Illustration: spatialship, future, space

AI Guardrails in 2026: Gaps and Who Owns Them

AI guardrails are runtime controls that filter what an AI system accepts, retrieves, says and does. They reduce risk but do not remove it: filters miss novel prompts, and wrong answers still bind the company. Treat each guardrail as an owned control with a named owner, a test and an evidence trail.

In February 2024 a British Columbia tribunal ordered Air Canada to pay a passenger roughly CAD 650, plus interest and fees, after its website chatbot gave wrong bereavement-fare advice. The tribunal treated the chatbot as part of the airline’s website and held the airline responsible for what it said, according to the American Bar Association’s summary of Moffatt v. Air Canada. The amount was small. The principle was not. If your booking assistant, loyalty chatbot or concierge agent gets a cancellation rule wrong, your company answers for it, whatever filter sits in front of the model.

What do AI guardrails actually do?

A guardrail is a runtime check wrapped around a model. In a typical assistant it sits at five points in a request, and each point blocks a different failure. The failure names below come from the OWASP Top 10 for LLM Applications (2025 edition), which lists ten risks.

LayerWhat it checksFailure it targets (OWASP 2025)
InputPrompts and uploaded content before the model sees themLLM01 Prompt Injection
RetrievalWhich documents and records the model may pullLLM02 Sensitive Information Disclosure
OutputThe answer before the guest sees itLLM09 Misinformation, LLM05 Improper Output Handling
ActionWhich tools the agent may call: refunds, rebooking, creditsLLM06 Excessive Agency
UsageRate, token and spend limitsLLM10 Unbounded Consumption

A guardrail is a control, not a product feature. It has a rule, a threshold, an owner and a log. Many teams buy the filter and skip the other three, which leaves a filter nobody tunes and nobody can defend in an audit.

What do guardrails miss?

Guardrails fail in three predictable ways.

First, they miss prompts nobody anticipated. The NIST Generative AI Profile (NIST AI 600-1, July 2024) notes that restricting outputs may still produce harmful recommendations in response to less explicit or novel prompts. It also asks organizations to define the kinds of queries an application should refuse and to evaluate possible circumvention of safety measures on a regular basis. A filter built last spring does not know this spring’s jailbreak.

Second, they check content, not correctness against your own policy. A toxicity filter will pass a fluent, polite, wrong answer about a refund window. In the Air Canada case the correct policy sat on another page of the same site, and the tribunal held that customers should not have to cross-check one part of a website against another. The control that matters is grounding: the assistant answers from the current fare rules or cancellation terms, cites the version it used, and refuses when it finds no source.

Third, they drift. A model update, a new prompt template or a newly connected tool changes what the assistant can do, and the guardrails written for the old setup no longer cover it. An agent that gained a rebooking tool last month has an Excessive Agency problem that no input filter will catch.

Who owns each guardrail?

This split is written for a VP Engineering or Head of Digital Product at a hotel group, airline, online travel agency or tour operator. Adapt the names to your org chart; keep the principle. The person who accepts the risk owns the threshold, and engineering owns the mechanism.

GuardrailDecision ownerBuilds and runs itEvidence to keep
Input and prompt-injection filtersHead of SecurityPlatform engineeringBlock-rate log, red-team results
Grounding sources (fare rules, cancellation terms, loyalty terms)Revenue or customer-policy ownerProduct teamSource version cited per answer
Output rules (what the assistant must refuse)Head of Digital ProductProduct plus legal reviewRefusal criteria, sampled answer reviews
Action limits (refund caps, rebooking rights)Finance and operationsEngineeringApproval thresholds, action audit log
Monitoring and drift testsHead of AI governance or riskEngineeringTest schedule, results, incident log

Two owners per row is deliberate. When a guardrail has one owner, either security tunes it until the assistant is useless, or the product team loosens it until it is unsafe. Splitting the threshold decision from the mechanism keeps both pressures visible.

What does a guardrail map look like for a booking assistant?

