reorganised and how much additional activity becomes worthwhile as costs fall.
What the technology changes: technological bias
Technological bias describes how a technology changes the relative productivity of different inputs. Economists assess this by comparing the inputs’ marginal products: the additional output produced by adding a little more of one input while holding the quantities of the others constant.
If a technological change raises the marginal product of labour relative to capital, it is labour-biased. If it raises the marginal product of capital relative to labour, it is capital-biased (capital includes productive assets such as compute and GPU capacity, software systems and AI infrastructure, not simply money available for investment). Here, “biased” simply means tilted towards one input.
It is important to note that the technological bias is a property of the production technology and describes a change in productive possibilities. It does not, by itself, determine the quantities of labour and capital that a company will choose to employ.
What the company chooses: factor use
Factor use describes how much labour, capital or another input a company employs after introducing the technology.
Unlike technological bias, factor use is an outcome of the company’s response. It depends not only on the technology, but also on input prices, the ability to substitute between inputs, demand for the company’s output and the scale at which the company chooses to operate.
It helps to separate two margins.
Factor use per unit of output
If a technology reduces the labour required to produce each unit, it is labour-saving for a fixed production target. The company can produce the same amount using fewer hours of labour.
Total factor use
A company that can produce something more cheaply may decide to produce more of it. Total labour use therefore depends on both the labour required for each unit and the number of units produced:
Total labour hours = labour hours per unit × units produced
A technology can therefore be labour-saving per unit while still increasing total labour use, or labour-using in total, if the resulting expansion in output is large enough.
The substitution effect and the scale effect
Economists separate the forces behind these outcomes into the substitution effect and the scale effect.
The substitution effect captures how a company changes the combination of inputs used to produce a given amount of output. A new technology may allow capital to perform work that previously required human time, reducing the labour needed for each unit.
The scale effect captures the additional demand for inputs created when lower production costs make a larger volume of output worthwhile. The company may produce more, serve more customers or undertake activities that were previously uneconomic.
If the expansion in output is not large enough to offset the saving per unit, total use of the input falls. If the expansion is large enough to more than offset the per-unit saving, total use rises and the technology is factor-using in total.
The terms factor-saving and factor-using can therefore be ambiguous unless the relevant margin is stated. “Labour-saving per unit of output” and “labour-using in total” can both describe the same technology.
How technological bias and factor use fit together
Technological bias and factor use answer different questions:
Productivity effect: Which inputs become more productive?
Technological bias: Which input’s marginal product increases relative to the others?
Factor use per unit: How do input requirements or the input mix change for a given amount of output?
Total factor use: How much of each input does the company employ after output and production have adjusted?
These are related questions, but they are not competing descriptions of the same effect. Technological bias describes how the production technology changes. Factor use describes how the company reorganises production in response.
Applying the framework to generative AI
Consider a generative AI coding assistant that helps software engineers write, test and review code more efficiently.
If the assistant increases what an engineer can produce in an hour, it increases the productive capacity of the labour working with it. For a fixed amount of output—for example, a given number of software features—the assistant reduces the engineering time required per unit. In this sense, the technology is labour-saving for that production target.
This does not mean that the company will necessarily use less engineering labour overall. Lower development costs may lead it to build more software. If that expansion outweighs the engineering effort saved on each unit, the company will use more engineering labour overall: the technology will be labour-saving per feature but labour-using in total.
The company might use its released capacity to shorten development cycles, improve quality, address technical debt or start projects that were previously too expensive. It might build more features, customise products for more customers or enter new markets.
A plausible pattern for generative AI
Suppose the coding assistant reduces the engineering time required per feature by 30 per cent and raises the marginal product of engineering labour relative to the supporting capital. Suppose also that the company responds by doubling the number of features it develops and that adoption requires more compute and AI infrastructure to support model inference.
Under those assumptions, a plausible description is:
Labour-biased, because the assistant raises the marginal product of engineering labour relative to the supporting capital. That relative-productivity effect is a stated assumption in the example, not a conclusion that follows from the productivity improvement alone.
Capital-using in total, because the company uses more compute and AI infrastructure to support model inference.
Engineering-labour-saving per feature, because each feature requires fewer engineering hours.
Engineering-labour-using in total, because the expansion in software production more than offsets the reduction in labour required for each feature.
These descriptions are compatible because they answer different questions.
