Organizations Rush Into AI Spending While Missing Critical Bottlenecks

Enterprise AI spending is accelerating at an unprecedented pace. Global AI investment grew 47% year-over-year in 2026, with projections reaching $2.5 trillion and 72% of enterprises already running AI workloads in production. Yet there's a growing paradox: individual developer productivity has improved, but organizational throughput hasn't meaningfully changed—and spending is often detached from measurable returns.

Companies like Uber exhausted their entire 2026 AI budget in just four months, while Microsoft recently halted Claude Code integration due to budget overruns. In both cases, executives acknowledged an inability to clearly connect expenditure to productivity gains. Individual usage, it seems, is not automatically translating to organizational outcomes.

The Theory of Constraints#

The disconnect becomes clearer through the lens of systems thinking. Every production system—whether a factory or a software organization—is ultimately constrained by a single bottleneck, the slowest step in the workflow. Optimizing everything else without addressing that constraint doesn't accelerate throughput; instead, it creates backlogs, longer cycle times, and compounding quality issues.

The software development life cycle works the same way. If code review is the bottleneck, flooding the system with more AI-generated code won't increase delivery speed. It will only pile work behind the constraint. Bottlenecks vary widely: they might be testing, deployment, architecture reviews, or something unique to each organization and team.

A 2025 DORA study of 5,000 tech professionals confirmed this pattern. While individual effectiveness rose at the top, software delivery throughput didn't budge—because the surrounding systems had unaddressed bottlenecks that prevented end-to-end gains.

The Five-Stage AI Maturity Framework#

To move beyond this stagnation, Lizzie Matusov, co-founder and CEO of Quotient, developed a maturity framework grounded in first-principles research. The framework synthesized 15 peer-reviewed studies, interviews with researchers behind frameworks like DORA and SPACE, and field discovery across 100+ organizations prioritizing AI.

The model describes five progression stages, each defined across six dimensions: enablement (learning and skill growth), policy and governance, validation and testing, embedding in workflows, workflow automation, and data context and access.

Stage 1: Ad Hoc Adoption Usage is driven by individual experimentation with no formal guidance. Developers paste context into chat windows, self-review outputs, and manually trigger tools. AI exists outside normal workflows, with minimal organizational guardrails.