User's avatar

Empathetic AI Lab

10-Minute AI System Readiness Check

Expose where your AI system’s responsibility hasn’t been fully designed — and the risks that emerge.

Judy Ossello (AI Mechanic)'s avatar
Judy Ossello (AI Mechanic)
Mar 18, 2026
∙ Paid

Your system already has a job. This shows you what it actually is.

I built my first empathetic AI agent from scratch in two days.

But, I spent the next few weeks trying to fix it.

Building isn’t the hard part. Creating an endless series of guardrails to fix a design problem is the hard part.

When an AI system’s job isn’t clearly defined, the system fills in the gaps.

Scope expands.
Authority drifts.
Behavior starts to change in ways that technically “work” but feel increasingly wrong.

Designing around responsibility doesn’t have to be the hard part.

In fact, I want to make responsibility design so straightforward and valuable that you’ll never want to build without it.

Each question targets a different part of how an AI system is defined.

Together, they give you a first-pass assessment of whether responsibility is actually defined—or left to interpretation.

What You’ll Notice When You Run It

As you answer each question, you’ll likely see:

  • where definitions are missing

  • where behavior is assumed instead of specified

  • where responsibility could shift under pressure or execution

Each question maps to a different part of the AI System Reliability Blueprint.

It shows you where to look next and gives you insight on where your system design needs to go deeper.

Responsibility Readiness Check

1. What job is your AI system responsible for?

Write one sentence describing the system’s job.

Then ask yourself:

  • What is the system not allowed to do?

  • When must the system refuse or escalate?

If the boundary isn’t clear, the system will infer one.


2. What behavior counts as success?

Think beyond task completion.

Ask:

  • When should the system assist?

  • When should it advise?

  • When should it stop or defer?

Many systems define tasks but never define behavioral expectations.


If you’re already noticing gaps… that’s the signal.

Most builders don’t realize their system’s job is undefined until they try to answer these questions.

That’s where behavior starts to drift.


3. Where does responsibility move in your system?

List every component your AI interacts with.

Examples:

  • humans

  • agents

  • tools

  • APIs

  • workflows

Then ask:

What responsibility moves across those boundaries?

If the answer is unclear, responsibility can quietly shift inside the system.


4. What happens when the system doesn’t resolve the situation?

Imagine:

  • repeated attempts

  • unresolved requests

  • user distress or urgency

  • lack of progress across turns

Ask:

  • Does the system change behavior—or keep trying?

  • When is it required to stop, defer, or escalate?

  • What is it explicitly not allowed to become under pressure?

👉 If this isn’t defined, pressure will reshape the system’s role.


5. What happens when output becomes action?

When the system produces an output that could trigger a real-world change:

  • Does it require a clear human decision before action?

  • Or can outputs pass directly into tools, workflows, or other agents?

  • Is there a pause, confirmation, or explicit handoff of ownership?

👉 If this isn’t enforced, responsibility can shift at the moment of execution.


6. Where could your AI accidentally become responsible for something it shouldn’t?

Describe one realistic scenario.

Examples:

  • the system gives advice outside its role

  • a tool is triggered without confirmation

  • an agent reinterprets instructions

👉 If this scenario is easy to imagine, you’ve identified a responsibility risk.


How Was It?

If some of those questions were hard to answer—that’s the signal.

Most builders assume they can fix behavior later with better prompts or more guardrails.

You can’t fix a poorly defined job with more code.

That’s how systems drift:
→ patching behavior
→ expanding scope
→ accumulating hidden responsibility

Most AI systems aren’t designed around responsibility.

So the system defines it on its own.

That’s where unpredictable behavior starts.

Was This Useful?

If this exercise surfaced questions about your system, this lab is built to help you answer them.

If you discovered something interesting while answering these, feel free to reply or comment — I’m collecting examples as the lab continues to study how AI systems behave in practice.


This is for you if:

  • your system works, but feels unpredictable

  • you’ve added guardrails but behavior still drifts

  • you can’t clearly explain what your system is responsible for

You Don’t Need More Guardrails. You Need a Defined Job.

If this check surfaced gaps, your system is already operating with undefined responsibility.

That’s where drift starts.

In the Responsibility Design Lab, we take one system and fix this directly:

→ define its job in one sentence
→ identify where it breaks under pressure
→ set boundaries it can actually hold

Small group. Real systems. No theory.


Set up a free Calendly with a 20-minute discovery call slot.

Design Lab Discovery Call


Related research from the lab

Your Prompt Was A Note, Not A Job

Your Prompt Was A Note, Not A Job

Judy Ossello (AI Mechanic)
·
Apr 16
Read full story
Why Every AI Agent Needs a Charter Before You Build It

Why Every AI Agent Needs a Charter Before You Build It

Judy Ossello
·
Feb 19
Read full story

Paid subscribers can work through it on their own with a virtual whiteboard prompt designed to help you sketch out your AI design ideas with these concepts in mind.

User's avatar

Continue reading this post for free, courtesy of Judy Ossello (AI Mechanic).

Or purchase a paid subscription.
© 2026 Judy Ossello · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for great culture