Terms coined and defined here. Each entry links to the work where the concept is developed. Cite the term, link the entry.
The category error behind enterprise AI failure. Organisations treat agents as software projects: build, test, deploy, close. Agents are probabilistic systems whose behaviour changes in production, so the project closes exactly when the real work begins. The trap is structural rather than a skills gap, which is why bigger budgets and better models have not moved the failure rate.
Named in The Pilot Trap (April 2026).
The operating discipline that takes over where project management ends: the ongoing practice of monitoring, evaluating, governing, and adapting AI systems through their production life. Project management assumes a defined end state and deterministic behaviour. AI systems have neither. Deployment is not the finish line; it is the first day the discipline is needed.
Developed in the essay System Stewardship and in The Pilot Trap.
The continuous, variable, compounding cost of running AI systems in production. Traditional software carries a build cost and a decaying maintenance cost. Agentic systems carry this third cost, incurred every month the system runs, and it is the cost most business cases omit. Programs that ignore it discover it in the year-two budget review.
Defined in the executive briefing The Operating Tax and Newsletter Issue 8.
The month at which an AI agent’s cost per outcome exceeds its value per outcome. Value per outcome is typically fixed at deployment while cost per outcome compounds, so the two curves cross quietly. Programs measured on aggregate savings do not see the crossing until the economics have already turned.
Developed in the article The Inversion Point.
The condition in which each task an AI agent completes costs more than the human equivalent, once the full operating cost is counted. Aggregate dashboards can show savings while unit economics run negative, because volume hides the per-task loss. The corrective metric is Cost Per Successful Outcome.
Named in Newsletter Issue 4 (April 2026) and developed in The Pilot Trap.
The governing metric of production AI economics: the full operating cost of an agent divided by the outcomes it completed successfully, not the tasks it attempted. Counting attempts flatters the system; counting successful outcomes prices failure, retries, escalation, and human correction into the number the CFO actually needs.
Developed in The Operating Tax.
The predictable gap between a flawless demonstration and production failure. Demos run on curated inputs, stable context, and a forgiving audience. Production runs on adversarial inputs, drifting context, and an unforgiving cost line. The more perfect the pilot, the less the organisation has learned about the conditions that will break it.
Named in Newsletter Issue 2 (March 2026).
The continuous performance practice that replaces episodic retraining. Retraining treats an underperforming agent as a broken build to be reshipped. Coaching treats it as a performing system whose behaviour is observed, evaluated against defined outcomes, and corrected in small, continuous adjustments, the way an operating leader manages a team rather than the way a project manager closes tickets.
Named in Newsletter Issue 5 (April 2026).
The surfaces across which system stewardship operates: Business Alignment, Integration Architecture, Governance and Compliance, Observability and Evaluation, and Orchestration and Operations. An enterprise that cannot name an owner for each surface does not have an AI operating model; it has deployments.
Defined in the AI Role Operating Framework (AI ROF™) whitepaper.
The roles that carry the five control surfaces: the AI Delivery Manager, the AI Architect, the AI Product Manager, and the AgentOps Lead. These are accountabilities, not headcount: in smaller organisations one person may hold two, but an unheld accountability is where production AI fails first.
Defined in the AI Role Operating Framework (AI ROF™) whitepaper and The AI Delivery Manager Blueprint.
An operating architecture in which intelligence, human and artificial, is the organising principle of the enterprise rather than a capability bolted onto existing functions. Co-authored work with Vinod Bijlani, Mark Cameron and Sharma Madiraju.
Developed in the ICE paper series, published separately.
