The industry's bottleneck moved from model capability to deployment, and money follows the bottleneck. Engineers who can land a model inside a live business are scarce, so the role pays a shortage premium: real now, and temporary.
The shape of an argument, not measured data. Pans sideways on a phone.
An engineer from a company with a famous model is sitting in a business park, waiting on a data export. When it lands, four per cent of the customer-ID column will contain a person's name: somebody built a form without validation in 2014. Her teammates at headquarters are shipping into a clean repo. Theirs looks like the more serious work. It isn't.
Why the job exists now
- The model is readyGood enough to sit inside a real workflow
- The company is notMessy data, bespoke auth, an unwritten definition of done
- The gap is a jobSomeone has to stand in it and ship
The models became deployable faster than customer environments became ready. Every frontier lab now hires forward deployed engineers (what the job is), and the field had its own track at AI Engineer World's Fair 2026; Latent Space's trends recap names it among the themes of enterprise adoption.
This has happened before
Electrification, 1890s to 1920s, after Paul David's account.
Robert Solow saw the computer age everywhere except the productivity statistics. Paul David's answer was electricity: factories swapped the steam engine for one big motor and gained little, and the gains came decades later, when the building was rearranged around a motor per machine. The model is the motor. Rearranging the work is this job.
The money changed hands
| The pilot era | Now | |
|---|---|---|
| What was scarce | Model access, GPU quota | People who can land one inside a live business |
| What the buyer asks | “Show us what it can do” | “What is in production, and who opens it every morning?” |
| Who signs | An innovation budget | A line owner with a target and a deadline |
| What gets you hired | Can you build it | Can you find the real problem, then ship the small version |
Nothing in that table was a model release. A demo can be built by someone who never meets the customer; adoption cannot.
What agents took, and what they left
Widths are illustrative, not measured.
Turning a well-specified ticket into correct code is the part whose price is falling fastest. This job is the complement: decide what to build where nobody can tell you, then own whether it works. That is also why the interview spends its time on the customer round rather than on syntax (The Customer Round).
Why the window shuts
A shortage premium is real because it is temporary.
IBM's systems engineers went from free with the machine to a priced profession after the 1969 unbundling, then to an industry with ladders and queues. This is the weakest of the claims here: if the tooling improves faster than expected, the gap closes sooner. But the residue is other people's history, and no tool absorbs that.
Five questions for the first call
| Ask | Good answer | Bad answer |
|---|---|---|
| How many accounts at once? | “One, sometimes two during a handover.” | “It depends on the quarter” |
| Who carries the pager after go-live? | “We do for a set period, then a named owner there.” | “Support picks it up” |
| What happens to the code? | “The third time we build it, it becomes product.” | “It stays in the customer's repo” |
| Travel, as a number? | A number of weeks a month | “Some travel” |
| Who was promoted out of this team last? | Names, and where they went | “The team is quite new” |
The last one is the most diagnostic. A team that cannot name its own promotions is one where the work is illegible to the promotion committee.
Do you fit
- Has the half that takes years
- Needs one project where you met the users and cut the scope
- Has the room, the scope, the trust
- Needs one production system you shipped and then carried
- Has the instinct for evaluation
- Needs a tolerance for plumbing and scanned documents
Nobody arrives with both halves. The missing half is what the interview tests (the loop, stage by stage).
Free starting points: the system design framework and your first agentic system. Then failure modes, human in the loop and cost per request.
She will get her export. Two days later she will find that the four per cent tracks a legacy sales channel, which tracks the customers who complain most, which is why nobody trusted the dashboard. That finding is worth more than the model was.
Customer contact reads as the tax you pay to reach the real work. Preparation does not fix that. The premium is temporary, the travel is real, and the metric is partly outside your control.
Why does the argument say the premium is temporary, and what would make it end sooner?
It is a shortage premium: few people can do the work yet. It compresses as a standard loop, a prep industry and better tooling create supply; faster tooling closes the gap sooner.
Agents write more of the code each year. Why does that strengthen the case for this role rather than weaken it?
They cut the cost of writing code, not the cost of knowing what to write. The share of the work that is finding the real problem and owning the result grows.
A hiring manager cannot name anyone promoted out of the team. What does that tell you?
The team's work may be invisible to the people who decide promotions, so the role may be a dead end inside that company, whatever the pay.
- The bottleneck moved from model capability to deployment, and the pay followed it.
- Electrification is the precedent: the gains came from rearranging the work, not from the motor.
- Agents made code cheaper and left the expensive half, knowing what to build, unchanged.
- The premium is a shortage premium; ask the five questions before you take it.