The average large enterprise currently operates between stage two and stage three of AI maturity, while growth-stage companies hover between stages three and four, according to a new framework published by Quotient that maps how engineering organizations advance their use of artificial intelligence. The framework, presented by Lizzie Matusov, co-founder and CEO of Quotient, identifies five distinct stages of AI adoption and explains why individual developer productivity gains aren't translating into organizational delivery improvements. The model draws on synthesis of 15 research papers, interviews with productivity framework designers, and field discovery across more than 100 organizations prioritizing AI in 2026.
Global AI spending reached $2.5 trillion in 2026, reflecting 47% year-over-year growth, according to the presentation. Seventy-two percent of enterprises now run at least one AI workload in production. Uber depleted its entire 2026 AI budget within just four months, one-third of the expected timeframe, yet the company's CTO reported being unable to establish a stable relationship between expenditure and productivity output. Microsoft announced plans to discontinue Claude Code by month's end due to budget overruns, similarly unable to quantify return on investment. The 2025 DORA report, examining roughly 5,000 technology professionals, found individual effectiveness increased while software delivery throughput showed no meaningful change, attributing the gap to unresolved bottlenecks in systems surrounding individual AI use.
The five-stage model defines maturity through six characteristics: enablement, policy and governance, validation and testing, embedding and workflows, workflow automation, and data context and access. Stage one represents ad hoc experimentation with informal self-review and manually triggered workflows. Stage two, assisted development, features basic policies and emerging best practices but creates bifurcation between early adopters and skeptics. Stage three, standardized workflows, introduces formal training, clear organizational policies, and automated checks that validate AI-generated code as systems adapt to increased volume. Stage four, supervised automation, involves engineers configuring agents to execute bounded tasks with high oversight, such as handling low-priority support tickets while humans review and deploy. Stage five, end-to-end autonomy, describes AI orchestrating multi-system workflows with agents possessing ownership equivalent to senior engineers, requiring continuous policy evolution and production-linked feedback loops.
The framework identifies bottlenecks as the primary obstacle preventing individual AI effectiveness from becoming organizational delivery gains. Matusov's presentation applies the theory of constraints to software development, arguing that every system is limited by a single rate-limiting step and that increasing throughput elsewhere won't improve output without addressing that constraint. When bottlenecks persist, work-in-progress inventory explodes, lead times extend, and quality problems compound. Organizations advancing from stage two to stage three experience particularly visible bottlenecks because this transition shifts focus from individual tasks to team systems and development harnesses. A Concordia University study examining 30 development tasks found code reviews consumed 60% of token usage due to back-and-forth between developers and agents, with tokens disproportionately spent on input rather than output, illustrating what researchers termed a communication tax. The presentation argues infrastructure built in stages one through three becomes essential for successfully deploying agents in stage four, as handing control to agents without proper validation systems and quality harnesses creates unacceptable risk.
The framework recommends organizations target stages three to four as a 2026 goal rather than rushing to stage five, and advocates measuring outcomes rather than activity metrics like token consumption. Token usage provides weak signals of productivity because each token carries literal dollar cost, incentivizing spending rather than results, the presentation warns. One senior engineer at a major technology company described deliberately inflating token metrics by asking AI to retrieve documentation ten times slower than manual lookup, prototyping abandoned features solely to burn tokens, and defaulting to agents even when manual work would be faster. Amazon shut down its token leaderboard over the weekend, with a senior vice president writing in an internal memo that teams should use AI to solve customer and business problems rather than for its own sake. The presentation recommends instead measuring speed, quality, and ease, or using the SPACE framework covering satisfaction, performance, activity, communication, and efficiency to identify friction points. When usage grows but delivery outcomes stagnate, that gap signals new bottlenecks requiring diagnosis and resolution as a continuous process throughout maturity advancement. The most effective engineering organizations both thoughtfully improve AI usage across the software development life cycle and systematically resolve the bottlenecks limiting their outcomes. Organizations that race toward full autonomy without building the necessary testing harnesses, validation systems, and quality controls risk losing the customer trust that depends on software stability—a trade-off that becomes irreversible once humans can no longer manually support the load agents have assumed.