Teams
Apprenticeship, Internship, or Rotation: Which Survives AI?
The choice between internship, apprenticeship and rotation comes down to ramp length. Where a beginner is useful in weeks, an internship still works as a short paid trial, and the unpaid version is rarely worth running. Where usefulness takes a year or more, run a rotation: full salary, several teams, a real conversion decision at the end. Where the ramp is licence-gated or regulated, an apprenticeship is the only shape with required hours, instruction and a rising wage written into it. AI shortened the tasks, not the ramp.
The takeThe line everyone cuts is the rotation, because it is the only shape that shows up as full salaries. That is backwards. What all three really buy is a named person willing to read a beginner's work and say what is wrong with it, and that supervision is the scarce good now, not the seat. A model will write the draft. Deciding a junior is ready to send one out unread is still somebody's afternoon. If the pattern holds, the programs still standing in five years are the ones that budgeted supervisor hours as carefully as headcount. Shape follows who is willing to teach.
Where Olive fits
Open a role and see what the work shows
The six dimensions Olive reads describe what a working ramp is supposed to produce: framing before generating, demanding a source for the claim a decision rests on, keeping the judgment that should not be handed over, building something between the brief and the answer, refusing output with a stated reason, and testing a claim against something outside the conversation. Olive is employer-purchased and built for hiring, so it settles those for a candidate on an assignment written for their occupation and read by a human reviewer, rather than tracking someone already inside a program.
Rank your shortlistWhich program shape survives AI-shifted entry-level work?
Rotations survive best, apprenticeships survive where the ramp is long and regulated, and internships survive as short paid trials rather than as talent pipelines. The reason is arithmetic. Each shape buys a different quantity of supervised time, and AI took most of the work a beginner used to be genuinely useful at in the first three months. Buy the shape that covers the gap between the start date and the day the work ships unread.
| Internship | Apprenticeship | Rotation | |
|---|---|---|---|
| Typical length | 8 to 12 weeks | 1 to 4 years | 12 to 24 months |
| What you pay | Intern rate, one summer | A rising wage schedule, plus instruction | Full salary from day one |
| What it buys | A look at the person | A defined skill set, certified | Context across several teams |
| Converts because | The summer went well | The hours are finished | They already know where things are |
| Fails when | The work was busywork | The occupation is not apprenticeable | Nobody owns the third rotation |
The internship is the cheapest look at a person and the thinnest evidence about them, and the AI shift hit it hardest. An eight-week project that ends in a slide deck nobody reads was always a weak signal; it is now a slide deck a model could have produced on day two.
The apprenticeship is the most expensive and the most durable, because it is a registered legal instrument rather than a name for a long internship. A time-based registered program requires at least 2,000 hours of on-the-job learning, recommends a minimum of 144 hours of related technical instruction for each year, and must carry a progressively increasing wage schedule with an entry wage at or above the federal minimum 3.
The rotation sits between them and is the only one of the three that treats the first year as a hiring decision already made rather than as an audition. That is also why it is the shape most often cut: twelve rotational hires at full salary is a real line in next year's headcount plan, and it outlasts the enthusiasm of whoever proposed it.
How long does your field take to make someone useful?
Long enough that the answer is usually published. O*NET sorts every occupation into five Job Zones by preparation required: Zone 2 needs a few days to a year of on-the-job training, Zone 3 needs one or two years combining on-the-job and informal training, Zone 4 needs several years, and Zone 5 mostly assumes the skills arrive with the person 2. Start from that number, then correct it with your own.
Your own measure is one date: the first month a new hire's work leaves the team without a second reader. It is observable, it differs by function inside one company, and nobody has to agree on a definition of ready to record it. Ask three managers for the last three hires in their function and you have the number in an afternoon.
What falls out is a rule that has nothing to do with headcount:
- Ramp under three months. An internship is proportionate. The program is a trial, and the honest output is a hire-or-not, not a curriculum.
- Ramp of six to eighteen months. A rotation. Anything shorter ends before the person is useful, which means you paid the full training cost and collected none of the return.
- Ramp gated by a licence, a supervised-hours requirement or client trust. An apprenticeship, because those ramps are defined in hours by somebody outside your company and the shape has to match.
