What Agentic AI Actually Means for Careers Guidance, and Why the Difference Matters

There is a phrase that gets used a lot in education technology right now, and it rarely means what people think it means. “AI-powered” has become a label you can put on almost anything, from a chatbot that matches keywords to a genuinely reasoning system that checks data, takes actions, and only then replies. For a sector that is trying to solve a real and urgent problem, that distinction is not academic. It is the whole ballgame.

The problem is well documented. Alan Milburn’s review of young people and work described careers guidance as deeply unequal, and the single most cited barrier young people face when trying to get into work. Government data this year revealed over 32,000 so-called “phantom NEETs,” teenagers whose activity local authorities simply cannot account for. Meanwhile, one careers adviser is often responsible for hundreds of students, spread across year groups with wildly different needs. The people working in this system are not short of commitment. They are short of hours in the day.

Our first version got this wrong

This is where the conversation about AI gets interesting, and where it also gets misleading. A chatbot that answers “what is a degree apprenticeship” with a static, pre-written paragraph is not solving the capacity problem. It is just moving the same generic advice into a different interface. We built our first version of an assistant this way, and it worked about as well as you would expect. It matched intents, it replied with templates, and it treated a Year 11 student unsure of their next step exactly the same as a Year 13 student mid-way through an interview process.

What changed the picture for us was building something agentic rather than conversational. The distinction is simple to state but harder to build. A conversational system generates text. An agentic system decides what the question actually requires, goes and checks something real, whether that is a student’s live application pipeline, a set of current vacancies, or a school’s own engagement data, and only replies once it has an answer grounded in that data.

Honesty over hype

The difference shows up most clearly in the moments where a generic system would be tempted to fabricate. When a student asks for digital marketing apprenticeships in London and none exist on the platform, the honest answer is that there are not any right now. A good coach does not paper over that. Ours checks, says so plainly, and then broadens the search to find the closest genuine match elsewhere in the country. That small moment of honesty over hype is, in our experience, the difference between a tool students trust and one they quietly stop using.

The same principle applies on the school side, where the operational cost of good careers guidance is often a spreadsheet problem in disguise. Asking how many students are at interview stage, or what a school’s Gatsby benchmark scores currently look like, used to mean an export and some manual cross-referencing. We built our school assistant to answer these questions directly, scoped automatically to each school’s own data, and in testing it classified and answered all the educator questions we put to it correctly. That is not a small thing for a careers lead already stretched across a full teaching timetable.

This does not replace a human adviser

None of this is a substitute for a human adviser, and we do not think it should be marketed as one. The system we built is explicit about its own limits. When two students share a similar name, it asks which one rather than guessing, because guessing about a child’s record is not a risk worth taking. Emotional support is handled the same way: acknowledge how the student feels first, then move toward something practical, rather than either dismissing the feeling or leaving the student stuck in it.

Devolution changes who is responsible, not what is possible

There is a broader policy point buried in all of this. As skills and employment support get pushed further into the hands of mayors and combined authorities, and as tools like the government’s new Risk of NEET Indicator try to catch young people earlier, the systems doing the day-to-day work with students need to be able to act on data, not just describe it. Devolution changes who holds responsibility for a young person’s outcome. It does not automatically create the operational capacity to do anything differently at the individual level. That capacity has to be built, and increasingly, it is being built through tools like this one.

We are publishing the full detail of what we built, including unedited examples from live testing, because we think the sector benefits from seeing what agentic AI actually looks like in this context, rather than relying on the marketing language that surrounds it. The technology is not the interesting part. What it lets a careers adviser, a school, or a young person actually do with their time is.

By James Kimble, Head of Partnerships at Apprentago