No. A model's guardrails decide what the model will say: they filter harmful content and refuse certain requests. Governance decides what the agent may do in your business, for which customer, up to what limit, and who answers for it. That decision belongs to named people, is enforced outside the model at the moment of action, and is recorded.
What do model guardrails actually do?
Model guardrails are trained behaviour and filters added by the model provider. They stop the model producing dangerous instructions, abusive text or certain kinds of personal data. Some deployments add their own input and output filters on top. Our explainer on what guardrails in AI are covers the types.
They are useful. They are also general. A model provider does not know your refund policy, your approval limits, which suppliers are on your master file, or that this customer's account is frozen. Guardrails cannot enforce rules they have never seen.
Why can't the model enforce your rules if you tell it them?
You can write your rules into the agent's instructions, and a good model will usually follow them. "Usually" is the problem. A model's compliance with instructions is probabilistic, can be argued with, and can shift when the provider updates the model. A prompt injection hidden in a document can override it.
There is also a question of evidence. When an auditor asks why the agent refunded 900 to a customer, "the model decided it was allowed" is not an answer. You need a record showing which rule applied, who set it, and that it was checked. That is the case made in guardrails are not governance.
Who decides, where is it enforced, and what is recorded?
| Question | Model guardrails | Governance |
|---|---|---|
| Who sets the rule? | The model provider | Your process owner, within platform limits |
| What does it cover? | Content: what the model says | Actions: what the agent does, for whom, how far |
| Where is it enforced? | Inside the model, probabilistically | Outside the model, before the action reaches a system |
| Does it know your context? | No | Yes: your limits, customers, sources |
| What is recorded? | Often nothing you can inspect | Every verdict, with the rule that decided |
| Who answers for it? | Nobody in your organisation | A named owner |
The principle is to govern the action, not the model. Whatever the model proposes, the action is checked against your bounds before it happens.
A worked example
Take a lender whose collections agent drafts messages to customers in arrears. The model's guardrails stop it writing threats or abuse. Good.
One customer writes that she has just been diagnosed with a serious illness and asks for a payment break. The model, being helpful, drafts a reply offering a three-month holiday on interest. Nothing in its guardrails objects: the text is kind and appropriate.
The lender's governance does object. Agreeing forbearance is outside the agent's role scope, and the customer's message signals possible vulnerability, which the lender's policy routes to a specialist team. The draft is held, the verdict names the role scope bound, and a specialist calls the customer. The guardrails judged the words. Governance judged the action.
What about agents supplied by vendors?
Vendor agents often arrive with the vendor's own guardrails and a claim that they are "safe". The same answer applies: the vendor's controls are inputs to your governance, not a substitute for it. Our answer on how to govern AI agents supplied by vendors covers what to ask for.
Which controls make governance hold where guardrails stop?
- Bounds outside the model. Role scope, limits and sources checked on every action.
- Named ownership. Each agent has an owner who sets and answers for its bounds.
- Evidence from your own cases before the agent acts, not the provider's benchmark.
- A filed verdict for every action, so decisions can be reconstructed.
- A management system around it. Frameworks such as the NIST AI RMF Govern function for AI agents describe the policies, roles and accountability that sit above any single control.
The topic guide, agent governance is a bound, not a model, sets out the full argument.
Questions people ask
Should we turn off model guardrails since they don't govern?
No. Keep them as one layer. They catch harmful content your own rules may not anticipate. Just do not rely on them for decisions about your business.
Can we fine-tune a model to know our rules?
You can shape its behaviour, but the result is still probabilistic and changes with each retrain. Rules that matter should be enforced outside the model, where they hold regardless.
Do platform guardrails count as governance?
Platform guardrails enforced at the moment of action are part of governance: they are one of the bounds every action must meet. The difference is where they are enforced and whether they are recorded.