Inputs, Outputs & Effects
Every AI Product must declare the contracts through which it receives context, produces intelligence, and, where applicable, causes digital or physical effects.
For informational AI this may look like a conventional input/output contract. For agentic or physical AI, an output-only model is insufficient: the Product may issue decisions, plans, tool calls, control commands, or physical actions.
Input Declaration
An AI Product should declare, as applicable:
- input types and modalities;
- machine-readable schemas/contracts;
- preprocessing or contextualization requirements;
- constraints such as size, language, timing, or environment;
- sensitivity and governance classifications;
- required Data Product, knowledge, sensor, or upstream AI Product inputs.
Output Declaration
An AI Product should declare, as applicable:
- scores;
- classifications;
- predictions;
- generated text/media/code;
- recommendations;
- explanations/evidence;
- decisions;
- plans;
- structured intelligence objects.
Schemas, uncertainty/confidence semantics, error handling, and compatibility expectations should be machine-readable where feasible.
Intelligence Effect Contract
Where the Product can cause action beyond informational output, AIPROD should describe the Intelligence Effect (IE) boundary.
Examples include:
- Digital Action — invoke an API, modify workflow state, create/update a record;
- Physical Action — issue a control instruction or directly actuate a physical system;
- Decision — commit or authorize a governed outcome;
- Plan — produce a sequence of actions for another actor/system to execute.
Effect declarations should link to Governance & Policy, Agency / Autonomy semantics, and relevant AIPDS controls.
Contract Characteristics
AIPROD should describe:
- deterministic vs probabilistic behaviour where meaningful;
- interface/effect stability across versions;
- uncertainty/confidence semantics where applicable;
- error/fallback behaviour;
- authorization and policy constraints on actions;
- human approval boundaries where applicable.
Example
Fraud Detection AI Product
- Input: governed transaction feature contract.
- Output: fraud score + classification + evidence.
- IE: Recommendation / Alert.
- AR: Recommend.
Autonomous Navigation AI Product
- Input: sensor/world-state context.
- Output: route/trajectory plan.
- IE: Physical Action where directly controlling actuators.
- AR: Autonomous Action within delegated safety boundaries.
Summary
- Inputs define what context the Product consumes.
- Outputs define informational intelligence results.
- Effects define decisions or actions the Product can cause.
Principle: For AI, the contract boundary may describe not only what is returned, but what the intelligence is authorized to cause.