In partnership with
AI didn't take your job. It quietly handed you a second one — unpaid — and this week the market put numbers on it. Today's issue is about getting that invisible work on the record before someone else prices it for you.
The shadow job — OpenAI read 800,000+ ChatGPT work messages: one in six is someone doing another occupation's task — and among job-specific messages, it's 43.5%
The cheaper-than-me math — research lab METR published a way to price the exact point where an AI agent becomes the cheaper hire — and where you stay the part that can't be automated
The repricing — Visa cuts 2,600 in a profitable year, while Class-of-2026 grad hiring is forecast up 5.6% — the split runs through your next review
The doubled engine — Claude's new model roughly doubled in capability with no change to the bill, plus the 5-minute prompt that puts it to work on your worst task
One in six work messages to ChatGPT is someone quietly doing a job they were never hired for
Tuesday, 4:52 p.m. A marketing manager pastes a Python error into ChatGPT and asks it to fix the code behind her campaign dashboard — the job of the data analyst whose role was never backfilled in March. She isn't a programmer, nobody asked her to become one, and her title and pay haven't moved.
OpenAI just counted how common she is. Its economics team read 800,000+ work messages from US ChatGPT business users and matched each one against the official list of what every job actually involves. The result: one in six work messages (16.8%) is someone doing another occupation's task.
Set aside the generic asks — "summarize this," "fix my grammar" — and look only at messages about specific job tasks. There, the crossover runs 43.5%. The borrowed work is skilled work: financial calculations and tech troubleshooting land in the top three borrowed tasks for every field outside their own.
Who works outside their own job the most | Of their job-specific AI messages, the share doing another occupation's tasks |
|---|---|
Customer experience | 77% |
Designers | 75% |
HR | 69% |
Engineers (the lowest) | 28% |
The study is OpenAI's own, not peer-reviewed, and it stops at the messages — it can't see titles or paychecks. Our read: work that has no name can't be rewarded, and when the reorg comes, it can't be defended. Naming it first, on your terms, is the whole game.
What it means for you:
→ The 2-minute shadow-role audit. Paste: "Here are the tasks I did this week outside my job title of [title]: [list them — 30 seconds of typing]. For each, write a one-line resume bullet and one sentence I can say in my next performance review." Email the best bullet to yourself for review season.
→ Say one out loud this week. "I've been covering our analyst work since March" in a 1:1 is visibility on your terms — before a reorg defines it on someone else's.
→ Manage a team? Before opening the next requisition, ask what's already being absorbed — the most-borrowed tasks (finance math, tech troubleshooting, marketing materials) may already be getting done off the books.
If AI already lets you do a job you were never trained for, the next question is uncomfortable: when does it stop needing you in the loop at all? Last week, that question turned into arithmetic.
"When is the AI cheaper than me?" is now arithmetic — run it before your boss does
Lunch break. An operations manager builds a quiet little spreadsheet: column A, what an AI agent would cost to run her weekly vendor report; column B, the minutes she'd spend catching its mistakes; column C, what her own hour costs the company. This month, for the first time, A plus B comes in under C — and she sees the opening: she found out first, so she gets to be the one who automates it.
The break-even she just found almost has an official name. METR — a nonprofit lab that stress-tests AI agents — calls its version the expenditure horizon: up to a certain budget of work, the AI agent is the cheaper hire; past it, you are.
METR priced agents and human experts in the same currency — dollars — on a hard machine-learning engineering problem that experts had spent two years competing on. The best agent it tested stayed the cheaper hire only up to about $3,300 of work per task. Last year's models: effectively $0. The newest models weren't tested yet, which means $3,300 is a floor, not a ceiling.
The expenditure horizon | Where the agent stops being the cheaper hire |
|---|---|
Last year's models | $0 |
Best agent METR tested | ~$3,300 per task |
The newest models | not yet tested |
Two details matter more than the dollar figure. Only 50–70% of what the agents produced was good enough to keep, and their self-reported wins shrank once humans re-checked them. That's why our worksheet adds the line METR deliberately leaves out: your checking time.
What it means for you:
→ Paste this — the AI supplies the estimates you don't have (2 min): "Here are my 3 most repeated weekly tasks: [list]. For each, estimate what an AI agent would cost per run, how many minutes I'd likely spend checking its output, and compare against my hourly cost of $[X]. Rank them: automate first, automate never."
→ The gears, if you want them: (AI cost per run + your checking minutes at your rate) vs. your cost to just do it. Where checking time dominates — judgment calls, client-facing work — you're the part that can't be automated.
→ Where the AI wins, automate that task yourself this week and put your name on the saving. The last section hands you the exact prompt that builds it. The person who runs this math first gets the credit; the person who never runs it gets the memo.
Companies run a cruder version of this ledger at company scale — and this week the job market printed both of its columns at once.
Speak naturally. Send without fixing.
Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.
Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.
89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.
