Stefan Wolpers has released the AI Workflow Inventory, a critical new artifact for the A3 Delegation System that challenges teams to move beyond individual AI usage to a collective, documented understanding of their workflow. The core premise is that perceived knowledge of team AI habits creates false confidence; while individuals know their own prompts, the sum of team activity often remains invisible, undocumented, and unowned. This gap is what Wolpers defines as 'AI Debt,' a liability that accumulates when recurring tasks run on models without clear ownership or verification standards.

The Anatomy of an Inventory Row

The inventory is designed as a one-page canvas that forces specificity. Each row must include an identity strip, the task defined as a verb and object, the process it belongs to, and a provisional task class named by output and audience. Crucially, the schema separates the person running the task from the tool used, recognizing that staff turnover and model updates are frequent while the underlying task remains stable. For example, a row might list 'Transcribe photos of Retrospective sticky notes' run by a Scrum Master using GPT-5.6 Sol, with the output staying internal. This structure prevents the common pitfall of conflating the operator with the infrastructure.

Exposing the Unapproved Shortcut

A key design choice in the inventory is the requirement to list unapproved shortcuts and unattended jobs. Wolpers argues that if teams only record sanctioned uses, they miss the 'AI Debt' sitting in informal workflows, such as a nightly migration test generation set up by a developer who has since left the company. The inventory explicitly asks where output goes (internal, external, or customer-facing) and what data enters the model, distinguishing between the stakes of the result and the exposure of the input. This distinction is vital for security and compliance, as a task labeled 'internal' might still involve sensitive candidate names being uploaded to a third-party model.

Provisional Classes and the Grouping Test

The system uses five refinement rules to group individual tasks into 'provisional task classes.' These classes are defined by shared review standards, reviewers, and model tiers. If any of these three factors differ, the tasks must be split into separate classes. This approach acknowledges that teams often lack settled decisions on AI delegation; the inventory is designed to carry this uncertainty rather than force premature categorization. By starting with a 60-minute session to capture eight rows, teams can identify the workflows that individual recollection misses, creating a shared basis for the subsequent 'Assist, Automate, Avoid' decisions in the A3 framework.

Key Takeaways

  • The AI Workflow Inventory is a free canvas designed to document recurring AI tasks, preventing 'AI Debt' from accumulating in unowned workflows.
  • Rows must separate the operator from the tool to account for staff turnover and model changes without breaking the audit trail.
  • Teams are required to list unapproved shortcuts and unattended jobs, not just sanctioned uses, to get an accurate picture of AI exposure.
  • Task classes are provisional and grouped by shared review standards, reviewers, and model tiers, allowing for uncertainty in early stages.

The Bottom Line

If you can’t name every recurring AI task your team runs, you don’t have a workflow—you have a liability. Documenting these shortcuts is the only way to turn chaotic AI usage into a manageable, auditable engineering practice.