The Local Optima Trap

In the AI Story

Goldratt considered local optimization the most destructive error in management. Not because it is irrational — each department is behaving rationally by its own metrics — but because the aggregation of locally rational behavior produces globally irrational outcomes. The machining department runs at full speed because its metric is utilization. The assembly department drowns in partially machined parts it cannot process. The paint department starves because assembly cannot pass enough completed units forward. Each department is 'efficient.' The factory is a mess.

The AI transition has created conditions for local optimization at a scale and speed Goldratt never witnessed but would have recognized immediately. A development team adopts Claude Code. Individual engineers report dramatic productivity improvements. Each engineer generates more code, builds more features, completes more tasks. The team's velocity, measured in story points per sprint, doubles or triples. Management celebrates. Dashboards glow green. But downstream processes — product evaluation, quality assurance, user testing, market validation — have not been augmented by the same tools. The rate at which generated features can be evaluated for market fit has not changed. The rate at which deployed features can be tested by real users has not changed. The rate at which the market can absorb innovation has not changed.

The engineering team is running at full speed. Everything downstream is drowning. Product managers face review queues that grow faster than they can process. QA, still operating at human speed, faces testing backlogs that compound weekly. User researchers, whose work requires the irreducibly slow process of watching humans interact with software, cannot keep pace. Customer support, confronted with a product that changes faster than its documentation can be updated, struggles to help users navigate features that were deployed before anyone determined whether they were usable.

Each upstream improvement — each increase in generation speed, each new AI capability the engineering team adopts — makes the downstream situation worse. The engineering team's velocity is a local optimum. The system's throughput — the rate at which value reaches the user — is constrained by the downstream bottleneck that no one is watching because all organizational attention is focused on the engineering team's spectacular metrics. This is the Local Optima Trap in its purest form, applied to the defining technology transition of the era.

The pathology compounds. As unevaluated features accumulate in the product, coherence degrades. Features interact in ways no one anticipated because no one had time to anticipate them. User experience becomes inconsistent — some features polished, others rough, quality determined not by deliberate standard but by the random variable of whether evaluation time was available before deployment. The product becomes, in a word designers use with a specific shudder, 'bloated' — not because any individual feature is wrong but because the aggregate exceeds the product's capacity to maintain coherence.

Origin

Goldratt encountered the Local Optima Trap in virtually every manufacturing consulting engagement in the 1980s and 1990s. His diagnosis was that cost accounting and departmental budgeting together create structural incentives for local optimization that no amount of exhortation to 'think systemically' could overcome without a corresponding change in measurement frameworks. His advocacy for Throughput Accounting followed directly from this diagnosis: if you want systemic behavior, measure the system, not the parts.

Key Ideas

Local rationality, global irrationality. Each department optimizing its own metric produces aggregate dysfunction that no individual department can see from within.

Measurement systems drive behavior. Organizations that measure utilization get utilization, even when utilization degrades system throughput.

The AI transition amplifies the trap. Tools that accelerate one part of the system without accelerating downstream processes create the exact conditions Goldratt diagnosed, now at software speed.

Engineering velocity without downstream capacity is waste. Features generated faster than they can be evaluated, validated, and absorbed into a coherent product are inventory, not throughput.

The cure is subordination. Non-constraint resources must operate at constraint-pace, even when this means 'underutilization' by local metrics.