Reshuffle∗

Singapore was the envy of the world.

Late spring, 2021. Virus-tight borders. Near-obsessive contact tracing. A vaccine rollout the rest of the world was envious of.

Then the second wave broke.

Not because the vaccine failed. Not because the contact tracing failed.

An off-books visa channel. A discretion-loving clientele. KTV lounges that quietly opted out of contact tracing.

The system had blind spots the tools could not see.

This is what AI would have seen.

Not a smarter contact tracer. A different layer of the system entirely.

Now watch what happens to the labels ↓

Same nodes. Same edges. Same hidden layer the system was blind to.

This isn't a COVID story. It's the AI story. It always was.

"It wasn't undone by a failed vaccine or flawed contact tracing, but by a system whose hidden fault lines and interdependencies couldn't be managed, even with the most accurate technologies or the most precise policies."

— Sangeet Paul Choudary, Reshuffle

Most of what people debate about AI — its IQ, its consciousness, whether it can pass a bar exam — is the wrong frame.

The story is what the system can finally see.

Sangeet goes deeper on this in Chapter 2 of Reshuffle ↗

This part is about your job.

Pick the role closest to yours in the simulator on the left. The numbers below change. The lesson does not.

A job isn't really a job.

It's a bundle of tasks — drafting, looking up, judging, owning the relationship, verifying. They travel together because, historically, they were cheaper to keep in one person than to split.

Drag the AI capability slider up.

Watch which tasks commoditize first. The first to go aren't the ones you'd guess.

Try it before you scroll ↑

The job didn't disappear. It unbundled and then rebundled — around three new constraints:

Human judgment, where the call is genuinely yours. AI orchestration, a new role that didn't exist five years ago. Risk and verification, because AI can't audit itself.

This is what Sangeet calls the reshuffle.

Now zoom out.

This same unbundle–rebundle dynamic plays out at the organisation level (which functions get pulled into a central platform team, which get pushed out). And at the value-chain level (which firms become the new orchestrators, which get commoditized into suppliers).

It's the universal mechanic.

You don't need an AI strategy. You need a strategy for the conditions AI creates.

— Sangeet Paul Choudary, Reshuffle

The work isn't predicting which jobs AI will replace. The work is identifying where the reshuffle creates new constraints — and positioning yourself there.

Sangeet's full taxonomy is in Chapters 4–7 ↗

By 2019, Alexa had won.

200 million Echo devices sold. 100,000 third-party skills. Amazon owned e-commerce, Prime, and the most ambitious voice ecosystem ever built. Every visible advantage.

Then you tried to order a pizza.

"Alexa, order a pizza." She told you to enable the Domino's skill. Then link your Domino's account. Then say it again — but with this precise syntax. Cannot ask follow-up questions. Cannot chain to another skill.

By turn three, the smartphone was faster.

Alexa asked partners to build on her platform. She never asked them to work together.

100,000 skills. All isolated. Each one a private contract between Alexa and a partner. None of them connected to each other. The platform never solved the coordination problem its users actually had.

Then ChatGPT showed up.

Same fragmented internet. Same composite user intent. But LLMs handled the connection — they composed actions across services without anyone having to agree on a schema first.

This is coordination without consensus. AI's real ecosystem superpower.

"Control points aren't created through dominance. They're created through dependence."

— Sangeet Paul Choudary, Reshuffle

Who wins when AI restacks the knowledge economy is not whoever owns the most customers, the most data, or the most distribution.

It's whoever earns the right to resolve the coordination burden that no one else can.

Sangeet's strategic taxonomy is in Chapters 8–12 ↗

The robot was ready in 1990.

It could pick, weld, and assemble faster, cheaper, and more precisely than any human. The "lights-out factory" was technically feasible by the early '90s.

Thirty years later, most factories still have humans on the line.

The robot itself wasn't the problem. The system around it was.

Workforce retraining. Supply chain integration. Exception handling. Capital expense. Each one a cost the customer had to absorb before the tool delivered anything.

Tools externalize complexity. They don't absorb it.

Solar power lived this exact gap.

By 2010 photovoltaic panels were efficient and cheap. Adoption was stuck — homeowners couldn't finance the $30k upfront cost. Then someone invented Power Purchase Agreements: install free, pay only for the energy generated.

Better panels didn't unlock solar. Better financing did.

A solution has three traits:

Accessible. Easy to find and adopt without specialised expertise.

Usable. Fits real workflows. Doesn't force the user to rewire their day to use it.

Reliable. Works when it matters. Failure isn't a tutorial step.

Anything that fails one of these is still a tool. Not a solution.

"AI value capture is less about technology and more about managing risk."

— Sangeet Paul Choudary, Reshuffle

Tool providers compete on what AI can do. Solution providers compete on whether the customer can absorb it.

The second contest is harder. It's also where the value lives.

Sangeet on tools vs. solutions — Chapter 9 ↗

In 1971, the barcode was just a checkout tool.

Both Walmart and Kmart adopted it. Same head start, same scale, same access to the new technology.

Most store managers thought they were getting faster checkout lanes. Few realised they were staring at the future of retail.

Kmart used the barcode to speed up checkout. The data stayed at the register.

Walmart used the barcode to build a system. A private satellite network. Real-time data flowing to suppliers. Automatic restocking based on what just sold.

Same tool. Different architecture.

The data didn't just speed up the queue. It shifted who controlled the supply chain.

Walmart pressured suppliers to meet tighter delivery windows. Standardised packaging. Integrate with their logistics. Brands could no longer dictate shelf placement based on marketing power — they had to prove themselves with every scan.

The customer-facing interface became the most valuable control point.

By January 2002, Kmart filed for Chapter 11 with $4.6 billion in debt.

Today Walmart is the largest retailer in the world. $611 billion in revenue. 2.1 million employees.

Same technology. Opposite outcomes.

"The difference wasn't the barcode itself, but the architecture each company built around it."

— Sangeet Paul Choudary, Reshuffle

Today every company has access to the same AI models. The same APIs, the same vendors, the same off-the-shelf agents.

The fork is the same fork. Whoever builds the architecture around the model — the data flows, the supplier coordination, the customer-facing decision point — wins the next two decades.

Sangeet on power and architecture — Chapter 3 ↗