Check where your AI system looks fine but actually breaks.
This is a small tool built from a bigger question:
What actually fails when an AI system fails?
Not the model.
Not the interface.
The system around it.
ControlGap generates a pre-mortem for AI products.
Before launch, it tries to show:
- How this system will fail
- Where users will trust it too early
- Where nobody is really responsible
Most AI products do not fail in obvious ways.
They fail quietly.
The output looks clean.
The interface feels reliable.
The system responds fast enough that nobody questions it.
And then decisions are made on top of that.
This project comes from the same thinking behind my book:
"The Interface Is Not the System: Why Control, Trust, and Responsibility Break at Scale in AI Systems"
The core idea is simple:
What users see is not where control actually lives.
And when that gap grows, failure becomes invisible.
You give it a product.
Either by filling in the fields or by providing a link.
It returns a structured analysis:
Not generic risks, but specific situations where the system breaks in real use.
Where trust is created, why it feels justified, and when it becomes wrong.
Hidden risks, long term effects, and responsibility gaps that do not show up in normal product thinking.
Not just UI fixes, but points where human control needs to exist.
This is not a prompt experiment.
It is an attempt to model:
- how people stop questioning systems
- how AI output becomes “good enough” to act on
- how responsibility disappears in automated workflows
If a system fails but nobody notices, that is not a small bug.
That is a design decision.
- AI hiring tools
- AI writing assistants
- internal knowledge systems
- automation agents
- enterprise workflows
Anywhere AI output is used as a decision input.
- Extract product intent and behavior
- Infer how the AI is positioned in the workflow
- Simulate failure modes
- Map trust and responsibility gaps
- Generate a system-level report
The analysis is based on available inputs.
- URL extraction depends on public content
- some scenarios are inferred, not verified
- it does not observe real user behavior
This is a thinking tool, not a monitoring system.
AI UX is often reduced to usability.
But the harder questions are different:
- Who is in control when the system is right most of the time
- What happens when it is wrong and nobody knows
- Where does accountability actually sit
These questions usually come too late.
I have been designing SaaS and AI systems for over 10 years.
Most of the real problems were never about usability.
They were about control, trust, and responsibility.
This project is part of that work.
I also wrote:
"The Interface Is Not the System: Why Control, Trust, and Responsibility Break at Scale in AI Systems"
Portfolio: nellypourmehr.com
LinkedIn: https://www.linkedin.com/in/nellypourmehr/
If you are building an AI product and something feels unclear about control or trust, feel free to reach out.