The skill gap is not technical; it is knowing what an agent may do

The AI skill gap most workplaces face is not technical. Few staff will ever need to build or code an agent. Many will need to know what an agent is allowed to do, what evidence it relied on and when a person must decide instead. That is authority literacy, and it is the gap that matters for most jobs.

Why do people assume the gap is technical?

Because AI arrived as a technology story. The public conversation focused on models, data and engineering. Training followed: courses on how models work, how to write prompts, how to connect tools. Job seekers concluded that the way to become employable in AI was to become more technical.

For a small number of roles, that is right. Someone has to build and integrate agents. But most people who work alongside agents will never do that. They will receive an agent's drafts, approve its held items, answer customers about its decisions and notice when it goes wrong. None of that needs code. All of it needs an understanding of what the agent is permitted to do.

What does "knowing what an agent may do" involve?

Four ideas, each simple to state and harder to apply.

Depth. How far an agent may act on each capability: 1 read, 2 draft for a person to send, 3 execute routine work within bounds, 4 move money or take actions that are hard to undo, 5 coordinate other agents. An agent might draft refunds at depth 2 and send tracking updates at depth 3. A worker who does not know which is which cannot tell whether something has already happened or is waiting for them.

Bounds. Every action meets several limits at once: the agent's role, its certificate, its sources, the business's own limits and the platform's guardrails. The lowest wins. When an item lands in your queue, knowing which bound sent it there tells you what to do.

Verdicts. Allowed, held (waiting for a named person), refused (outside the agent's bounds) and withheld (refused at the platform floor for every agent). A worker who treats a refusal as a fault to work round has misunderstood the system.

Provenance. Where each fact came from: sealed from the system of record, signed by a named party, declared by someone unverified, or external. An external claim cannot authorise a payment. Knowing this lets a non-technical worker spot most of the cases that should be stopped.

Why is this the harder gap to close?

Because it is not taught by default. Technical courses are widely available. Courses on authority, oversight and evidence are fewer, and the ideas are newer. Many people working with AI tools have never been told that an agent's right to act is limited, earned and expiring.

It is also harder to self-teach. You can learn to code from tutorials. Learning what an agent may do requires seeing real decisions: what was proposed, what the bounds allowed and what a person decided. That needs a workspace and realistic cases.

And it is invisible when absent. A worker who lacks technical skill knows it. A worker who lacks authority literacy usually does not, until they approve something they should have stopped.

How does this fit with soft skills?

Closely. Authority literacy is a hard skill in the sense that it has specific content you can learn and be tested on; the page on what a soft skill is draws the line between the two. But applying it draws on qualities usually called soft: judgment, care, confidence to say no, clear writing. That mix is why people from non-technical backgrounds often do well.

The page on the most valuable AI skill for non-technical workers argues a similar point from a different starting place, and what a skills gap is explains the general idea.

What is the trade-off?

Technical knowledge still helps. Someone who understands roughly how models produce text is less likely to trust a fluent but wrong answer. Someone moving towards building agents will eventually need more systems knowledge. The argument here is not that technical skill is useless, only that it is not where most people's gap lies.

"Non-technical" also does not mean easy. Authority literacy takes study and practice. The four ideas above are simple to list, but applying them consistently under time pressure is a real skill, and employers can tell the difference between someone who has read about it and someone who has done it.

How do you explain this in an interview?

Plainly, and with an example. "I don't build agents. I know what an agent is allowed to do and how to tell when it has gone beyond that. For example, if a payment is held because the bank details came from an emailed request, I know the agent could not release it, and I know the check I need to make before I can." That kind of answer shows the skill in a sentence.

The page on how people with no technical background explain AI agent skills in an interview gives more examples. Whatever you claim, make it checkable, as argued in why every AI skill on your CV should be checkable.

What should you do differently?

If you have been planning to close your AI gap by learning to code, ask first whether the jobs you want require it. Most review, operations and customer-facing roles do not. Put your effort into the four ideas above instead: learn them, practise applying them on realistic cases and get evidence an employer can check. A free introductory course, such as Pryme Intelligence Academy's Foundations track, covers depth, the decision chain and certification in about four hours and is a sensible first step.

The overview of AI skills for job seekers sets out what else employers will ask you to prove.