# The AI-Operable Product

> Formative Labs' framework and open Agent Skill for assessing whether a product can be safely understood and operated by an AI agent.

Canonical page: <https://labs.formative.ai/ai-operable-product>

## Summary

Most products are human-operable. Almost none are AI-operable.

A product that cannot express how it operates cannot be reliably reasoned about by an AI operator. AI-operability is a property of the product, not of the model or vendor stack attached to it.

## Definition

An AI-operable product exposes enough explicit operational context that a human or AI operator can:

1. Observe what is happening.
2. Reason about why it is happening.
3. Determine what can safely be done.
4. Verify the result.
5. Learn from the outcome.

## The seven operating primitives

1. **Outcome Intent**: What does success mean? The customer outcome the product should produce.
2. **Journey Context**: Where is the user? The business context needed to reconstruct the customer journey.
3. **Workflow Map**: How does work move? Progress across services, events, workflows, databases, and human handoffs.
4. **Recovery Path**: How does failure resolve? Designed retry, compensation, reconciliation, rollback, replay, failover, or escalation.
5. **Policy Boundary**: What must not happen? Authorization, approvals, regulatory controls, privacy constraints, and blast-radius limits.
6. **Safe Actions**: What can be done? Actions with explicit preconditions, authority, scope, verification, and reversibility.
7. **Feedback Loop**: How does the product learn? How production evidence changes code, tests, runbooks, policy, automation, or the journey itself.

## AI-OPERABLE DONE

The assessment reports evidence-backed PASS, PARTIAL, FAIL, or UNKNOWN results. It does not calculate an aggregate score.

- Outcome measurable
- Journey observable
- Failure explainable
- Recovery designed
- Actions governed
- Support enabled
- Learning feeds back

## What this framework does not conflate

- Observability is not operability.
- System health is not product success.
- A service topology is not a workflow map.
- Detecting failure is not recovery.
- An executable API is not a Safe Action.
- Telemetry alone is not a feedback loop.

## Resources

- [Open AI-Operable Product Agent Skill](https://github.com/khangtoh/ai-operable-product)
- [Digital Operations Diagnostic](https://labs.formative.ai/fde#fde-cta)
- [Formative Labs articles](https://labs.formative.ai/articles)