AI isn't deleting jobs, it's repricing them: 2,600 cuts at a profitable Visa — and grad hiring forecast UP 5.6%
Tuesday morning, July 28. An all-staff memo from Visa's CEO lands across a 34,100-person company: about 2,600 roles — 7%, concentrated in technology and product — are being cut, with AI credited for having "reduced repetitive work and accelerated product development." This is not a struggling startup; the memo went out hours after another profitable quarter.
A week earlier, Monday.com cut 620 people — 20% of its workforce — in what its co-CEOs called "the most painful decision in the company's history," redirecting the savings into AI. The same market is buying on the other side: the campus-recruiting association NACE forecasts Class-of-2026 hiring up 5.6%. And when the Strada Institute — a workforce-research nonprofit — asked ~1,500 hiring leaders, nearly three times as many expect AI to increase entry-level hiring this year as decrease it.
The repricing, in three numbers | |
|---|---|
Visa, July 28 | 2,600 cuts (7% of staff) |
Share of US layoff notices naming AI (layoff tracker Challenger) | 7% of notices in January → 40% by May |
Class-of-2026 hiring forecast (campus-recruiting body NACE) | +5.6% |
The market has stopped asking how senior you are and started asking which side of the automation you're on. The paid side has a price tag: Glassdoor's average for an entry-level AI engineer is $128,013.
What it means for you:
→ Rewrite one resume line right now (2 min). Paste: "Rewrite this bullet so it leads with the AI workflow I built and the time it saves: [your bullet + what you actually did]." "Familiar with AI tools" is wallpaper; "built the workflow that saves 5 hours a week" is the profile getting offers.
→ If your role is the routine version of your title, the ledger above is your map — automate the routine slice yourself and become the person who did, before the company-scale version does it for you.
→ Advising a new grad — or are one? Lead the resume with one working AI project, not the GPA. Don't have one yet? The 20-minute playbook below is the project. Demonstrable output is what that $128K entry market is actually buying.
The durable seat on the paid side is "the person who wields the tool" — and days ago the tool got twice as good while the bill stayed flat.
The top scorer on AI's hardest test just doubled and your bill didn't move — learning to spend it is the cheapest raise you'll get
Sunday night. A freelance consultant drags three years of client reports — about 700 pages — into a single Claude conversation and types: "Draft next week's report the way I would." Twenty minutes of checking later, Monday's four-hour task is a 30-minute review; her rate didn't change, but her hourly income just did.
What made that possible is Claude Opus 5, released by Anthropic on July 24. Three things about it survive translation into plain English:
→ It holds your whole archive. One conversation now takes in a million tokens — roughly three-quarters of a million words, or years of your own documents at once — enough to feed it your back catalogue and ask for output in your voice.
→ It's roughly twice as capable on the hard stuff. On Frontier-Bench, an industry test built from problems most AI models still fail, it scored more than double its predecessor. It isn't ahead everywhere — on one coding benchmark, OpenAI's GPT-5.6 Sol still beats it.
→ Your bill didn't move. Per-use pricing is unchanged from the model it replaces, and on Claude's Max plan Opus 5 simply became the new default — so if you were already paying, the doubling came free.
That engine sits under everything above: the second job you're absorbing, the cheaper-than-me math, the repricing. The difference between reading about it and getting paid for it is about five minutes.
What it means for you:
→ The 5-minute version — do this one tonight: "Take the top task from my shadow-role audit: [task]. Ask me for the inputs you need, run it with me right now, and finish with a short checklist of what I should verify in your output."
→ The 20-minute weekend version: "Here are my 3 recurring tasks: [list]. For each, build a reusable playbook — the inputs you need from me, the steps you'll run, the checks I should do on your output. Then run task #1 with me now." It turns the break-even ledger into a working system.
→ Tool decision, made: on Claude's Max plan, Opus 5 is already your default — nothing to flip, nothing extra to pay. On ChatGPT or Copilot? Stay put and run the same two prompts there; the workflow is the win, not the logo.
That closes the loop: you were already working the second job — this issue was about getting it named, priced, and paid. Name the work, run the math, move to the paid side of the split, let the doubled engine carry the load.
Pick two. Start tonight.
☐ Paste the shadow-role audit and get your resume bullets drafted for you. (2 min)
☐ Paste the cheaper-than-me prompt on your 3 most repeated tasks — the AI supplies the estimates. (2 min)
☐ Rewrite your top resume bullet to lead with a working AI result, not "familiar with AI tools." (2 min)
☐ Run the 5-minute delegation prompt on your number-one task. (5 min)
Don't bookmark. Don't "save for later." Pick two. Start tonight.
If this named something you've been feeling, forward it to the colleague who's been quietly covering a second role since the last reorg — the one whose job outgrew their title.
Reply with one word — which one is you: shadow (working the unpaid second job), math (ran the cheaper-than-me numbers), split (felt the repricing), or double (tried the new engine)?
That's it for tonight.
See you Sunday.
— Rohit