Topic

AI and work

What changes when software does part of the job: roles, teams, and who checks the work.

20 essays

A layer cake with a wedge missing rests on a thin metal board. The removed slice sits beneath the board as the sole pedestal supporting the much larger cake.The Strike at the Bottom of the Stack

Everyone keeps arguing about whether AI will take our jobs. The livelier fight is over who gets what the boom produces, and it's breaking out at the physical chokepoints the AI economy can't copy, move, or route around: a fab in Taoyuan, a zoning board in Ohio, a memory line the models can't exist without.

An unprinted blister pack contains six miniature magnifying glasses. One foil pocket is peeled open, with its magnifying glass partly pulled out.It's Not the Job That's Dying. It's the Discount.

American employment looks healthy, and the smart commentary says AI refuses to kill jobs. But what got squeezed from both ends this autumn is not an occupation. It is an arbitrage: the discount that made Bengaluru the default address for routine tech work. My bet is that the arbiter of who does tech work is no longer wages or visas but the cost of supervision, and the world's largest IT services employer is already selling its answer by the unit.

A tilted glass pitcher pours a broad ribbon of cream, terracotta and plum liquid. The three layers inside the pitcher remain distinct stripes as they bend through the spout and flow downward.The New Product Surface Is Where the Work Turns

The next AI product advantage may not be a better answer. It may be better placement: carrying customer and domain context into the moment where work turns into action.

A silver necklace made entirely of large lobster clasps lies in an oval on brown leather. At the front, one clasp is open, with the neighboring clasp disconnected just outside its mouth.The Unit of Creation Is No Longer the Asset. It's the Operating Loop.

Adobe’s June 2026 expansion is not “another creative AI feature drop.” It is a product shift that relocates the surrounding workflow into the suite itself. When creative and marketing platforms can execute outcome specs with editable artifacts and clear escalation boundaries, the unit of creation b…

A bicycle wheel stands on a pale green floor. All its steel spokes fan from the central hub into the lower-right quarter of the rim, leaving the rest of the wheel’s interior empty.Hire the Phase, Not the Title: What the Mini-CEO CMO and the AI Role Boom Have in Common

Read the ChatGPT Work marketing page with the titles erased and it stops looking like a product page. It looks like an org chart. OpenAI describes a loop: customer insights and brand standards in, briefs and assets out, performance reports back in, recommendations for the next launch. The frontier…

A glass cloche covers an ivory-frosted cake. Its thick magnifying front enlarges a jagged crack exposing dark cake crumb, while the decoration at the sides appears small and pristine.When the Artifact Arrives Finished: ChatGPT Work and the New Acceptance Loop

ChatGPT Work feels like a marketing speed upgrade until you notice the real shift: the agent collapses drafting into finished multi-format artifacts, and the only remaining human step is deciding whether to ship. That acceptance gate becomes the new work primitive.

A leather walking boot has a rubber sole that extends from its heel into four accordion-folded cream paper panels, each embossed with the same boot-tread pattern.The 11% Problem Is a Product Spec, Not a Readiness Gap

I keep coming back to IBM’s “11% readiness” stat. Most coverage reads it as a human literacy gap. My read is different: the missing piece is the product instrument panel that makes agent autonomy observable, auditable, and re-approvable. Treat the control gap as a spec, not a training problem.

A brass key lies on pale stone. Twelve overlapping, worn key heads are fused into a broad cluster that narrows into one shaft and a single bright, toothed blade.Domain Expertise Is the Agent Multiplier. Institutional Memory Is the Interface.

Across roughly 400,000 Claude Code sessions, Anthropic found domain expertise predicts how much useful work an agent produces per instruction. OpenAI’s Codex adoption shows the same bottleneck moving into departments with real domain fluency. When execution gets cheaper, power shifts to what your o…

A tobacco-brown tailored jacket lies flat, with a broad panel extending from its right side and opening into cream reinforcement, seam allowances and copper-colored lining.When the Loop Runs Itself, Acceptance Becomes the Job

OpenAI’s autonomous chemistry demo and Adobe’s creative agent expansion look like pure capability leaps. But the real shift is leadership. When the loop plans, executes, and iterates, humans move from performing steps to designing acceptance: what the loop is allowed to do, what artifacts prove out…

An ivory porcelain dinner plate assembled from tightly fitted jigsaw pieces, with interlocking seams continuing across its concave center and raised rim, rests on dark wood.The Micro-Work Episode: Why Desktop Agents Need a Stopping Rule, Not a Chat Window

Gemini Spark on macOS turns desktop agents from “chat you keep alive” into workflow infrastructure. But the advantage only lands when you redesign work as micro-work episodes: bounded inputs, a spec, a named done artifact, and a stopping rule.