The third case is the one people misread. A ramp gated by regulation did not compress when AI arrived, because what is gated is the authority to sign, not the speed of the draft. An audit senior still needs the hours. A paralegal still needs someone who will put their name on the position. This is the same triage that decides which roles actually need AI skills, run against time instead of task exposure.
What did AI actually take from the entry-level job?
The practice reps, not the job. Stanford's payroll analysis puts employment of 22-to-25-year-olds in the most AI-exposed occupations about 19% below where it would have been had it kept pace with less-exposed peers, and the gap arrives through reduced hiring rather than through separations 1. Declines concentrate in occupations where AI substitutes for human tasks; where it complements them, employment is flat or rising 1.
Read that as a statement about training, because that is what it is. The first-draft memo, the comparables screen, the meeting summary, the first pass at a ticket queue: those were never valuable output. They were the reps a junior person did badly for six months until they did them well, and they were affordable to a company because they also produced something. A model now does them at a standard that clears the bar the junior was being measured against.
So the entry-level job did not get easier. It lost its bottom rung and kept its ceiling. The work that remains for a person in month two is the work that was previously reserved for month eighteen: deciding whether the draft is right, noticing the number that cannot be true, refusing an answer that reads well. That is a harder job to start in, and it is why new grads who have never worked without AI can look fluent and be unable to say which part of their own output they would defend.
The program design consequence is blunt. A shape whose value came from cheap junior output (the unpaid internship, the two-month analyst pool) is buying something that no longer exists. A shape whose value came from supervised judgment under a named owner still works, and costs what it always cost.
What can each shape teach, and what can't it?
An internship teaches whether someone can be taught. It cannot teach judgment under consequence, and in the Second Circuit an unpaid one legally must not try: the seven-factor primary beneficiary test asks, among other things, whether the intern's work complements rather than displaces the work of paid employees, and whether the internship is limited to the period of beneficial learning 4. Consequential work is displacement. Pay the intern and the question goes away.
That is the real reason unpaid internships aged badly. The tasks that kept them defensible (shadowing, practice work, the deck for internal use) are the tasks a model does now, so the program is left choosing between illegal usefulness and legal irrelevance. A paid short trial escapes the trap entirely and is a better instrument anyway.
The apprenticeship teaches a defined skill set, verifiably. Its cost is rigidity: a written work-process schedule, an instruction requirement, a wage progression, a probationary period capped at 25% of the program length or one year, whichever is shorter 3. That structure is worth buying when an outside body already defines the ramp and worth nothing when the occupation is not apprenticeable. It also produces the only credential of the three that means something to the next employer, which is a retention risk and an honest recruiting advantage at once.
The rotation teaches context, which is the part AI has not touched. Knowing which team owns the pricing table, which number the CFO distrusts, and which customer complaint is a pattern rather than an incident is not in any model's training data about your company. It is also the reason rotations convert well: the person is not being hired at the end, they are being placed.
What none of the three teaches on its own is whether the person is any good with an assistant that will do all of it for them. That is a check, not a curriculum, and it belongs at the same checkpoints as everything else, which is a different decision from whether to hire the capability or train it.
Design the program around the check, not the curriculum
Write down what a person must do unsupervised at month three, six and twelve, then build the shape backwards from those three checks. A curriculum measures attendance. A check measures whether the ramp worked, and it is the only part of a program that can tell you to extend it, shorten it, or convert someone early. Decide the checks before you decide the length.
1. Name the unsupervised act, not the topic. "Files a variance commentary that goes to the controller unedited" is a check. "Understands the close process" is a memory of a conversation. 2. Put one check inside the AI work itself. Give the person a task with an assistant available and a wrong answer buried in the source material, then read what they did: what they framed before generating, which claim they demanded a source for, what they refused, what they tested against something outside the chat window. Self-report tells you nothing here, and neither does output quality: the output is the assistant's. 3. Cost the program per converted hire, not per participant. Twenty interns at 30% conversion and twelve rotational hires at 85% are not the same purchase, and the second is usually cheaper per person who is still there in year two. 4. Give every rotation a named owner with a hiring stake. The failure mode of every rotational program is the third placement, where the receiving manager did not choose the person and does not want them. Fix it by making the receiving manager sign. 5. Check the check. If everyone passes month three, the bar is decorative. Assessments built for early-career candidates fail the same way, and for the same reason: a bar written to be passed measures the writing.
