The Operating Model Was the Upgrade, Not the AI

2026-07-01 · Source: Towards AI - Medium · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Intermediate, extended

Summary

A 2025 randomized controlled trial by METR found experienced open-source developers were approximately 19% slower when using early-2025 AI tooling, despite expecting a 24% speedup. In contrast, fortiss's platform engineering team, using the same class of AI model, achieved dramatically faster output and cadence building their Punctilious Platform, a production-ready, multi-tenant system. This discrepancy highlights that the true advantage from AI lies not in the model itself, but in the "operating model"—the surrounding system of methods, structure, and governance. fortiss's approach integrates seven layers, including agile delivery, structured work items, encoded skills, and durable memory, to eliminate repeated work and provide auditability for regulated sectors. While fortiss reported a 22–28× output per engineer-hour (a ~99% imputed proxy) and 3.5%–17.9% rework, these are from an n=1 observational study, not a controlled trial.

Key takeaway

For AI Architects and Software Engineers evaluating AI integration, recognize that raw model capacity is less critical than the surrounding operating model. You should prioritize building robust systems with structured work items, durable memory, and strong CI delegation to maximize AI effectiveness and ensure auditability. Start by implementing one layer, such as structured work items or curated memory, for two weeks to measure its impact on reducing repeated work. This approach shifts your focus from renting model benefits to owning a sustainable, auditable AI-driven development process.

Key insights

The true advantage of AI in software development stems from the operating model built around it, not the AI model itself.

Principles

• AI's impact depends on the operating model, not just the model.
• Unscaffolded AI incurs repeated work, silently costing time.
• An owned AI advantage stems from the operating model, not the model.

Method

The MIA (Model, Instantiate, Apply) method, applied to software development with AI, involves modeling work, instantiating with reusable assets, and applying a governed loop.

In practice

• Implement structured work items for clear AI task boundaries.
• Curate durable memory to prevent AI re-learning past lessons.
• Use "eval data" (tests) to verify AI output against domain standards.

Topics

• AI Operating Models
• Platform Engineering
• Developer Productivity
• Software Auditability
• Context Management
• MIA Method

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Software Engineer, AI Architect

Related on AIssential

• From AI projects to an operational capability — AI becomes an operational capability when it impacts P&L, business KPIs, and is supported by modern architecture and governance.
• Why Experience Matters Most in the AI Era — AI accelerates code generation but amplifies the need for human judgment and experience to prevent rapid accumulation of technical debt.
• Measuring AI Ability to Complete Long Software Tasks — AI's software task completion ability doubles every 7 months, potentially reaching month-long tasks by 2029.
• 📈 Data to start your week: Inside the AI boom – jobs, jargon & jittery uptime — AI is poised to automate junior roles, but widespread adoption faces cost and compute constraints.
• AI & The Harness Theory — An AI operating system can transform individual AI utility into replicable team capability.
• The AI Stack Is Moving From Model Access to Model Operations — The AI stack is evolving from single-model access to multi-model operations, requiring a dedicated management layer for diverse provider complexities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.