A charcoal sweater lies flat with its lower torso unravelled into loose yarn. A bright red cord passes through a row of exposed stitches, holding the intact chest and sleeves above the unrav…When the Creative Loop Never Stops, Approval Stops Working

I keep coming back to a leadership shift almost nobody is naming directly: once the creative loop never stops, approval becomes the bottleneck. Jasper’s end-to-end GEO agent makes brand representation a managed, always-on surface. The real question is whether your team can govern that loop with ass…

An ivory running shoe lies on its side, revealing a dark rubber sole molded as one giant fingerprint. Beside it, a matching fingerprint-shaped tread impression is pressed into reddish clay.When the Loop Runs Faster Than the Record

The debate about AI ad labels misses the frontier shift: creative loops can now iterate, test, and publish faster than humans can reconstruct an auditable archive. In 2026 politics, narrative coherence becomes an operator-owned discipline, not an external compliance fix.

A red retractable tape measure lies extended across pale wood. Its steel blade is blank except for a short cluster of black measurement ticks beside the end hook.The Agent Productivity Gap Is a Measurement Problem

The best agent stack scores barely above 30 percent on real-world post-production tasks. The gap is not a model verdict. It is an instrumentation verdict. Scaffold your agents with trace fields, step-conditioned evaluation, and weekly outcome review so your team can learn instead of hope.

A fine silver necklace lies on aubergine felt, its chain attached to a large opened oyster shell containing a single cream pearl.The AI SDR Debate Is a Category Error. The Handoff Artifact Is the Product.

The AI SDR argument is stuck on autonomy versus human selling. The real bottleneck is later: what the agent hands to a human at the exact moment a buyer decides whether the risk is worth it.

A large rounded gray boulder overhangs the seat of a wooden dining chair. The seat bows beneath it, but the chair remains intact with all four feet on the floor.Goal Mode Just Changed What 'Done' Means

OpenAI made Codex goal mode broadly available across the app, IDE, and CLI. Gartner simultaneously established enterprise AI coding agents as an evaluable category and named OpenAI a Leader. Read together, they flip the operating contract: “done” can no longer be a sentiment. It must be an inspecta…

A mustard-yellow yarn ball on a blue surface has a strand extending left into an open loop, widening into a knitted band that returns and wraps around the ball.The Background Agent Economy: Why Persistent Execution Is the New Operating Model

The most important agent in your company is the one you never open a chat window with. Gemini Spark runs continuously inside Gmail, turning the inbox into persistent execution. The advantage is no longer what the model can answer. It is what the loop can do, remember, and prove when no one is watch…

A navy suit jacket hangs against an aubergine wall, its sharply tailored silhouette covered with exposed canvas chest panels, ivory reinforcement tapes and outward-facing stitched seams.The Conversion Test: What Cisco's Backlog and Layoffs Really Demand

Cisco’s $5.3B AI infrastructure backlog and simultaneous layoffs look contradictory to the public. The market read a single decision: reallocate capital and labor toward agentic workflows, then accept the short-term optics. The problem is measurement. This essay introduces the conversion test: dema…

Thick sourdough slices stand together as an arch on a limestone slab, with dark crust around the outside and pale crumb framing the open center. Small stone stops brace its feet.The AI Workforce Bifurcation: Redesign or Retreat

GM laid off over 10% of its IT department while hiring for AI-native development, data engineering, and agent workflow roles. The real signal is workforce bifurcation: AI advantage compounds only when leaders redesign decision rights, evaluation artifacts, and accountability close to the work.

A severely pruned citrus tree stands without a pot, its few leafy branches ending in pale cut stumps above a dense cylindrical mass of circling exposed roots.The Real AI Jobs Crisis Is a Design Problem, Not a Trade Problem

Cisco’s $5.3B AI backlog alongside thousands of layoffs is not a proof that AI “replaces jobs.” It is a proof that organizations are accumulating integration debt: the gap between buying AI speed and redesigning work. The operator move is a job-family selector audit plus evidence gates before scali…