The clearest measure of whether companies are actually using artificial intelligence may turn out to be a hiring line. Employers in the United States posted more than 7,300 jobs for forward-deployed engineers in the first nine months of this year, against fewer than 200 in all of 2023, according to labor analytics firm Lightcast, which compiled the figures for CNBC on October 15.
The advertised median salary is $195,456, more than three times the national median for all jobs. The title describes an engineer who sits inside the customer’s organization, adapting a model to how a specific business works, and who is graded on whether the customer keeps buying rather than on whether the software was delivered on time.
The three largest vendors are staffing the role at a scale that has no parallel in enterprise software. Amazon Web Services has committed $1 billion to its program. Anthropic has set aside $100 million and says it will train 10,000 forward-deployed engineers by the end of 2027. OpenAI’s deployment business began with more than $4 billion of investment.
Where the openings sit is as revealing as how many there are. Deloitte has posted 872, the most of any employer, and Accenture has posted 688. Google is third with 499. Consulting firms led their AI ambitions with headcount long before they led with revenue, and the demand for engineers who can translate a model into a working process has landed on the firms that already sell into corporate operations.
The role exists because models do not deploy themselves. A bank that wants to route claims through a language model has to connect it to systems built decades ago, decide what the model is allowed to see, and define what a correct answer looks like in a domain where the definition varies by department. That work is done by people, on site, over months.
The hiring wave is coming at a moment of doubt about the spending that produced it. This was the week when investors sold the companies that supply artificial intelligence infrastructure, after Barclays downgraded several data center hardware makers on October 14, citing valuations and the timing of enterprise adoption. Dell fell 11 percent that day, and chip makers including Nvidia, AMD and Micron were sold in premarket trading on October 15.
Forward-deployed hiring is the counterargument. If the problem were that customers could not find a use for the software, the industry would not be paying nearly $200,000 a year for engineers whose job is to find one. If the problem were that the software works and customers simply need to buy more chips, the hiring would not be concentrated in firms that sell services alongside models.
The economics of the model are unproven. A forward-deployed engineer is a cost per customer that scales with the customer base, which is the opposite of the software margin that made the industry famous. Vendors accept it because enterprise contracts are large and because a failure to reach production is the most common way an AI sale is lost.
The training commitments suggest the companies expect supply to stay tight. Anthropic’s plan to train 10,000 engineers in roughly two years is an attempt to manufacture a labor pool that universities are not producing, and it is a bet that the role will persist rather than be automated away by the products it installs. The same firms are building agents meant to do parts of this work.
Analysts said the salary figure is itself a signal. A median of $195,456, advertised, means employers are competing for a small number of people who understand both machine learning and how a specific industry runs, and that they have concluded it is cheaper to pay that sum than to lose a deployment. Advertised pay is also a marketing number, and the true cost of a hire includes the months before an engineer produces anything.
For customers, the effect is that the cheapest path to a working AI system now runs through a vendor’s staff. That inverts the usual expectation that software arrives finished. It also means a company that buys from two vendors ends up with two sets of embedded engineers, each with an interest in the tools they were sent to install.
For employees, the role pays well and demands relocation, travel and tolerance for customer politics. It also builds a skill that is not tied to one model, which is unusual in a field where the underlying technology changes every few months.
What the numbers do not show is how many of these positions were filled, how long deployments took, or whether the customers renewed. The count of openings is an input, and the output that matters is revenue per customer, which none of the vendors has broken out. Until they do, the hiring line is the best available evidence that the software is being used for something.
