Guardrails are the enforcement layer of a Zoho Zia Agent. Unlike instructions, which guide how the agent approaches tasks, guardrails set firm boundaries the agent cannot cross regardless of context. When a guardrail and an instruction conflict, the guardrail always wins.

Built-in Guardrails

Two default guardrails are available and enabled on every agent from the outset. You can toggle either on or off during agent creation or from the agent details page at any time.

• Fairness and Bias. This guardrail monitors agent responses for potential biases and adjusts content that could treat any individual or group unfairly.
• Toxicity Check. This guardrail prevents the agent from producing harmful, offensive, or toxic content in any response.

Custom Guardrails

Beyond the built-in protections, you can define custom guardrails specific to your organisation's requirements. Custom guardrails are configured at Step 4 (Additional Settings) during agent creation, or from the agent details page once the agent exists. They fall into two categories:

• Do's. Affirmative behaviours the agent must consistently follow in all interactions.
• Don'ts. Hard boundaries the agent must never cross, regardless of how a user phrases their request.

Examples of Effective Do's

• Always include ticket IDs when referencing support cases.
• Cite specific document names from the knowledge base when drawing on reference material.
• Request user confirmation before modifying any records.

Examples of Effective Don'ts

• Never share employee contact details externally.
• Never provide legal, medical, or financial advice.
• Never compare products negatively against competitors.
• Never fabricate information that is absent from the connected knowledge bases.

Writing Effective Guardrails

Specificity is the most important quality in a guardrail. A rule such as "Be professional" is too vague to be reliably enforced. Effective guardrails are concrete and testable — you should be able to clearly determine whether the agent has complied with the rule in any given response.

Guardrails vs. Instructions

Instructions guide operational behaviour and allow the agent to apply contextual judgement. Guardrails are non-negotiable and are evaluated before instructions. If an instruction and a guardrail pull in different directions, the guardrail takes precedence — without exception. Use this distinction deliberately: put flexible behavioural guidance in instructions and immovable rules in guardrails.