Written for Sponsors, program leads and operators who have to explain why a case has stalled, or who want to refuse a pattern before it is funded.
What's in this framework
This 22-page document is a diagnostic framework for sponsors, program leads and operators who have to explain why a case has stalled, or who want to refuse a pattern before it is funded. The library has sometimes described it as a root-cause analysis, a survey of post-mortems, and a ranking of why AI projects fail. It is not a survey. It does not publish frequencies. It does not claim to have measured a sector.
What it is, instead, is eleven failure patterns, F1 through F11, written so they can be recognized in a proposal, a pilot, or a system that is already running. The patterns are a demo treated as an outcome; no signed baseline; data that cannot carry the work; a pilot that cannot travel; governance after the fact; every idea in the first wave; the wrong mechanism; built, then left; the people who do the work were never in the design; a number that cannot be questioned; and an incomplete cost of ownership.
It is written so a workshop can apply the patterns to a case of your own. A companion workbook holds the eleven as a register: for each, a sentence, a named owner, and a piece of evidence. A completed workbook is a diagnostic, not a finding that a program has failed, and not a certificate that the next one will succeed. The readiness assessment, the workflow playbook, the ROI guide and the governance framework in this library are companions. This diagnostic will not stand in for them.
How the patterns are meant to be used
Apply them before a case is funded, and to work that is already underway. A pattern that is present is a reason to pause, to redesign, or to refuse — not a reason to add a slide. The document does not say that every stalled case has all eleven, and it does not say that avoiding all eleven is a guarantee.
Several patterns point at companions in this library. No signed baseline, and a number that cannot be questioned, point at the ROI method. Data that cannot carry the work points at the data-strategy guide. Governance after the fact points at the governance framework. The wrong mechanism points at the agentic-workflows paper and the workflow playbook. Built, then left, points at handover: a system without an operator, a switch-off, and a trail is a pattern, not a later preference.
The people who do the work were never in the design is a pattern we see when a case is designed in a room the operators were not in. Every idea in the first wave is a pattern we see when a program funds a portfolio instead of a first case that can travel. A pilot that cannot travel is a pattern we see when the successful room cannot be repeated, because the data, the people, or the process were special to that room. The diagnostic asks you to write those facts down, honestly, before the next case is funded.
Eleven failure patterns
A diagnostic you can apply to a proposal or to a stalled case. Not a survey, and not a ranking by frequency. The full wording of each pattern is in the PDF and the workbook.
F1
A demo treated as an outcome. A polished demonstration is taken as evidence that the work is done, or that it will travel.
F2
No signed baseline. There is no agreed starting point, on a named process, against which a result can be examined.
F3
Data that cannot carry the work. The source, the definitions, the quality, or the permissions will not support the job.
F4
A pilot that cannot travel. What worked in one room depended on people, data or process that will not be there the next time.
F5
Governance after the fact. Inventory, counsel, an operator and a switch-off are promised for later, after “value” has been shown.
F6
Every idea in the first wave. A portfolio is funded instead of a first case that is bounded and can travel.
F7
The wrong mechanism. A goal-directed agent is chosen where a reliable workflow should stay, or the reverse.
F8
Built, then left. There is no operator, no handover, no switch-off, and no trail that can reconstruct a day.
F9
The people who do the work were never in the design. The case was designed in a room the operators were not in.
F10
A number that cannot be questioned. A claimed result has no method a second person could examine.
F11
An incomplete cost of ownership. The cost that was approved is not the cost of running, changing and eventually leaving the system.
A pattern that is present is a reason to pause, to redesign, or to refuse. Avoiding all eleven is not a guarantee that a case will succeed.
Get the guide and workbook
Enter your name, work email and company. The files appear on this page immediately. A backup download link is also emailed when email delivery is available.