The numbers in this example are illustrative. Set yours from your own loss and complaint data.

Take a hotel group’s assistant that answers rate and cancellation questions and can modify bookings. Its map has four lines:

  1. Ground every rate and cancellation answer in the live policy store. If retrieval returns nothing, the assistant says so and hands off to an agent.
  2. Cap automatic credits and refunds at a fixed amount, for example $100, and route anything above it to a human approver.
  3. Scope retrieval to the authenticated guest, so a crafted prompt cannot surface another guest’s reservation (the LLM02 case).
  4. Log the prompt, the source version, the answer and any action for every conversation, so a disputed answer can be reconstructed.

Each line names a layer from the table above, a decision owner and the evidence it produces. That is the unit an auditor, a regulator or a claims team will ask to see. The same map feeds the mapping exercise that frameworks such as the NIST AI RMF organize around its Govern, Map, Measure and Manage functions.

How do you test that guardrails still work?

Run a fixed test set every time the model, the prompt template or a connected tool changes. Start with 20 to 50 questions: one adversarial case for each OWASP entry that applies to your assistant, plus the policy questions your guests ask most. Track three numbers: refusal accuracy, grounded-answer rate and escalation rate. A drop in any of them after a change is a release blocker.

Testing has a cost, and so does the guardrail itself. Every extra classifier call adds latency and spend, so track the cost per prevented incident next to the cost per conversation. The AI performance and ROI view is where those numbers belong, rather than in a separate engineering spreadsheet.

Regulation is moving the same way. Frameworks and laws increasingly expect controls that are documented, assigned and tested, which is the same evidence a guardrail map produces. Mapping each guardrail to the requirement it supports, whether in the EU AI Act, ISO/IEC 42001 or the NIST AI RMF, avoids producing that evidence twice. The AI regulations overview lists the frameworks Fruggr covers, and the Trusted AI page shows how the pieces connect. Verify current legal text for your markets before relying on any mapping.

What should you do this week?

  1. List every customer-facing and employee-facing AI assistant, including AI features built into SaaS tools your teams already use.
  2. For each one, write down which of the five layers has a control today.
  3. Name one decision owner per layer, using the table above as a starting point.
  4. Identify the policy sources the assistant must cite, and confirm which version it reads.
  5. Run a 20-question test, file the result, and schedule it to rerun on every model or prompt change.

Key takeaways

  1. A guardrail is a control with a rule, a threshold, an owner and a log; a filter alone is not a control.
  2. Filters miss novel prompts and wrong-but-polite answers, so ground answers in current policy sources and test for circumvention on every change.
  3. Split each guardrail between a decision owner who accepts the risk and an engineering owner who runs the mechanism, and keep the evidence.

Next step: Guardrails can only cover the AI you know about. See how Fruggr’s Shadow AI detection surfaces the AI tools, models and embedded AI features in use across your organization, then map a guardrail owner to each one.

FAQ

What are AI guardrails?

AI guardrails are runtime controls placed around an AI system. They check inputs, retrieved data, outputs, actions and usage limits so the system stays within the rules the organization has set.

Do guardrails make an AI chatbot legally safe?

No. Guardrails reduce the chance of a harmful or wrong answer, but the company stays responsible for what its assistant says, as the Moffatt v. Air Canada decision illustrates. This article is not legal advice; check the rules in your markets with counsel.

Who should own AI guardrails?

Split ownership by layer. The business or risk owner who accepts the exposure sets the threshold, and the engineering team that builds the mechanism runs it. A governance or risk function should own monitoring and drift tests.

How are AI guardrails different from an AI gateway?

A gateway routes and controls access to models and often handles logging, keys and rate limits. Guardrails inspect the content and the actions: what goes in, what comes out and what the agent is allowed to do. Many setups run both.

How often should AI guardrails be tested?

Test them whenever the model, prompt template, retrieval source or connected tool changes, and on a fixed schedule in between. New jailbreak techniques appear without any change on your side.

This article was drafted with AI assistance and reviewed by our editorial team before publishing.