Documentation, Transparency and Human Oversight

Write down what it is for

Keep a short system card for each AI feature: what it is meant to do, what it must not be used for, the data and models behind it, evaluation results (including weak slices), known limitations and who owns it. Tell users they are interacting with AI and how reliable it is. For decisions that significantly affect people, such as loans, hiring, health or discipline, keep a human decision-maker who can review and override the system, and keep a record of why the decision was made.

A system card outline

One page is enough to start. Keep it next to the code and update it when the system changes.

Name / owner         Support assistant v3 / Customer Care team
Intended use         answer questions from the published help centre
Not for              legal, medical or financial advice; account changes
Model + data         <model name>; help-centre articles (updated weekly)
Evaluation           250 cases; 91% grounded-correct; Hindi slice 84% (gap noted)
Safeguards           PII masking, moderation, grounding check, rate limits
Human oversight      low-confidence answers go to an agent
Known limits         weaker on mixed Hindi-English; may miss recent changes

Disclose AI use clearly

A short label such as "AI assistant: answers may contain mistakes, check important details" sets correct expectations and reduces over-reliance.

Quick check: Who should make high-stakes decisions such as loan approval?

• The AI alone, always
• A human who can review and override the AI
• A random number generator
• Nobody