Bottleneck Shifts Upstream

Core Concept

When engineering velocity dramatically increases (through AI, automation, or process improvements), the bottleneck doesn’t disappear — it moves upstream to product definition, design decisions, and leadership alignment. Teams that invest only in engineering speed without addressing the upstream pipeline end up with fast developers who have nothing to build.

This is Amdahl’s Law applied to the software development lifecycle: optimizing the fastest stage yields diminishing returns when a slower stage dominates.

The Pattern

Before AI: Engineering is the bottleneck. Features take weeks/months to build. Product and design can stay ahead of dev capacity.
After AI adoption: Engineering throughput 2-5x. PRs merge same-day. Tech debt gets paid down aggressively.
The shift: Product can’t generate specs fast enough. Design can’t produce mockups fast enough. Leadership can’t make decisions fast enough. Engineers idle on “what should I build next?”
The new constraint: Decision-making velocity, not execution velocity.

Where I’ve Seen It

A peer engineering leader (Mar 2026): “We are currently bottlenecked by product and design… Leadership can’t decide things quickly enough which experiment they want to run.” His team is running out of tech debt to pay down — a problem he finds “a little scary.”
Drucker’s warning: “Nothing so useless as doing efficiently that which should not be done at all.” When you can build anything in a day, choosing the right thing becomes the only question that matters.

Implications

For VP Engineering roles

The relationship between VPE and CPO becomes the critical interface:

The VPE↔CPO relationship will matter more than the VPE’s relationship with any individual engineer
The conduit function (exec strategy → engineering execution) must flow in both directions: engineering velocity creates pull on the product pipeline
If the team adopts AI successfully, the pressure moves to the CPO’s ability to feed the backlog with well-scoped, validated work

For small companies

Don’t invest in AI engineering velocity without also investing in product/design velocity
An “orchestration mesh” vision tries to solve this by making product people self-serve — but that requires its own infrastructure
Feature flags and experiment frameworks become essential — they let you ship fast and learn fast, which keeps the product pipeline moving

Amdahls-Law — the theoretical foundation (sequential bottlenecks limit maximum speedup)
Input-Constrained-vs-Output-Constrained — explains why the bottleneck shifts: automating input-constrained work reveals the output-constrained bottleneck
Effectiveness-Over-Efficiency — when execution is cheap, choosing the right problem is everything
Testing-Infrastructure-As-AI-Enabler — the prerequisite for engineering velocity gains
Right-Problem-Leverage — 100x comes from problem selection, not speed