The Framework
The Causal Stack
A diagnostic for the six patterns of AI failure, and how to design out of them.
Premise
AI optimises whatever signal you give it. Most companies get the signal wrong.
The result is systems that work. They hit metrics. They ship. They get used. And then the impact stalls.
The Causal Stack is a framework for identifying which of six failure patterns is currently dominating an AI system, and what to do about it. It comes out of building and deploying AI inside organisations like Shell, KLM, ABN AMRO, and RTL, and watching where the systems stalled.
The framework grades every AI system against five operational dimensions. When one or more collapse, the system settles into one of six patterns. The pattern is the diagnosis. The dimension that broke is the operational fix.
The Five Dimensions
Every AI system is graded against five operational dimensions. Each one has to hold for the system to compound.
- 01
The Problem
Is there a real, named, valuable business problem this AI is solving? Or was the tool chosen first and the problem retrofitted?
- 02
Tech-Fit
Is the architecture and model class right for this problem? Right size, right deployment pattern, right level of capability used.
- 03
The Signal
Does the measured signal actually move the strategic outcome? Or is it tracking activity (adoption, throughput) that doesn't chain to business value?
- 04
The Loop
Does the system itself learn from observation? Or are humans iterating manually on prompts and configs while the system stays static?
- 05
The Moat
Is the system accumulating proprietary signal a competitor cannot replicate? Or is your edge mostly off-the-shelf capability that anyone could rebuild?
The Six Patterns
When one or more dimensions collapse, the AI system settles into one of six patterns. The pattern names what is wrong. The redirect names where to look.
- 01 Pattern
The Facade
Looks like AI. Isn’t.
Diagnosis
Your AI looks like AI. Yet, it's running scripted work behind a generative interface. The capability you're paying for isn't being used.Where to look
Consider: what would change if you let the model do the work it's actually good at? - 02 Pattern
The Ghost
It works. Nothing moves.
Diagnosis
Your AI works. It hits its metrics. Yet, you can't point to where it's actually moving the business.Where to look
Consider: which decision in the company would be different tomorrow if this system disappeared? - 03 Pattern
The Misfire
Aimed well. At the wrong thing.
Diagnosis
Your AI is executing well, but most probably it's optimising the wrong thing. The metric moves; the strategic outcome doesn't.Where to look
Consider: if you traced your dashboard back through the chain to the business outcome, where does the link break? - 04 Pattern
The Wrapper
Your edge is rentable.
Diagnosis
Your AI works, but your edge is mostly prompts and integration. A small team could rebuild what makes you different in a few weeks.Where to look
Consider: what would your AI know about your customers that no foundation model can replicate? - 05 Pattern
The Mandate
Adopted widely. Anchored to nothing.
Diagnosis
Your AI is deployed widely. The team measures who uses it. Nobody can pin down what it's actually changing for the business, apart from cost/tool.Where to look
Consider: what specific job is the tool doing, and who would notice if it stopped? - 06 Pattern
The Compound
On the way. Almost none are.
Diagnosis
Your AI is close to compounding. Almost no system is fully there. You have the foundation. The radar shows the gap.Where to look
Consider: what would have to be true for this to keep compounding for the next five years?
Run it on your AI system
Seven questions. Ninety seconds. A reading on which pattern is dominating yours.
The Author
The Causal Stack is by Niya Stoimenova, a Principal AI Architect and keynote speaker. Her work starts after deployment, when leaders realise the system is not improving what actually drives the business.
She has built and deployed AI inside organisations like Shell, KLM, ABN AMRO, and RTL, and now works with teams to design systems that improve in the right direction. PhD in AI Reasoning, TU Delft.
Run this live