Guardrails
Filters, rules, and runtime checks that keep an AI system within desired safety or workflow boundaries.
Which checks belong around a read-only assistant?
A fictional visitor assistant answers questions from public exhibit notices. Its support team wants an incorrect answer to be discoverable and correctable.
- Before retrieval
Accept a visitor’s question and select only sources the public service may read. Restrict the service account to its intended privileges.
- During lookup
The application makes a permitted read request for the current notice and records the source version. An unavailable source leads to an explicit unknown.
- Before display
Check that a claimed time appears in the cited notice and that the notice covers the relevant date. Route conflicts to staff.
What did the guardrail catch?
Log the specific decision and examine both blocked useful answers and unsupported answers that passed. A model answering in a reassuring tone offers no evidence that the lookup or check ran.
What Guardrails Can and Cannot Do
Guardrails reduce risk, but they do not guarantee perfect safety. A weakly designed system can still fail even with many checks in place, and overly rigid guardrails can make the product frustrating or brittle. Good guardrails balance usefulness and control, with rules that are visible enough to audit and flexible enough to support real work.
The best guardrails are layered. A reliable system usually combines prompts, permissions, validation, monitoring, logging, and human review rather than relying on only one safety mechanism.
Follow the idea
Grounding links an answer to the evidence used.
AI Firewall · An AI firewall is one possible inspection layer.
Related concepts (6)
Related Yenra articles (5)
Source · NIST Generative AI Profile on application risk controls
Return to Guardrails in the A–Z glossary