Over the past year, nearly every major AI vendor decided that selling software was not enough. Amazon, Microsoft, Google, OpenAI and Anthropic have all launched units that send their own engineers to live inside a client's company and build the AI system by hand. Microsoft alone put 2.5 billion dollars and 6,000 people behind its version. AWS committed 1 billion dollars. Add Google, OpenAI and Anthropic's versions and the industry has pledged more than 15 billion dollars to this one idea in under a year.
The job title behind all of it, forward deployed engineer, did not come from nowhere. Palantir invented the role two decades ago to serve intelligence agencies whose data was too sensitive and whose needs were too unusual for a normal software sales process. Those engineers moved into CIA and Army facilities and wrote code on-site because nothing else worked. Every major AI company has now copied that playbook for ordinary businesses.
Gartner, the research firm most large companies rely on for technology advice, is not convinced the copy works as well as the original. It predicts that 70 percent of AI systems built this way will be thrown out by 2028. The reason is not that the technology fails. It is that the arrangement is structured badly from the start: companies get fast early results, then discover they have no one on staff who can run, fix, or improve the system once the outside engineers move on to the next client.
The price makes this worse. Gartner estimates a single project can cost more than 200,000 dollars every three months in fees for the engineers alone, separate from whatever the software itself costs. That bill does not stop once the system is built. Someone still has to maintain it, and if that someone is the vendor, the company is paying rent on its own AI system indefinitely.
There is also a straightforward shortage. Gartner counts about 2,000 of these engineers working today against demand that is close to four times that. When supply is that tight, quality slips. Gartner's own analyst noted that some companies have started calling ordinary consulting work "forward deployed engineering" simply because the label sells better, without the skill or project discipline the real thing requires.
None of this means the underlying push to get AI working inside real businesses is wrong. Most companies genuinely cannot do this alone yet; a widely cited MIT study found that the vast majority of corporate AI projects deliver no measurable financial return, mainly because the tools never get woven into daily workflows. Outside help often is the fastest way to get something running.
The lesson for any company considering this path is to treat the contract as the real product, not the engineers. Insist on a plan for who takes over once the vendor leaves, insist on owning the code and the knowledge, and ask hard questions about whether the firm pitching "forward deployed engineers" actually has that rare skill set or just a new name for the same consulting team.