Glossary
Definition
Guardrails are the enforced constraints on what an AI agent may read, do, and decide, including data scopes, approval requirements, and cost limits.
Guardrails turn a capable model into a deployable agent. They define the data the agent may read (and as whom), the tools it may call, the actions that require a person's approval, the topics it must escalate rather than answer, and the limits on how much it may spend or how often it may act. They also define what is logged.
The distinction that matters is where guardrails are enforced. Instructions in a prompt can be overridden by input the agent reads. Constraints enforced by infrastructure below the agent, at the data layer and the tool gateway, cannot. Buyers should ask which of a vendor's guardrails are prompts and which are policy.
Meridian enforces data scopes in the Data Fabric and tool and approval policies at the Gateway. Agents, builders, and prompts cannot remove them.
Read next
Long-form explainers and frameworks that use this term, dated and signed.
Buying and ROI
How credit-based pricing works for AI agents: what counts as a billable action, Meridian's published credit rates, three worked examples, budgeting, and guardrails against overage.
Aisha Bello6 min read
Finance operations
How the Close, Controls, Audit, and Planning Agents remove three business days from the month-end close: orchestration, reconciliations, evidence, controls, and a before-and-after close calendar.
Priya Raman6 min read
Get started
Start with one workflow, one approver, and one number to move. Most design partners were live in five weeks.