AI glossary · Safety and regulation

Guardrails

In one sentence

Guardrails are the safeguards placed around an AI model to keep it within agreed limits: checks on what goes in and comes out, limits on topics and actions, and rules for handing a case to a person when needed.

Also known as: AI guardrails · LLM guardrails · AI safeguards · safety rails

What are guardrails?

The word comes from the barriers along a road: they don’t choose your route, but they stop you going off it. In AI, guardrails are everything around the model that stops a mistake or misuse turning into a real problem: a made-up answer, leaked personal data or an action nobody approved.

Put it to work

AI agents that look after your customers

From €49/mo + VAT

Voice, WhatsApp and email with the phone line included: agent, numbers and CRM in one fee.

See plans and pricing

A language model works on probabilities. However carefully you instruct it, it can get things wrong. Guardrails start from that fact: rather than trusting the model blindly, you limit the damage when it slips.

Related to this term

UTIO agents

Available 24/7 in 26 languages
Set it up yourself for €0: built from your website
14-day free trial, no card needed

From €49/mo + VAT

See plans and pricing

How it works

They work in several layers:

Before the model. Check what comes in: off-topic requests, sensitive data that shouldn’t be processed, messages containing prompt injection attempts.
In the instructions. The system prompt sets the role, the allowed topics and what to do when the model doesn’t know.
In the data. The model answers from your documentation rather than from memory, using RAG. If it can’t find the answer, it doesn’t improvise one.
In the actions. The agent has only the permissions it needs. Anything sensitive needs a person to confirm it.
After the model. Check what goes out: format, data that shouldn’t appear, promises that can’t be kept.

The golden rule: anything that would be serious if it happened is blocked outside the model, with code and permissions. Instructions help, but a determined jailbreak can get around them.

Example

An online bike shop in Cardiff sets up an assistant for customers. It defines its guardrails like this:

It only talks about orders, delivery, sizing and returns.
It can check an order’s status but can’t change the address or issue refunds: it prepares those and a person approves them.
It never gives out another customer’s phone number or email.
If the customer is very upset or asks for a person, it passes the conversation to the team.

None of these rules needs a cleverer model. They need a well-built system.

Guardrails at UTIO

UTIO’s agents come with several of these safeguards built in. Autonomy is gradual: no AI, ask first, draft or automatic. The agent starts in draft mode, so a person reviews things before anything goes out. In the support vertical, anything that isn’t in your documentation doesn’t get answered: the agent takes a note and opens a ticket.

The voice agent says at the start of the call that it’s an AI; you can adjust the wording of that notice but not remove it. And you decide when the call goes to your team: by time of day, by topic or when the caller asks.

Frequently asked questions

Is a good system prompt enough as a guardrail?

It's one layer, but not enough on its own. Instructions to the model can be worked around or misread. Anything that must never happen, such as issuing a refund or deleting data, should also be blocked outside the model, with permissions and confirmations.

Do guardrails slow AI down?

Some add a little time, especially those that have a second model check the reply. So you choose them by risk: on sensitive matters it's worth it; for a question about opening hours, you don't need as much.

Are guardrails related to the EU AI Act?

Partly. Depending on the system's risk level, the EU AI Act requires transparency, human oversight and error control. Guardrails are one practical way of meeting those obligations.

Related terms

Prompt injection
Prompt injection is an attack in which someone slips instructions into text an AI model reads, such as a message, an email or a web page, to make it do something other than what whoever set it up intended.
Jailbreak
An AI jailbreak is an attempt to get a language model to bypass its safety rules or the instructions of whoever set it up, using messages designed to trick it or pull its attention elsewhere.
System prompt
A system prompt is the set of fixed instructions an AI model receives before any conversation starts. It sets the model's role, its tone, what it knows about the business and what it isn't allowed to do.
Human handoff
Human handoff is the point at which a chatbot or AI agent passes the conversation to a member of staff, together with the context it has gathered, because the case needs it, a rule requires it or the customer asks for it.
AI hallucination
An AI hallucination is a response from an AI model that sounds convincing but is false or not backed by any source: a figure, quote, price or reference the model invents while generating the most probable text.

Updated on 7 October 2026

← Back to the glossary