What Anthropic's Economic Index Quietly Says About Who Gets Paid

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Anthropic's Economic Index did not read like a product launch. It read like a quiet map of who keeps getting paid when models get cheap. The slides that stuck with me were not capability curves. They were about which tasks compressed into hours, which roles absorbed the leftovers, and which organisations still needed a human who could say no without getting overruled by a dashboard.

What the Index actually measured

The notes framed the Index around usage patterns, wage-adjacent tasks, and where Claude showed up in real workflows — not a leaderboard of clever prompts. That framing matters because clients keep asking me for "AI ROI" as if a single chat transcript were a P&L line item they can wave at a board.

I treat the Index as a prior: which job families show high model substitution pressure, and which still look like complementarity. My proposals then price the complementary work — harness design, evals, access control — instead of pretending a subscription replaces a team with a shared Slack channel and good intentions.

Hours compressed are not headcount deleted

When a three-week engineering task collapses into three hours, the spreadsheet screams savings. The operational reality is messier. Someone still owns the definition of done, the security review, and the apology when the agent was confidently wrong at 4 p.m. on a Friday.

I rewrite client asks that say "replace two juniors" into "compress the tedious middle while keeping a named owner." If they refuse the named owner, I walk. Unowned agents are how quiet incidents become loud press, and I do not want my name on that postmortem.

Who gets paid in my engagement model

I bill for brief design, golden-set construction, connector scoping, and incident runbooks. The model call is a line item, not the product. When a CFO asks why they are paying me if Claude is cheap, I show the Index-shaped answer: cheap tokens amplify whatever process you already have — including the broken ones that nobody documented.

Complementarity roles in my shop: threat modelling, data classification, human-in-the-loop UX, and teaching operators when to override. Those do not show up as autocomplete. They show up as fewer 2 a.m. pages and fewer "the bot sent what?" messages in the executive channel.

Pakistan context without the TED talk

Local clients often hear "AI" as either magic or mass layoffs. The Index helps me stay boring: some reporting, drafting, and triage work will compress; relationship-heavy sales, regulated decisions, and physical ops will not. I staff pilots accordingly and refuse to promise a fantasy org chart.

I also refuse to use the Index as cover for underpaying people who now "only" supervise agents. Supervision without authority is a trap. If the human can be blamed for the agent's mistake, they must be able to stop the agent mid-flight — with a big red control, not a polite Slack nudge.

How I brief from an Index slide

Pick one job family. List the tasks that look substitutable versus complementary. Attach a permission matrix. Attach a success metric that is not "number of prompts sent." Run a two-week measurement before anyone talks about headcount or "efficiency gains" in a town hall.

If measurement is refused, the conversation was never about economics. It was about optics. I do not optimise for optics with production write access sitting one OAuth scope away from a confused agent.

What I keep on my wall

A one-pager: substitution pressure by task, complementary work I sell, and the kill criteria for any agent pilot. When a vendor deck contradicts the Index with vibes and a stock photo of a smiling robot, the wall wins.

Cheap intelligence still needs expensive judgment. The Index did not make me cynical. It made me price the judgment correctly — and walk when a client wants the savings without the judgment.

A conversation I refuse

I refuse "the Index says we can cut twenty percent, so draft the layoff list." The Index is a map of pressure, not a warrant. My counter-offer is always a two-week measurement on one workflow with complementary roles named in writing. If they want the warrant without the measurement, they need a different consultant.

I also refuse to hide behind the Index when a pilot fails. Failed pilots are usually brief and permission failures, not proof that complementarity is dead.