Count the genuinely new jobs the last three years produced.
Prompt engineer, which was a real title for about eighteen months and is now a skill nobody lists. AI ethicist, which mostly existed already under other names. Head of AI, which this playbook has argued you should not hire.
The one that stuck is the forward-deployed engineer.
It started as a Palantir term for the people who went and sat inside the customer, learned the domain, and built against reality rather than a specification. This month it went mainstream. Brianne Kimmel circulated what she called the best overview yet written on the role, carrying the line that has been repeated everywhere since: FDEs eat pain and excrete product. Aaron Levie followed a day later, saying forward-deployed engineers are real, are not going away, and are happening at a scale never seen before.
It had been building. In July one investor described the profile every founder claims to be hiring for, and it is openly mythical: a computer science undergraduate who then spent two years at McKinsey and two more in product or business operations at a Series A startup. Almost nobody has that resume. Everybody is recruiting for it.
Why now specifically
Because the bottleneck moved, and it moved somewhere a product cannot reach.
The models work. The evidence across this playbook is unambiguous on that. What does not work is the gap between a capable model and a company whose processes were designed around people. Closing it requires knowing what the work actually is, which is written down nowhere and is frequently not what the org chart claims.
That is not a software problem, so it does not get solved by better software. Levie's framing is right: most enterprises need substantial support to apply model breakthroughs to their own workflows, and that support does not scale the way a SaaS seat does.
Which is why the market's shape is changing underneath. Sequoia's version is that the next very large company sells work, not software. Mercor went from $1M to $2B in run-rate in twenty-four months selling human judgment by the hour. AI consultancies are being acquired.
There is a good historical rhyme here. In 1996 every strip mall had a computer guy who set up your network and charged a monthly fee to keep it alive. SaaS spent twenty years killing that trade by making software configure itself. AI is bringing it back one tier up, because the thing that now needs configuring is not the software. It is the company.
Where I sit
This is my business model, so treat what follows as a practitioner's account rather than analysis from outside. I embed with one company at a time and ship function by function. Two things I would tell anyone considering the role, or the hire.
The job is three jobs. Understanding the business reality, which is mostly listening and takes longer than anyone budgets. Exercising judgment about what should exist, which is where most of the value is created and none of it is visible. Then building it. Most people are strong at one, passable at a second, and quietly outsource the third. The rare part is not the engineering.
The failure mode is becoming staff. An engineer who embeds and never productises anything is an expensive contractor with better tools. The discipline is extracting the reusable thing from every engagement, so the second company takes half as long as the first and the fifth takes a fifth. Without that, nothing compounds and you have bought yourself a job.
That second point is the one the current enthusiasm is going to get wrong. The market is about to hire a lot of very good people into roles with no mechanism for compounding, and in eighteen months a lot of them will be running the same playbook from scratch for the fourth time and wondering why it feels like agency work.
If you are the one hiring
Stop looking for the mythical resume. It does not exist in the volume you need, and the two-years-at-McKinsey part is a proxy for something you can test directly: can this person sit with your operations team for a week and come back with a description of the work that your operations team agrees with?
That is the whole screen. Someone who can do that and also build will outperform any credential combination on the job description.
AI made the models work. Somebody still has to make the company work.
Function by function, in your office.
Book a free Diagnostic: 30 to 45 minutes, no deck, no pitch. We pick one function, describe how the work actually happens today, and find the loop worth rebuilding first.
Book the Diagnostic →