Redesign the entry-level engineering role for an AI-era team: a week-by-week apprenticeship alternating AI-free fundamentals with AI-verification reps, a mentor load model, and the business case that answers "just stop hiring juniors" — using the junior-employment evidence honestly. The provided draft plan has a hole you must find and fix.
Grace Nakamura is Director of Engineering at Pinewood Fintech, a fictional 110-engineer company whose CFO just asked the quiet question out loud: "If AI does what juniors did, why are we hiring juniors?" It is not a rhetorical question — research on early-career employment in AI-exposed occupations shows real declines for 22–25-year-olds, on the order of −9 to −10% in the most exposed roles. Grace's position, which this lab defends with a spreadsheet rather than sentiment: the evidence argues for redesigning the junior role, not deleting it — because the tasks juniors learned on are automatable, but the seniors every org depends on still have to come from somewhere. An HR-provided draft plan is on her desk. It contains a structural flaw you must catch.
COST-MODEL SHEET (provided by finance, annualized)
Junior (L1) fully loaded: $118K
Mid (L3) fully loaded: $176K
Senior (L5) fully loaded: $234K
Contractor (mid-equiv) blended: $205K + ramp cost each engagement
Mentor time cost (current program): ~4 hrs/wk per junior x $112/hr
Junior time-to-net-positive (2023 baseline): ~9 months
Attrition: juniors promoted in-house stay 3.4 yrs avg vs
2.1 yrs for lateral mid-level hires
Senior pipeline note: 61% of current L5s were L1/L2 hires here
External evidence pack: early-career employment in AI-exposed
roles down ~9-10% (2022-2025); firms report shifting entry
tasks to AI; senior demand unchanged or rising
HR DRAFT: "AI-NATIVE JUNIOR PROGRAM v0.8" (excerpt)
Wk 1-2: Tooling bootcamp - master Copilot-style IDE + chat
workflows; ship first AI-assisted PR by day 4
Wk 3-6: Feature delivery with AI pair - juniors own small
features end-to-end using AI for design, code, tests;
velocity targets from week 3
Wk 7-10: Agent operations - juniors supervise coding agents on
backlog tickets; measured on tickets closed
Wk 11-12: Capstone - ship a feature 'at mid-level velocity'
using the full AI toolchain
Assessment: weekly velocity + AI-tool proficiency rubric
Mentorship: 'AI-first support model' - juniors ask the AI
before asking humans; mentor hours cut to 1/wk (cost saving)
Tagline: "Productive from day 4, mid-level output by month 3"
Audit the draft plan first — fifteen minutes, no AI. Read v0.8 the way you read Lab 05's PRs: what does a junior know at week 12 that they didn't at week 1? List what the plan trains (tool operation, AI-assisted shipping, agent supervision) against what it never trains. Write the one-sentence diagnosis of the structural flaw before continuing — the answer key checks whether you found it independently.
Name the flaw and its mechanism. The plan trains juniors exclusively with AI from day 4 — there is not one AI-free rep in twelve weeks. The mechanism to articulate: verification skill is downstream of fundamentals. A junior who has never debugged unassisted, never written the code themselves, never held a mental model of the system cannot tell plausible AI output from correct AI output — they can only measure it by whether it runs. The plan manufactures exactly the Gustav profile from Lab 12 (fluent operation, absent verification), then hands it agent supervision in week 7 and calls the output "mid-level velocity." Add the second-order flaw: cutting mentor hours to 1/wk and routing questions to the AI removes the human feedback loop that catches all of this.
Design the alternating apprenticeship — the core artifact. Twelve weeks, week-by-week, built on an explicit alternation: AI-free fundamentals reps (reading unfamiliar code, debugging without assistance, writing small features by hand, incident shadowing) interleaved with AI-verification reps (reviewing AI output against specs with seeded flaws, Lab 05-style provenance review, prompt-then-verify cycles on real tickets). A defensible shape: weeks 1–2 fundamentals only (codebase, debugging, one hand-written PR); weeks 3–8 alternating blocks (e.g., AI-free Mon/Tue, AI-assisted-with-verification Wed–Fri, each AI-assisted PR requiring an author walkthrough); weeks 9–11 supervised AI-leveraged delivery with sampled deep review; week 12 capstone with a verification exam — find the seeded flaws in an AI-generated change — alongside the delivery demo. Assessment: skill-change on both tracks (Lab 12 instrument), never velocity in the first eight weeks.
Build the mentor load model. Restore and structure the human loop: 4 hrs/wk per junior in weeks 1–6 (daily 30-min review of the AI-free work — that is where fundamentals are actually installed), tapering to 2 hrs/wk by week 12; one mentor per two juniors max; mentor time named in sprint capacity and recognized in calibration (Lab 03/11 logic). Cost it honestly in the sheet: the delta between 4→1 hrs and your model is the number the CFO will ask about, so compute it — roughly $17K per junior per year at the provided rate — and carry it into the business case as a cost you chose.
Build the business case — use the evidence honestly. This is graded on intellectual honesty: the −9–10% early-career decline is real and you must not explain it away. The argument that survives: (1) the decline shows the old junior role — ticket-grinding on automatable tasks — is genuinely dying; hiring juniors into that role is indefensible, which is why the redesign exists; (2) the senior pipeline math: 61% of Pinewood's L5s were internal juniors, internally-grown engineers stay 1.3 years longer, and the market for laterals prices at $176–234K plus ramp — model the 5-year cost of "stop hiring juniors" (all-lateral senior supply at rising market rates, plus the option value lost) against the redesigned pipeline (junior cost + mentor cost + faster time-to-net-positive with AI leverage, plausibly 9→6 months with verified fundamentals); (3) the risk case: an org with no junior intake for 3 years has no L3 bench in year 4 and buys everything at senior-market spot prices. Run the sheet both ways; present the range, Lab 09 discipline.
Define the redesigned role itself. One page: what an L1 at Pinewood now does (verification-grade review contributor, AI-leveraged delivery under sampled review, fundamentals reps as protected work — not a smaller mid-level); what they are measured on (skill change, verification catch-rate, explanation quality); what they are explicitly not measured on (velocity vs AI-equipped mids, tickets closed by their agents); and the promotion signal to L2 (passes the verification exam, explains every line they ship, catches a real seeded-or-natural flaw in review).
Red-team the business case as the CFO. Prompt: "You are a CFO. AI tools close most of the junior-output gap. Attack this case for continuing to hire juniors." The two attacks to answer in the memo: "contractors cover gaps without pipeline cost" (answer: $205K blended plus per-engagement ramp, zero retention, zero pipeline — the sheet shows the crossover) and "the Harvard-style numbers prove the market has decided" (answer: the market data shows firms deleting the old role; it says nothing about firms that redesigned it — and Pinewood's own 61% statistic is the cost of following the herd).
Assemble the pack. Role definition, week-by-week apprenticeship plan with the alternation visible, mentor load model with costs, business-case sheet with both scenarios and the range, and a one-page memo to the CFO leading with the honest read of the evidence.
A junior program that is AI-assisted from day one does not produce faster juniors — it produces confident operators with no way to know when the machine is wrong, which is the most expensive employee profile you can manufacture.
A complete lab has all of the following:
AI didn't make juniors obsolete — it made the old junior role obsolete. The pipeline survives by alternating hand-built fundamentals with verification reps, keeping humans in the feedback loop, and defending the whole thing with a spreadsheet that concedes the market data and wins anyway.