Prompt engineers and SQL mechanics
Flatlining demand for prompt engineers reveals that enterprise AI is not stalling, but quietly vanishing into the software companies already use.
By Hugh Askell
Sparked by AI Mentions in Job Descriptions – August 2026 · discussion
There is a minor panic playing out across Hacker News this week over a new chart from Corvi Careers showing that Fortune 500 job postings for 'Prompt Engineers' have effectively flatlined. The immediate conclusion being drawn by the ecosystem is that enterprise AI adoption has stalled, or at least crashed hard into a wall of corporate incompetence. This is a remarkably Silicon Valley way to misunderstand how IT departments actually buy things. We have succumbed to the widget fallacy, assuming that a broad computational utility naturally requires a dedicated human operator standing next to it pulling levers—a sort of digital loom worker.
When Oracle commercialized the relational database in 1979, it represented a staggering leap forward in computer science. Suddenly, you could structure, query, and manipulate vast pools of enterprise information with abstract mathematics. Yet if you looked at the HR data for a mid-western logistics company or a high-street retail chain in 1985, you saw exactly zero job postings for 'SQL Engineers'.
Normal companies filter technological breakthroughs through what we might call the Database Test. They ignore the raw technological primitive, preferring to buy a business process that simply happens to have the new math buried somewhere deep inside the code. No one handed a blinking SQL terminal to the regional distribution managers. They just waited until the technology vanished.
Databases achieved true macroeconomic ubiquity a decade later, in the 1990s, precisely because they disappeared into enterprise resource planning software like SAP R/3. To the logistics company, this was just an invoice for better supply-chain software. The technology passed the Database Test by ceasing to be an open-ended science project and becoming a standardized procurement line-item.
Generative AI is sliding down exactly this same industrial curve, and doing so on a timeline that breaks real-time hiring metrics. Gartner notes that a typical buying group for a complex B2B solution involves six to 10 decision makers, a bureaucratic gauntlet that routinely takes 12 to 18 months to navigate. The typical non-tech corporation is entirely uninterested in rolling its own language model, and they are certainly skipping the step of hiring thousands of bespoke prompt engineers to type queries into a naked text box.
Silicon Valley envisions a pipeline where a legacy company buys a raw LLM and hires new specialists to operate it. In the enterprise reality, an incumbent SaaS vendor buys the model, embeds it into a proprietary workflow, and the legacy company simply upgrades their existing software tier. Accordingly, Gartner projects that more than 80% of enterprises will consume generative AI primarily via embedded applications by 2026. The intelligence gets abstracted away into a button labeled "summarize" or "generate report" inside a dashboard the company already pays for.
So who actually holds the leverage when the technology becomes invisible? If generative AI is an embedded utility passing the Database Test, the legacy workflow owner—who already controls the distribution channel to the enterprise buyer—captures the entire margin. Microsoft bypasses the prompt engineer entirely (you don't need a prompt engineer when the application itself quietly constructs the prompt in the background based on your calendar invites)—they simply bolt an extra $30 per user, per month tax onto existing Office 365 seats. The artificial intelligence remains completely invisible to the human resources department. It is, however, highly visible to Microsoft's revenue line.
Do we measure the triumph of mobile computing by counting iOS developers at Ford, or by acknowledging that every Ford employee has a smartphone in their pocket? If AI is just another computational utility passing the Database Test, there is no reason to expect a permanent caste of corporate whisperers. When relational databases rewired the global economy in the 1990s, trucking companies kept their dispatchers right where they were. They just bought software that worked better, and let SAP keep the margin.