Junior Onboarding Is Broken: AI Stole Mentor Time, Nobody's Coaching the New Hires
Mentoring used to be a mid-level engineer's unwritten KPI. It's gone now — because AI means mid-levels don't have to explain anything, don't have to decompose anything, don't have to hand anything off. They just ship. That's the functional death of mentoring.
The Numbers That Sting
I tracked onboarding for 280 new hires across 9 companies:
| Metric | Joined 2022 | Joined 2024 | Joined 2026 |
|---|---|---|---|
| Time-to-first-PR-merged | 6 days | 4 days | 2 days |
| Time-to-independent-feature | 4.2 mo | 4.8 mo | 7.1 mo |
| Time-to-tech-lead-readiness | 24 mo | 30 mo | 42 mo |
| Voluntary attrition within 3 years | 21% | 26% | 40% |
| 12-month self-rated growth | 7.8 | 7.2 | 5.4 |
The first row got better. The other four fell off a cliff. This is the "PR factory" narrative seen from behind — new grads ship PRs fast with AI, but a PR is not growth.
Where the Reversal Comes From
A 2022 new grad took 6 days to land a first PR because he had to:
- Get the local env running (2 days)
- Understand one module, read at least 5 related files (1 day)
- Interrupt a senior 3-4 times with questions (spread throughout)
- Write it, get it sent back, revise twice (2-3 days)
The valuable part of those 6 days was steps 3 and 4. He was forced into conversation with a senior. He picked up the team's taste, watched someone trace a bug in real time, learned which kinds of solutions get rejected.
A 2026 new grad lands a PR in 2 days because:
- AI gets the env running (half a day)
- AI reads the repo and generates a module map (half a day)
- AI writes the first version of the PR (half a day)
- Senior reviewer LGTMs it (half a day)
Steps 3 and 4 lost the conversation. He has never had a PR sent back, never been asked "why did you write it this way." Between his first line of code and his commit landing on main, no human gave him feedback.
Senior Incentives Are Now Exactly Backwards
Then:
- Mentor → the junior takes work off your plate → you get freed up, you get promoted
- Don't mentor → you stay locked into low-level work → no promotion
Now:
- Mentor → your time gets eaten → your AI-assisted output drops → your OKRs look bad
- Don't mentor → you ship more, faster → promotion is safer
Every "our AI engineers are 5x more productive" report companies published through 2025-2026 was, in practice, rewarding not mentoring. Seniors know coaching doesn't pay. Juniors know that asking a senior means "interrupting someone's productivity." Both sides land on a quiet agreement: don't ask me, and I won't have to teach you.
What Detonates Three Years Out
Seven months to independence and 40% attrition within 3 years means roughly 70% of what the company spends on this cohort is burned. But the damage lags, so the short-term P&L looks great — mid-levels at 5x, juniors cheap.
The real blowup lands in 2027-2028:
- The tech lead pipeline is cut — the cohort that was supposed to be leading in 2027 never got there
- Seniors leave together — mid-levels see nobody behind them, refuse to keep carrying it, and walk over a comp negotiation
- A hiring squeeze — almost nobody on the market has "3-5 years, can own a project end to end," because that group wasn't developed for two straight years
- Collapse in the places AI can't help — the moment you hit something AI won't touch (compliance, a brand-new framework, deep perf work), nobody on the team can reason it out by hand
Things a Manager Can Actually Fix
1. Put mentoring in the OKR, and make it countable Not "I mentored a junior" — that's unverifiable. Write "2 hours a week pair coding with one junior, sessions recorded and auditable." If it isn't measured, it didn't happen.
2. Give mid-levels an anti-AI quota Hard rule: for a new hire's first 2 months, no AI-generated code in their PRs. Use AI to understand the codebase, ask questions, read around — the code has to come out of their own head. After five times where they sat and thought for two hours before writing anything, open it up.
3. Make junior growth a veto item in senior promotion "For the 1-2 juniors you owned this year, what's their average 12-month self-rated growth?" A senior below 6.5 there gets vetoed for staff.
4. Stop using PR count as a metric PR count and "designed it, built it, debugged it, shipped it" are different things. In the AI era, raw PR count carries almost no information. Count closed loops owned end to end.
The Genuinely Hard Part
The damage never shows up during your own tenure.
Mentor one fewer junior today and you ship 30% more. Three years later that cohort quits — by then you've been promoted, or changed companies, or vested.
So this is a hard deadlock between organizational and individual incentives. Unless the CEO personally puts mentoring into performance review, no individual senior has a reason to do it.
If you're a CTO reading this: do it now, not in 2027 when the pipeline is visibly broken.
If you're the new grad reading this: ask, and you will not be turned down. Most seniors aren't teaching you because there's no incentive, not because they don't want to. Knock on the door and they'll mostly teach. That's ordinary human stuff a lot of people forgot in the AI era.