Short answer
A guardrail is a hard limit that holds no matter what, while a guideline is recommended practice that can bend to fit the situation. In AI, guidelines shape how an agent behaves and guardrails decide what it can never do.- Guardrail
- A limit enforced by the system, whatever anyone asks
- Guideline
- Advice followed with judgement, which can bend
- Guardrails for
- Money, promises, private data, actions that cannot be undone
- Guidelines for
- Tone, wording, length, the order of steps
Guidelines and guardrails are often mentioned together because they do different halves of one job. Guidelines describe the right way to work. Guardrails make sure that even when something goes wrong, it cannot go wrong past a certain point.
The difference at a glance
The test is what happens when the rule is broken.
| Guideline | Guardrail | |
|---|---|---|
| What it is | Recommended practice | A hard limit |
| Enforced by | The judgement of whoever follows it | The system itself |
| Can it bend? | Yes, when the situation calls for it | No |
| If ignored | The result is worse | The action is blocked |
| Example | "Reply within four hours" | "Refunds over $200 need a manager" |
Guidelines and guardrails in AI agents
For an AI agent in a Shopify store, the split looks like this.
- Guidelines: keep replies short, use the customer's first name, offer store credit before a refund, suggest a size from the fit notes.
- Guardrails: change an order only for a customer who has proved it is theirs, never claim to be a person, never refund more than the order allows, send refunds for approval.
The important detail is where a guardrail lives. An instruction in the agent's prompt is a guideline, however firmly it is worded, because the model decides whether to follow it. A guardrail sits outside the model: a check that runs on every reply or action and blocks the ones that break the rule.
The practice this page rejects is writing every rule into the prompt and calling the result guardrails. Writing "never" in an instruction does not make it a limit.
Guardrails in business and law
Managers use guardrails for the boundaries inside which a team decides freely: a budget limit, a discount ceiling, a list of things that always need sign-off. Inside them, guidelines describe the preferred way, and people use judgement. In law and policy, guardrails usually means limits set by regulation, such as rules on how AI systems may use personal data.
In Aide, sign-off works the same way: a large refund requested by a teammate with the Agent role waits for an Owner or Admin, who gets an "Order action needs approval" notification and approves or rejects it on the order.
Guardrails vs guiderails
The physical words follow the same logic. A guardrail stops something from going over an edge: a barrier on a road or a railing on a balcony. A guide rail, or guiderail, steers movement along a path, like the rails a lift car or a sliding door runs on. In some US regions, road crews call roadside barriers guide rails, so the two words can name the same object.
As metaphors, guiderails sit closer to guidelines: they steer without stopping anything.
How to decide which one a rule should be
Ask what one failure would cost.
- Could it cost money, break the law or expose private data? Make it a guardrail.
- Could it be undone easily? A guideline is usually enough.
- Does the right answer depend on the situation? Keep it a guideline, or it will block the cases it was not written for.
- Does breaking it once mean losing a customer's trust? Make it a guardrail.
In Aide, the answer can differ by topic: on a topic set to Drafts, an order action the shopper confirmed still waits for the team to approve, while on an AI Agent topic Aide acts within the matching procedures.
Where Aide fits
In Aide, Instructions are the team's always-on rules, such as "Offer store credit before a refund", and the agent follows them in every conversation. The limits sit outside the model: every reply is checked before it is sent and rejected if it promises or does something the topic's AI level does not allow, or claims to be a person; order actions are only for shoppers who prove the order is theirs; and refunds wait for a teammate by default. See AI customer service guardrails.