Audit Your Own Job for AI (the Prompt I Run on My Role)

Stop asking "will AI take my job." Ask "which tasks inside my job should it already be doing." The data backs the reframe. Anthropic's Economic Index tracked the split all year: in January 2025, augmentation (AI helps, you decide) led automation (AI does it) 55% to 41%; by August automation had crept ahead 49% to 47%; then in November 2025 augmentation retook the lead at 52% to 45%. Two things are true at once — automation's long-run share keeps rising, and most people still reach for AI as a partner, not a replacement. Nobody's handing over a career. They're handing over tasks. As an EM, that's exactly how I think about delegating to a team, so I run the same audit on myself.

Why task-level beats job-level

Almost every role is a thin slice of what you were actually hired to do, buried under a pile of overhead. Salesforce's State of Sales puts a number on the sales version of this: reps spend only about 30% of the week actually selling. The other 70% goes to updating records, chasing data, and prepping for meetings. Your job has a different ratio, but the same shape — a little high-value core, a lot of connective busywork. That 70% is the automation prize, not the 30%. So the useful question isn't "can this job be automated," it's "which slices of this week can, starting today."

The two questions behind every bucket

I score each recurring task on two axes, and two axes only: how much of my judgment does it genuinely need, and what does a mistake actually cost. Run every task through those two questions and it falls into one of three buckets.

• Automate fully — repetitive work with a low cost of error. A wrong output is annoying, not damaging. This is where AI runs end to end, ideally on a schedule or a trigger. For me that's formatting the weekly sprint summary, sorting incoming asks into the right board, turning release notes into a changelog, first drafts of routine status messages. If getting it slightly wrong just means a five-second fix, hand the whole thing over.
• AI assists — work that needs a sliver of your judgment, but where the machine should carry most of the weight. It drafts, you sign off. It researches three vendors, you pick. AI builds the interview debrief, you make the call. You stay in the loop, but your minutes on the task collapse — an hour of prep becomes a five-minute review. Most of a manager's week lives here once you're honest about it.
• You stay in control — work that moves a relationship or carries a real consequence. Performance conversations, a tense negotiation, a hiring yes/no, an incident post-mortem where trust is the actual product. AI can brief you before you walk in, but it does not act inside these moments.

The prompt below is deliberately stricter than your gut, because the gut always guards too much. It interviews you first — you don't need a clean task list to start — then sorts, then ranks.

Copy this prompt

Paste it into Claude with five quiet minutes (before the workday starts is ideal). It interviews you, then sorts and ranks.

You are my automation auditor. Interview me, then sort my work week into three
buckets and hand me a ranked plan. Be decisive: most people under-automate out of
habit. Push back when I claim a task "needs me" without a real reason.

STEP 1 — INTERVIEW (one question at a time, wait for my answer):
1. My role, and a rough hour-by-hour normal week.
2. What I do every single day without fail.
3. Weekly/monthly work: reports, reviews, planning, 1:1s.
4. Tasks I procrastinate on most.
5. Where I copy info from one place to another.
6. Tasks that touch another person's feelings, money, or trust.

STEP 2 — SORT each task into exactly one bucket:
- AUTOMATE FULLY: repetitive + low error cost. AI does it end to end.
- AI ASSISTS: needs some judgment. AI drafts/researches/preps, I approve;
  my time on it must drop >=70%.
- I STAY IN CONTROL: moves a relationship or needs real judgment. AI only briefs.
If I push a task to "control" with no relationship/judgment reason, challenge me
once with the honest counter-argument, then respect my call.

STEP 3 — DELIVER:
1. A table: task | bucket | hours/week | one-line "how" (scheduled prompt, saved
   Claude skill, a connector, or an agent) for every AI-touched task.
2. Top 3 to automate first, ranked by (hours saved) x (ease of setup). For each:
   plain setup steps I can do this week + the first ready-to-paste prompt.
3. Honest total hours/week I get back from just the top 3, and the bucket-three
   work I should reinvest that time into.
4. One task I probably think is automatable but isn't yet — and why.

What good output looks like

You want four things back, not a vague pep talk. A table of every task with its bucket, the hours it currently eats, and a one-line "how" — is this a scheduled prompt, a saved Claude skill, a connector to your calendar/email/docs, or a full agent. A ranked top three, scored by hours saved times ease of setup, each with steps you can actually finish this week and the first prompt written out ready to paste. An honest hours total you'd claw back from just those three. And the one task you think is automatable but isn't yet — that last item is the one that saves you a wasted weekend, so don't skip it.

After the audit: three rules that make it stick

Ship exactly one thing first. The audit hands you a top three, but week one is only for the #1. One automation that genuinely works beats three half-wired ones you abandon by Thursday — and the first real win is what funds the motivation for the next two.

Turn every bucket-two assist into a saved prompt. If "write my weekly update" is an assist, don't rewrite the prompt from scratch each Friday. Write it once — your format, your tone, the sources it pulls from — and turn that into a reusable Claude skill (or just a pinned note you paste when you need it). That's what actually drops the task from an hour to a review pass, permanently.

Defend bucket three on your calendar. The entire point of clearing buckets one and two is to pour the recovered hours into judgment work: deepening relationships, thinking about strategy, making the high-visibility calls that move your career. If the freed-up time silently refills with more admin, the audit changed nothing. Block the hours before they evaporate.

Do this today

Run the prompt, then ship only the #1 item this week. Re-run the whole audit every six months, because none of the buckets are fixed. A task slides from "assists" over into "automate" as the tools get better and your trust grows, and the people who re-audit twice a year keep finding hours everyone else leaves on the table.

Sources

• Anthropic Economic Index — Cadences (June 2026 report)
• Anthropic Economic Index — Economic primitives report
• Salesforce — State of Sales