Separate legal duty from model-provider discretion.

Video caption

Separate legal duty from model-provider discretion.

A fallback model cannot cure a legal ban.

Task class, jurisdiction, provider and action right.

My rule: One model response never settles the whole question.

#EricFieldNotes

Full written post / accessibility read

If an agent stops, leaders need to know what changed. The EU AI Act has specified prohibitions and obligations. A provider's usage policy sets its service conditions. Your application applies local roles and business rules. Conflating them turns a tractable incident into a vague 'AI compliance' problem.

For an authorized use that a provider declines, the team may have an approved alternate or human route. For a prohibited use, moving to another model does not make it allowed. For an application permission failure, changing providers does not grant the employee access.

For each consequential route, have legal and product owners record permitted use, data boundary, provider agreement, identity/role gate and escalation. Version the matrix. A later provider or law change should trigger targeted review, not an unrecorded prompt edit.

Stop the effectful tool until the applicable legal use, provider route and application permission are resolved. Do this because the business owns the action even when an upstream model is willing to generate it.

#EricFieldNotes

Four-beat scene transcript

1. Separate legal duty from model-provider discretion.

If an agent stops, leaders need to know what changed. The EU AI Act has specified prohibitions and obligations. A provider's usage policy sets its service conditions. Your application applies local roles and business rules. Conflating them turns a tractable incident into a vague 'AI compliance' problem.

Visual: The same failure symptom can require a different executive decision.

2. The recovery routes are not interchangeable.

For an authorized use that a provider declines, the team may have an approved alternate or human route. For a prohibited use, moving to another model does not make it allowed. For an application permission failure, changing providers does not grant the employee access.

Visual: A fallback model cannot cure a legal ban.

3. Build a decision matrix before launch.

For each consequential route, have legal and product owners record permitted use, data boundary, provider agreement, identity/role gate and escalation. Version the matrix. A later provider or law change should trigger targeted review, not an unrecorded prompt edit.

Visual: Task class, jurisdiction, provider and action right.

4. Require an explicit three-layer pass.

Stop the effectful tool until the applicable legal use, provider route and application permission are resolved. Do this because the business owns the action even when an upstream model is willing to generate it.

Visual: One model response never settles the whole question.

Research and claim limits

• OpenAI Usage Policies (S133)
• European Commission: AI Act enforcement (S135)
• Regulation (EU) 2024/1689, official text (S136)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.