Backend · Frontend · Architecture

AI Made Individuals 10x Faster. Teams Are Still Stuck.

By HiSt · Jul 3, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Organizations are pouring money into AI seats expecting a productivity revolution and getting only localized speedups. The real payoff requires restructuring collaboration interfaces, not just node performance—a lesson that separates teams who actually ship faster from those who just generate more drafts.

Summary

Individual workflows have been supercharged by AI, but organizational throughput remains flat because the real constraints are the dependencies between people—reviews, approvals, and handoffs. Three classic distributed-systems laws explain the ceiling: Amdahl's Law caps speedup at the non-parallelizable collaboration portion, Brooks's Law shows communication costs grow quadratically with headcount, and the Theory of Constraints proves that accelerating non-bottleneck nodes only piles up queues. AI was applied to the nodes while the edges stayed untouched.

The fix is to service-enable upstream capabilities so downstream teams can self-serve instead of waiting. A design team giving marketing a skill to generate its own materials, or a data team exposing a self-service dashboard builder, turns a blocking dependency into an API call. This mirrors Amazon's 2002 API mandate, but today the marginal cost of writing a skill is low enough that teams can start without a CEO's decree.

A common fear is that teaching colleagues will make you redundant. The shift is from a ticketing model—where your value is measured in tickets processed—to a PR model, where you become the owner of a specification. Anthropic's content design team did this with Clontent, an internal agent that handles 70% of copy tasks and escalates judgment calls to humans, making the team's expertise more visible, not less. The approach has clear boundaries: don't skill-ify decisions that require accountability, low-frequency collaborations, or anything that won't have a dedicated maintainer.

Takeaways

— Individual AI acceleration of 3–10x leaves team output cycles unchanged because coordination dependencies remain untouched.

— Amdahl's Law mathematically caps organizational speedup: if 40% of work is non-parallelizable collaboration, overall speedup cannot exceed 2.5x no matter how fast individuals become.

— Brooks's Law means communication paths grow as n(n-1)/2; AI did not remove a single one of those edges.

— The Theory of Constraints dictates that accelerating non-bottleneck resources only lengthens the queue in front of the real bottleneck, sometimes making delivery cycles longer.

— Measure waiting time for first responses across cross-team handoffs—the longest waits reveal the true bottlenecks, which are rarely the ones assumed.

— Service-enabling an upstream team's capability into a self-service skill turns a blocking dependency into an asynchronous interface call.

— Amazon's 2002 API mandate forced every team to expose functionality through standard interfaces; today, writing a skill is cheap enough to start locally without a top-down order.

— Anthropic's Clontent agent handles 70% of content tasks and escalates judgment calls to humans, shifting the team from ticket processors to specification owners.

— Skill-ification fails when applied to accountability-heavy decisions, low-frequency collaborations, or anything that will lack a dedicated maintainer.

— Leaders should record waiting times, target 70% coverage for skills, make skill-building a recognized work output, and set a regular maintenance cadence.

— Individual contributors can start by building a minimal skill for their own most frequent blocking dependency—a script, prompt template, or fill-in document—and let adoption spread organically.

Conclusions

AI's biggest organizational leverage may not be accelerating existing tasks but increasing the parallelizable fraction p—turning cross-team negotiations into single-person judgments.

The marginal cost of service-enabling a capability has dropped so far that the bottleneck is no longer technical feasibility but the political will to expose interfaces.

A 70% draft with a cheap human merge step yields higher throughput than a 100% solution that requires queuing; this inverts the traditional quality-first instinct in enterprise workflows.

Unmaintained skills are more dangerous than no skills because errors accumulate silently once manual review is bypassed, creating a false sense of automation safety.

The PR model of collaboration makes expertise more visible as usage scales—counter to the intuition that self-service tools diminish the expert's role.

Organizational structure follows interface design: when local teams start exposing self-service skills, the actual collaboration graph begins to reshape before any formal reorg is announced.

Concepts & terms

Amdahl's Law

A formula stating that the overall speedup of a system is limited by the fraction of work that cannot be parallelized. If 40% of work is serial, maximum speedup is 1/(1-0.6) = 2.5x, regardless of how fast the parallel portion becomes.

Brooks's Law

Adding manpower to a late software project makes it later, because the number of communication paths grows quadratically as n(n-1)/2, increasing coordination overhead faster than it adds capacity.

Theory of Constraints

A management principle stating that a system's throughput is determined by its single bottleneck. Improving any non-bottleneck resource only increases work-in-progress inventory without improving overall output.

Skill (in AI collaboration context)

A packaged capability—such as a prompt template, script, or agent configuration—that lets a downstream team self-serve a task that previously required waiting for an upstream expert, effectively turning a human dependency into an API-like interface.

PR Model vs. Ticketing Model

In a ticketing model, upstream experts process requests sequentially. In a PR model, downstream teams use self-service skills to reach a 70% solution and submit changes for upstream review, shifting the expert's role from executor to specification owner.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