Job Ads are Fluent in AI. Organisations aren't.

Job postings for roles most exposed to AI now ask for AI-related skills at roughly twice the rate of less-exposed roles. Nobody has yet shown that asking for the skill is the same as getting the capability.

A number from Bangkok

An IMF economist presented new research this week at the IMF-World Bank Group Annual Meetings in Bangkok, in a session on Asian labour markets. The analysis, built on roughly 700 million job advertisements, found that jobs highly exposed to AI carry close to double the demand for AI-related skills compared with jobs less exposed to it (Nation Thailand, 12 October 2026).

The skills employers are screening for fall into three groups: general AI literacy, specialised skills such as prompt engineering and working with AI agents, and the ability to connect AI tools to a company’s own internal data. That third one is the interesting one. It is not a skill you can certify from the outside. It depends entirely on what the organisation has already built.

The part the IMF couldn’t find

Here is what makes the finding harder to read as straightforward progress. Hiring demand for programmers actually fell by about 10 percent relative to other occupations over 2024 and 2025, even as AI-skill demand rose. And the IMF researchers said plainly that they have not yet found clear evidence that any of this AI-skills premium is moving wages, up or down, for the people who hold those skills.

So the job ad has changed. The thing the job ad is supposed to predict, namely whether the organisation gets more capable once that person is hired, hasn’t been shown to change at all. We’ve argued before that the operating model has to change before the skill set does; this is the same gap showing up at the hiring stage rather than the training stage. Developing Asia alone needs to create 150 million jobs over the next decade. A hiring criterion that hasn’t yet been shown to predict anything is a real problem to have at that scale.

Organising AI turned out to be the easy part

A separate piece of research published this month points at where the real difficulty sits. Ethan Mollick, who studies AI’s effects on knowledge work, wrote about watching AI agent systems organise their own work with less human design than he expected (One Useful Thing, 1 October 2026). In one example he cites, OpenAI used a swarm of agents to produce a mathematical proof in 88 hours, with the agents coordinating among themselves after humans set the goal and changed direction once.

The pattern he describes is that AI systems are getting better at figuring out how to organise themselves. What they still need from a human is the decision about where to point them in the first place. That is a judgement call, and it is exactly the kind of thing a job ad’s skills list cannot capture, because it depends on knowing what the organisation actually needs done, not what a model can technically do.

Where this shows up first

Put the two findings together and a specific prediction follows. The gap between what a job ad screens for and what an organisation can actually absorb will not show up as a pay gap first. Wages are slow to move and hard to attribute. It will show up as attrition and quiet underuse: people hired for AI fluency who spend their first year waiting for the organisation to catch up to the job description that hired them. It’s the same mechanism we’ve traced in why AI adoption stalls on managers, not employees: the skill gets hired for, but nobody builds the layer around it that would let it actually land.

I have watched this pattern before technology had anything to do with it. A new title gets written before the workflows, approvals, and manager habits around it change. The organisations that will show a return on this AI-skills premium are the ones treating the hiring criteria as the easy, fast-moving part of the system, and the slow, hard part as everything that happens after someone starts.

‍