This field guide asset explores the Industrialised Filtering pattern:

when optimisation systems determine who receives attention, consideration, or opportunity at a scale beyond direct human review. 

The Scratch Test

Reality Always Pushes Back

Every AI deployment begins as a claim.

That claim might be about reducing cost, improving productivity, increasing consistency, supporting better decisions or delivering better outcomes for citizens and customers.

For a time, that claim may appear entirely valid.

Then reality changes.

People adapt. Workflows evolve. Regulations change. New technologies emerge. Budgets tighten. Organisational priorities shift. Success itself changes behaviour.

The question is no longer whether the AI system still works.

The question is whether the original organisational claim still holds.

The Scratch Test is the disciplined practice of repeatedly asking that question.

Unlike a one-time governance review or deployment approval, the Scratch Test recognises that AI systems operate inside dynamic human organisations. Every deployment is continuously exposed to changing conditions that may strengthen, weaken or completely invalidate the assumptions on which the original decision was based.

Governable organisations do not assume that yesterday's approval guarantees tomorrow's success.

They deliberately revisit their claims.

Why It Matters

Many organisations monitor technical performance.

They measure uptime. Accuracy. Latency. Token consumption. And, user adoption.

These measures are important, but they tell only part of the story.

The Scratch Test asks a different question.

Does the original organisational claim still hold?

That simple question shifts attention away from governing models in isolation and towards governing outcomes over time.

It recognises that organisational success depends not only on the quality of the technology, but on the continued alignment between people, workflows, governance and AI.

The Governability Loop

Use Case as a Claim → Deployment → Gather Evidence → The Scratch Test → Review the Claim → Does the Claim Still Hold? 

The Scratch Test is not a one-time event.

It is a governance habit.

Every time evidence accumulates, organisations have another opportunity to revisit the claim they originally made.

Common Triggers

Reality rarely changes in only one way.

A Scratch Test may be triggered by:

• Organisational growth or increased scale

• New legislation or policy

• Economic pressure or changing token costs

• New AI models or platform capabilities

• Changes in user behaviour

• Unexpected workarounds

• New data sources

• Changing customer expectations

• Adversarial behaviour or misuse

None of these automatically indicate failure.

They indicate that the environment has changed.

Ask Yourself

Before assuming your AI deployment remains successful, ask:

• What was the original claim?

• What assumptions supported that claim?

• What has changed since deployment?

• Does the evidence still support the original claim?

• If we were making this decision today, would we make the same one?

• If not, what should change first: the people, the workflow, the governance, or the AI?

Continue Your Learning

Creator's Dispatch

Reality Always Pushes Back

A short reflection on why successful deployment is only the beginning of effective governance. 

[Youtube Video]

Field Guide Exercise

Stress the Claim

Download the worksheet and apply the Scratch Test to one of your own AI deployments or digital workflows. 

[Download PDF/Link to Interactive Version]

Worked Example

Policy as a Digital Resource

Why deliberately standardising on GPT-4.1 within Azure Foundry provided a more governable operating environment than continually chasing the newest frontier model.

The lesson is not "use an older model."

The lesson is:  Control the variables you can before reacting to the variables you cannot.

Key Takeaway

"Governable organisations do not assume their original decisions remain correct. They continually revisit the claims on which those decisions were based." 

Related Patterns

• Outcome Blindness → Are you measuring outputs or outcomes?

• The Default Human → Who was this system designed for?

• < Invisible Grounds → What assumptions remain hidden?

• Adaptive Organisation (Chapter 12) → How should organisations respond when claims begin to weaken?

This page is part of the Fourth Horizon Digital Field Guide accompanying Governable: The Last Human in the Room.

First published as part of the AI Pathfinder Field Guide.

Last reviewed: June 2026.