Note that this is a potential verdict for a particular use of generative AI, not a universal verdict on the technology. If the company did not expand production sufficiently, the same assistant could be engineering-labour-saving both per feature and in total. If it raised the marginal product of AI infrastructure more than that of engineers, it could be capital-biased rather than labour-biased.
For generative AI, then, the framework produces a clear but conditional assessment: identify the technology’s relative productivity effect, state which factor and margin are being discussed, and then examine whether the scale effect outweighs the saving per unit.
Applying the framework to agentic AI
A useful distinction is that conventional generative AI use often assists people with particular tasks, while agentic systems can carry out longer sequences of work: choosing features from the backlog, designing the feature, using available tools, building, testing and referring uncertain decisions to an engineer.
This changes the analysis because “labour” is no longer a sufficiently precise category. It becomes useful to distinguish between two broad types of work.
Execution work
Execution work includes writing code and tests, optimising, generating documentation, and updating the product/sprint backlog.
As agents become capable of completing longer sequences of these activities, less human time may be required for execution at a given level of output.
Complementary work
Complementary work includes defining objectives, exercising judgement, reviewing uncertain results, handling exceptions and taking responsibility for consequential decisions.
Agentic systems can increase the reach of people doing this work because each person can direct and evaluate a larger volume of machine-assisted activity.
These are categories of activity, not fixed categories of workers. A software engineer, lawyer, analyst or manager may perform both kinds of work. As AI capabilities develop, the balance within each role may change.
A software engineer, for example, may spend less time writing code and tests and more time on architecture, system constraints and evaluation. A domain expert may examine more cases while concentrating attention on ambiguity and exceptions. A manager may coordinate a larger volume of activity without directly supervising every step.
The economic effect is therefore not only a change in how efficiently existing tasks are performed. It is also a change in how tasks are grouped into roles, how workflows are coordinated and where human attention is most valuable.
A plausible pattern for agentic AI
Once labour is divided into different forms of work, several effects can occur simultaneously.
To make the framework concrete, suppose an agentic system can complete routine production steps that were previously performed and coordinated by people. The company invests in additional compute and agent infrastructure, increasing its use of model inference. Suppose the system raises the marginal product of that capital relative to human execution labour, increases the marginal product of complementary labour relative to execution labour, and produces enough additional activity to increase the need for human direction, evaluation and exception handling.
Under those assumptions, a plausible description is:
Capital-biased relative to execution labour, because the system raises the marginal product of AI capital relative to the marginal product of human execution work. Again, this is a stated assumption in the example, not a conclusion that follows simply from the use of more capital.
Capital-using in total, because the company uses more compute and agent infrastructure to support model inference.
Execution-labour-saving per workflow, because less human time is required to perform and coordinate routine production steps for a given amount of output.
Biased towards complementary labour relative to execution labour, because agents increase the productive reach of people providing judgement, direction and evaluation relative to people performing routine execution work.
Complementary-labour-using in total, because the assumed growth of AI-enabled production increases demand for direction, evaluation and oversight.
Agentic AI might reduce the execution time required to produce a piece of software while increasing the amount of software a company chooses to build. It might allow one experienced engineer to oversee more implementation while also making experienced engineering judgement valuable across a wider set of projects.
Whether the company ultimately uses more or less of a particular kind of labour depends on how these effects combine.
The same caveat applies to agentic AI: this is a potential verdict for a particular use, not a universal property of the technology. If agents did not raise the marginal product of capital relative to execution labour, the technological change would not be capital-biased on that comparison. If AI-enabled activity did not expand sufficiently—or if direction and evaluation could also be automated—total demand for complementary labour might not increase.
The important shift may be in the composition of work
The most immediate question about the impact of AI is often whether it will increase or reduce demand for labour. But that may be too broad to capture the more interesting changes.
The effects are likely to differ not only between occupations, but also among tasks within the same occupation. As some activities become easier to delegate, human effort can move towards setting direction, integrating knowledge, resolving ambiguity and evaluating results.
Generative AI can raise the productivity of individual tasks. Agentic AI can go further by reorganising how those tasks are combined into workflows. Its economic significance may therefore lie as much in changing the composition and coordination of work as in reducing the cost of any individual activity.
AI assistance: This article was developed with assistance from OpenAI GPT‑5.6 Sol, which helped structure and draft. The ideas and conclusions are my own, and I reviewed and edited the final text.