4,000 People Lost Their Jobs At Block. Dorsey Blamed AI. Here's What Actually Happened.

Block’s job cuts tied to AI miss the deeper point: AI doesn’t just automate tasks inside today’s org chart—it erases the coordination layer that makes many “knowledge work” roles necessary in the first place. The result is a structural shift in what work is, why it exists, and who gets to spend time on value creation.

In software and product companies, most day-to-day activity isn’t the product itself. It’s coordination overhead: writing PRDs and specs so decisions can be executed by people not in the room, running sprint planning and status updates to synchronize work across humans, preparing decks for executives who can’t read primary sources, filing tickets for cross-person tracking, and managing onboarding documents, handoffs, and ceremonies like retros and postmortems. Code and shipped products are the value; everything else is the cost of producing that value with an execution layer made of people.

Agentic coding systems change the equation because they can move from insight to code in a tight loop. When agents can read the codebase, progress artifacts, and git history quickly, the “handoff” problem shrinks dramatically. Onboarding tax—transferring institutional context from one human brain to another—falls away because agents can reconstruct context directly from inspectable artifacts. The same goes for coordination tax: if the state of work is continuously inspectable (often as code), there’s less need for separate meetings, documents, and status rituals.

The transcript anchors this in time-use data. Microsoft’s 2025 Work Trend Index is cited for an average employee spending 57% of time communicating and 43% creating, while Sana’s Anatomy of Work study lands near 60% “work about work.” Meeting time is described as 11.3 hours per week on average, tripling since 2020. The argument isn’t that people are failing at productivity; it’s that organizations have built mechanics—meetings, artifacts, alignment ceremonies—to compensate for human bandwidth limits and context fragmentation.

Once translation layers between humans disappear, work moves closer to the product itself. Instead of PMs shaping documents for engineers to interpret, PMs and designers can shape the actual artifact directly, iterating on design, implementation, and verification in the same workflow. The transcript points to Anthropic discussing a world without PRDs, where code becomes the living state of the project and separate ceremonies like sprint planning and design reviews lose their purpose.

This shift also creates a compounding “flywheel.” As agents take over execution, remaining work becomes more verifiable—because it’s expressed as machine-testable artifacts—so agents can do more, which reduces coordination further. That changes forecasting too: AI doesn’t merely replace some tasks; it changes the distribution of work by deleting coordination-heavy structures.

The transcript anticipates the pushback that “judgment-heavy work” stays human. It argues that much of what looks like judgment is actually verifiable preparation plus a thin layer of non-verifiable synthesis. Agents can automate the prep and leave the genuinely human parts: product vision, brand and meaning, customer care, systems-level architecture under uncertainty, and the design and debugging of agentic systems themselves. The people most likely to thrive are those with “agency” (treating it as a skill issue) and “ramp” (fast learning and curiosity).

The central takeaway is framed as good news: AI is showing that the job was never the job. By removing coordination overhead, organizations can concentrate human effort on higher-value work—work that gets closer to customers and the product, not closer to meetings and documents.