Britain is losing the work that teaches graduates judgment. Rebuild the first rung

THE SIGNAL IN ONE SENTENCE

The first rung was never glamorous. It was how you learned not to fall. Every profession has work that senior people remember as tedious. A trainee accountant checks a schedule line by line. A junior lawyer compares clauses. A new analyst builds the first ugly spreadsheet. A graduate engineer writes the draft that comes back covered in comments. These tasks produce something useful. They also produce the professional. That second output is easy to miss when an AI tool can finish the visible work in seconds. Remove the first draft, basic coding, routine research and standard reporting, and a company may gain speed. It may also delete the repetitions through which a new employee learns what a good answer looks like, where exceptions hide and when a confident output deserves suspicion. Professor Mohamed Abdel-Maguid, a pro vice-chancellor at Canterbury Christ Church University, put that problem plainly in an essay published October 26 by the Higher Education Policy Institute. His argument is not that employers should preserve busywork. It is that institutions need to replace the apprenticeship function that routine work used to provide. The evidence suggests the redesign has begun before the training plan is ready. The Institute of Student Employers surveyed 144 employers across sectors in January and February 2026. It found that 87 percent expected AI adoption to reshape graduate and apprentice roles. Forty-three percent said entry-level roles had already evolved in response to AI without being formally redesigned. Those are employer reports, not a causal experiment. They do not prove how much work AI removed or how many skills graduates lost. They do establish an awkward gap: a large share of employers see the job changing, while many have not rebuilt the pathway on purpose. PwC offers a second clue, with an important geographic label. Its 2026 AI Jobs Barometer analysed 2.4 million entry-level job postings in the United States. The entry roles most exposed to AI were seven times more likely than the least exposed roles to ask for traditionally senior, human-intensive skills such as leadership, creativity and face-to-face interaction. PwC says postings for these seniorised entry roles grew 35 percent since 2019, while other entry-level postings fell 10 percent. That is US posting data, not a measurement of British graduates. Still, it describes the same structural risk. Employers can ask a 22-year-old for judgment while automating the work that used to help build it. Call it the seniority paradox. The junior role becomes more interesting on paper. Less copying, fewer routine summaries, more client contact and harder decisions. That sounds like progress. But the person is now expected to exercise judgment before the organisation has shown how that judgment will be formed. Good judgment is not a personality trait that arrives with a degree. It is compressed experience. You review an output, make a choice, see the consequence, hear why your reasoning failed and try again. You encounter enough ordinary cases to recognize the strange one. You learn which rule can bend, which cannot and who needs to be called before either happens. AI can support that loop. It can generate cases, offer alternatives, expose contradictions and let a trainee practise more often. It can also short-circuit the loop if the trainee becomes a courier who moves model output from prompt to client. The useful redesign starts with an inventory. Take every junior task being automated and write down the hidden capability it used to teach. Drafting a memo may teach issue spotting and audience awareness. Reconciling figures may teach materiality and anomaly detection. Basic coding may teach how a system breaks at its boundaries. Client intake may teach which facts people omit until someone asks the right question. Then build a paid replacement for that learning. One option is supervised simulation using cases drawn from real work but stripped of private data. Give the trainee incomplete evidence and require a decision, not a polished paragraph. Let AI help, but ask the trainee to mark what they accepted, modified and rejected. A reviewer should discuss the reasoning before revealing what happened in the original case. Another option is progressive authority. Start with low-consequence decisions and a close review. Move to real client exposure with a senior person present. Increase scope only when the trainee can explain evidence, uncertainty and escalation triggers. The goal is not to keep someone junior forever. It is to make the path to independence visible. Universities can help by assessing judgment rather than merely banning tools. Give students unfamiliar evidence, incomplete records and conflicting claims. Ask them to produce a recommendation and an AI-use receipt. What did the tool contribute? Which source changed the conclusion? Where did the student disagree? That is harder to fake and closer to the job. Employers need receipts too. A company announcing an AI efficiency gain should be able to show what happened to the learning embedded in the removed task. Which capability did the task develop? What replaced it? Who supervises the replacement? Is the opportunity paid? How is progression measured? What happens when the tool is wrong? This is where equity becomes operational. If professional formation migrates into unpaid internships, private courses, family connections and extracurricular credentials, the missing first rung has not been rebuilt. It has been moved behind a toll gate. Students with money and networks will still find practice. Everyone else will be asked to arrive experienced. The UK government has already created a place to work on the problem. Its Early Careers Jobs Alliance, backed by £20 million, brings together government, employers, trade unions and young people. It begins in digital and technology and is meant to map how entry-level work is changing, identify good practice and help businesses redesign roles while maintaining routes into employment. That alliance should demand concrete artifacts, not another cloud of principles. Publish model job descriptions that pair AI capability with structured learning. Create sector-specific judgment ladders. Fund small employers that cannot build training systems alone. Measure access, completion, error recovery, promotion and retention. Include apprentices and graduates in the design, because they can tell the difference between a learning opportunity and cheap work wearing a nicer badge. The strongest caution is also the simplest: none of the current figures proves that AI alone caused a decline in entry-level hiring or a measurable loss of professional skill. The ISE survey records expectations and reported role changes. PwC analyses job-posting patterns in the United States. Abdel-Maguid's HEPI article is an expert argument that connects these signals. It is persuasive, but it is not a controlled study of British career outcomes. Different sectors will move at different speeds. A hospital, law firm, software company and local council do not share one apprenticeship model. Simulations can help, but they cannot reproduce every consequence of real responsibility. Mentors also need time, authority and recognition for teaching. A dashboard will not rescue a programme nobody is staffed to run. The wrong response would be to defend repetitive work because repetition once had a useful side effect. The better response is to keep the learning while improving the task. Let the machine prepare the first pass. Give the graduate the evidence, the exception and the decision. Require them to challenge the output. Put a real mentor on the other side of the review. Increase authority as judgment becomes visible. That is a first rung worth rebuilding.