One thing not to do: pick the shape first and reverse-engineer the checks to fit it. That is how a company ends up with a two-year rotational program in a function where people are useful in seven weeks, and an eight-week internship in a function that takes two years. Both are expensive, and only one of them is visible on a budget.
Common questions
Can an unpaid internship still work now that AI does the busywork?
Rarely, and the legal shape is the reason. The Second Circuit's primary beneficiary test weighs whether the intern's work complements rather than displaces paid employees, and whether the internship is tied to formal education and limited to the period of beneficial learning 4. The tasks that satisfied those factors are largely the tasks a model now does, which leaves an unpaid program choosing between work that is useful and work that is defensible. Pay the intern and the analysis stops mattering. A paid four-week trial gives you better evidence about the person and costs less than one bad hire.
How long should a rotational program be?
Long enough to cross the ramp plus one full placement. If people in the function are useful without a second reader at nine months, three placements of six months each is a program that ends before the third placement is worth anything to the receiving team. Two placements of nine to twelve months usually beats three of six. Judge it by whether the last rotation produces work the team would have paid a permanent hire to do; if it does not, the program is a tour, and a tour converts badly because no manager competes for someone they only watched.
Is a registered apprenticeship worth the paperwork for an office role?
Only when an outside body already defines the ramp in hours. A registered program carries real obligations: a written work-process schedule, at least 2,000 hours of on-the-job learning on the time-based approach, a recommended 144 hours of related instruction a year, and a rising wage schedule 3. That structure is a gift when licensure, supervised hours or an audit standard already dictate the sequence, and pure overhead when the ramp is whatever your manager says it is. For most office functions, a rotation gives you the same supervised time without registering anything.
What conversion rate should the program be judged on?
Your own last three cohorts, not a benchmark. The number that matters is cost per person still in the role at eighteen months, which folds conversion and early attrition into one figure and stops a program from looking successful because it made offers nobody accepted. Track offer rate, acceptance rate and eighteen-month retention separately, then divide total program cost by the survivors. A shape that converts at 40% and retains everyone can be cheaper than one converting at 80% into roles people leave.
Should an entry-level program require AI use or restrict it?
Require it, and check what the person did with it. Restricting it trains a way of working that does not exist in the job, and the graduate then meets an assistant for the first time with nobody watching. The useful constraint is not volume but evidence: the person should be able to say which claim they demanded a source for, what they refused, and what they tested against something outside the conversation. Volume of use is not a virtue. Someone who judged the model was the wrong instrument for a step and did it by hand made the right call.
How do you know the ramp actually worked?
One observation: work leaves the team without a second reader, and nothing comes back. Everything else (course completions, rotation feedback forms, a manager's confidence) measures the process rather than the person. Record the month it first happens for every cohort member, and the distribution tells you whether the program length is right within two cohorts. If half the cohort clears it at month four in a program built for twelve, the program is paying for eight months of supervision the work does not need.
References
- 1. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence ✓ digitaleconomy.stanford.edu ADP payroll data through June 2026: employment of 22-to-25-year-olds in the most AI-exposed occupations sits about 19% below its counterfactual, arriving through reduced hiring rather than separations, with declines concentrated where AI substitutes for human tasks and employment flat or rising where it complements them.
- 2. O*NET OnLine Help: Job Zones ✓ onetonline.org The five Job Zones and the preparation each requires: Zone 2 a few days to one year of on-the-job training, Zone 3 one or two years of combined on-the-job and informal training, Zone 4 several years of work-related experience, Zone 5 typically more than five years. Accessed 24 August 2026.
- 3. 29 CFR 29.5 - Standards of apprenticeship ✓ ecfr.gov A time-based registered apprenticeship requires at least 2,000 hours of on-the-job learning; a minimum of 144 hours of related technical instruction per year is recommended; program standards must include a progressively increasing wage schedule with an entry wage at or above the FLSA minimum, and a probationary period capped at 25% of program length or one year, whichever is shorter.
- 4. Glatt v. Fox Searchlight Pictures, Inc., Nos. 13-4478-cv, 13-4481-cv (2d Cir., amended opinion) ✓ govinfo.gov Adopts a non-exhaustive seven-factor primary beneficiary test for unpaid interns at for-profit employers, including whether the internship is tied to formal education, whether its duration is limited to the period of beneficial learning, and whether the intern's work complements rather than displaces the work of paid employees. No single factor is dispositive.
4 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.