AI is getting cheaper. So why are firms hiring people to install it?

Cheaper AI raises the demand for the people who install it, and their pay. The price of artificial intelligence is collapsing. For a fixed level of capability, the cheapest model that reaches it has become between nine and several hundred times cheaper each year, depending on the task. You would expect the people who sell and install AI to be among the first casualties: if the product gets cheaper and better every month, surely it needs less help. The opposite is happening. Job postings for forward-deployed engineers, the people AI companies send into client firms to make their models work, rose more than eightfold on Indeed between April 2025 and April 2026.

This is not a hiring fad. It follows from a feature of AI that is easy to miss when looking at token prices. An AI agent can be copied at the price of the compute it uses. The work of plugging it into a company cannot.

A job invented for a problem

The forward-deployed engineer was popularised by Palantir, whose software was too complex to hand over and leave. The firm sent engineers to sit with the client, learn the data and the workflow, and build the deployment on site. AI companies have revived the role for the same reason. OpenAI, Anthropic and Cohere have all built or expanded forward-deployed and applied-AI teams, and the job now appears at software firms, consultancies and banks.

What these engineers do is not exotic. They connect a model to the firm's databases and tools, decide which steps of a process an agent should take and which a person should keep, prepare the data, write the evaluations that tell the firm whether the agent is doing the job, and redesign the workflow around it. None of this is about making the model smarter. All of it is needed before the model can do anything useful inside a particular firm.

The part of AI that does not scale

Here is the economic point. Once a workflow has been built, agents can be added at constant cost: a thousand more invoices processed means a thousand more calls to the model, at whatever the model costs. The work of building the workflow is different. It is done once per task, it costs roughly the same whether the firm processes a thousand invoices or a million, and it has to be done by people who understand both the technology and the business.

In the language of economics, AI agents have a constant marginal cost in compute, and implementation is a fixed cost in organisation. In the research note this essay draws on, I build that single assumption into a standard model of automation, in which firms decide task by task whether a person or a machine does the work. Everything else follows from it.

Jevons, one step removed

In 1865 William Stanley Jevons observed that more efficient steam engines had increased Britain's consumption of coal rather than reduced it: cheaper power made it worth using in many more places. Cheaper AI has a Jevons effect too, but it falls on the input that makes AI usable rather than on AI itself.

The mechanism has two parts. When the price of agents falls, each workflow a firm has already built becomes cheaper to run, the firm cuts its prices and sells more, and every task it automates is now worth more. And tasks that were not worth automating at yesterday's prices become worth it today. Both push in the same direction: the firm wants to implement more workflows, and each one needs implementers. The model makes this exact. As agents get cheaper, the demand for implementation labour rises, whatever the firm's size or industry.

So the same technology is a substitute and a complement at once. It substitutes for the people who used to perform the tasks it takes over. It complements the people who deploy it. Asking whether "AI replaces workers" mixes up two groups whose fortunes move in opposite directions.

Where the shortage shows up

Demand for implementers rises whenever AI gets cheaper. What happens next depends on whether they can be found. The people who can do this work are scarce: they need engineering skill, an understanding of how businesses operate, and experience with a technology that barely existed three years ago. In the short run, their number is close to fixed.

When it is, the figure above shows the result. Cheaper AI does not lead to more automation, because there is nobody to build the extra workflows. It leads to higher wages for the implementers there are. Postings rise because demand rises, but hiring cannot follow, so the price adjusts instead. The value of all the workflows that would be worth building at the lower AI price, and cannot be built, goes to the people who can build them.

That has an uncomfortable implication for anyone expecting cheap AI to transform productivity quickly. With the pool of implementers fixed, a fall in the price of AI lowers costs only on the tasks firms have already automated. Figure 3 compares two economies hit by the same 50% fall in the price of AI. They differ only in whether implementers can be found.

Where implementers are scarce, the price cut lowers unit costs by about a sixth, leaves the number of automated tasks unchanged, and more than triples implementers' wages. Where they can be found, the same price cut doubles the number of automated tasks, roughly triples spending on compute and nearly doubles the cost reduction. Notice the last row: with the pool fixed, production jobs rise, because the firm grows on the back of cheaper existing automation without automating anything new. Displacement comes when the implementers arrive. I take up that point, and what it means for the productivity statistics, in a companion essay.

Will the boom last?

History offers both answers. Enterprise software created an implementation and consulting industry that has outlived several generations of the software itself. Other technologies created roles that were later absorbed into better tools: building a website once required a specialist and now takes an afternoon with a template.

The model points to the variable that decides between the two. Everything above assumes that the work needed to implement a workflow stays the same as AI gets cheaper. But AI is increasingly good at precisely that work: writing integrations, mapping data, generating its own tests and evaluations. If the implementation requirement per workflow falls as models improve, there are two opposing forces. Cheaper agents raise the number of workflows worth building. Smarter agents lower the labour each workflow needs.

When AI does not help implement itself, implementer employment rises steadily as AI gets cheaper. When it helps a little, employment rises at first and then falls back: a boom followed by a bust, while automation keeps advancing throughout. When it helps a lot, the boom never happens. Which world we are in is an empirical question, and an observable one. The signal to watch is not the price of tokens. It is how many person-weeks it takes to put an agent into production, and whether that number is falling faster than the price of AI.

What follows

For firms, the binding constraint on AI adoption is unlikely to be the model, which anyone can rent, or its price, which keeps falling. It is the capacity to implement: people, internal tooling and processes that make the next workflow cheaper than the last. Firms that build this capacity now are buying the scarce input, not the abundant one.

For workers, the scarce skill is the combination: enough engineering to build the integration, and enough knowledge of the business to know what the agent should and should not do. It is scarce because few people have both. How long it pays depends on how quickly AI learns to do the integration itself.

For researchers, the argument makes a falsifiable prediction. Large cuts in API prices are public, discrete and set by a handful of providers. If the model is right, postings and posted wages for implementation roles should rise in the months after each cut, and by more where such people are hardest to find. If they do not, the argument is wrong.

The usual question is whether AI will replace workers. The more useful one is which workers AI makes more valuable as it gets cheaper. For now, the answer includes the people who put it to work.