What are AI agent guardrails?

Checks that sit between what an AI agent drafts and what it's allowed to send or do. In customer support that usually means grounding answers in your own documents, blocking invented policies, limiting which actions it can take, and handing off when it isn't sure. Hay.chat runs up to three checks on every answer that states a fact.

How can you tell if customer service is AI?

Ask it. A well-run AI agent should say so, and in the EU, Article 50 of the AI Act has required it since 2 August 2026. Other signs are instant replies at any hour and the same phrasing every time you ask the same question. With Hay.chat, the greeting is yours to write, so say it there.

What does HITL stand for?

Human in the loop. It means a person reviews, approves or takes over part of an automated process. In Hay.chat that happens two ways: approval mode holds AI replies for a team member to approve, edit or block, and the guardrails escalate conversations to your team when an answer can't be backed.

What's the difference between guardrails and approval mode?

Guardrails are automatic. They check answers that state facts and escalate the unsure ones to your team. Approval mode is a switch you choose: every AI reply waits for a team member to approve, edit or block it. You can start with approval on and switch to auto-send once you trust what it's doing.

What does the customer see when a guardrail blocks a reply?

A short message that acknowledges what they asked, in their language, and a hand-off to your team. A separate model writes it from the last six messages and never sees the blocked reply. If that fails, Hay.chat sends the fixed fallback message you set in Settings. The blocked text is kept with the message, so you can see what was stopped.

Which messages skip the guardrail checks?

Stage 0 runs on every answer, even a goodbye, because that's where a bot slips in "I've cancelled it, glad I could help!". Stage 1 reads everything except greetings. Clarifying questions, closing messages, hand-off text and canned errors skip the fact check, since none of them state facts about your business.

Where is the guardrail log, and what's missing from it?

Each conversation has its own log entries for Stages 1 and 2, and the same details sit on every AI message. Developers can read them through the API, and a conversation export includes every message, its timestamp and its metadata. One gap: Stage 0 results go with the hand-off to your team but aren't written to the log yet.

What happens if a guardrail check itself breaks?

If the model behind Stage 0 or Stage 1 times out or errors, the answer goes through rather than stalling the conversation. We chose that over leaving customers waiting, and it's in the technical docs. It's a trade-off: a rare unchecked answer instead of a customer left with no reply.

Can I switch off Stage 0 or Stage 1?

No. Stages 0 and 1 aren't adjustable from the dashboard and are on for every workspace. What you can change is the fact check: the high and medium thresholds, the automatic recheck, the escalation switch and the fallback message, all in Settings, then General, under AI Confidence Guardrails.

Does Hay.chat publish a hallucination or accuracy rate?

No, because a single figure wouldn't mean anything for your store. Accuracy depends on your documents. Every reply that went through the fact check carries its confidence score, the three parts of that score and the documents it used, so you can measure your own. A low score usually points to a missing or muddled document.

AI guardrails that stop made‑up answers before your customers see them

AI guardrails in Hay.chat check every answer that states a fact before your customer sees it. Anything your documents can't back up goes to your team instead.

Hay.chat

Guardrails on

Hi! Do you ship to Australia?

Sure, let me check that for you.

The draft reply said "Yes, we ship to Australia." The shipping policy doesn't say that, so the guardrail blocked it.

A second draft said the same, so the question was escalated to your team.

I can't confirm Australia from our shipping policy, so I've asked the team. You'll have an answer shortly.

The draft said yes. Nothing in the shipping policy backed it, and the customer never saw it.

What happens when the AI isn't sure?

It doesn't get to guess. An answer that states a fact goes through up to three checks before it's sent. Fail one and it gets one retry, then it's escalated to your team.

Stage 0 Did it claim something it didn't do? A reply that says "I've refunded you" is checked against the tools that ran this turn. A failed call doesn't count. If it fails: one more try, then your team.
Stage 1 Did it invent a product or a policy? Blocks made-up products and policies, competitor suggestions and off-topic answers. Critical violations are always blocked. If it fails: one more try, then your team.
Stage 2 Can your documents back it up? Scored on grounding (60%), retrieval quality (30%) and the model's own certainty (10%). 0.8 and up: sent. 0.5 to 0.79: rechecked with more documents. Below 0.5: your team.

The exact numbers are in our technical guardrails documentation.

Can it get things wrong?

A large blue speech bubble stopped at a lowered barrier while a small glowing speech bubble passes through toward a doorway

Does the AI lie to customers?

A language model can state something untrue and sound completely sure of it. When a check catches a bad answer, the customer gets a short message that acknowledges what they asked, in their language, and a hand-off to your team. That message is written without seeing the blocked reply, so the bad claim can't sneak back in.

A tall bookshelf of blue books with a few empty slots and one yellow book sliding into a gap

How often is it wrong?

We don't publish an accuracy rate, because there isn't an honest one that would mean anything for your store. It depends on your documents. Every checked reply carries its confidence score and the documents it used, so you can measure your own. Expect more hand-offs in the first weeks, while your documents have gaps. The safe AI guide has the week-by-week picture.

A small speech bubble tied by a thread to a glowing line in an open rulebook, while a second bubble floats loose with a frayed thread

Can it make up a policy, like the Air Canada chatbot did?

That's what Stages 1 and 2 are for. In 2024 a tribunal held Air Canada to a refund rule its chatbot had made up (Moffatt v. Air Canada). In Hay.chat, a policy your documents don't contain is flagged, retried once, then escalated. Even one that sounds right, like "apply within 90 days", scores low if your page says otherwise. More in what works and what fails with AI chatbots.

What you control

A neat stack of blue coins rising to a thin cap line on a post, with one yellow coin held above the line, apart from the stack

Is it safe to let it act?

Safe enough if you set the limits. Actions follow your playbook rules, so a refund cap is a line in a refund playbook: over the cap, it goes to your team. Stage 0 catches any reply that claims an action that didn't run. Approval mode holds replies, not actions. More on the human approval page.

A blue machine with visible gears inside a clear glass case, one gear glowing yellow, a strip of ticked paper feeding out of a slot

Who can see what it decided?

Your team can, conversation by conversation. Each time Stage 1 or 2 runs, the log records whether it passed, what it found, the reviewer's reasoning, the score and the documents used, plus the original text of any reply that was replaced. Developers can read all of it through the API.

A large blue dial with a cyan arc and a single yellow marker set high, a small slider beside it

Can I change the thresholds?

Yes, for the fact check. Settings, then General, has the high and medium thresholds (0.8 and 0.5 by default), the recheck, the escalation switch and the fallback message. Raise the thresholds and more answers go to your team. We'd leave escalation on, because with it off a low-scoring answer is sent as written.

How do other AI support tools handle made-up answers?

All four check answers against your content in some form, so the difference is in what they publish. Only one says it checks that a claimed action really happened.

Swipe the table sideways to see every column.

How Hay.chat, Fin, Zendesk AI agents and Ada describe their handling of made-up answers: checking answers against your content, checking claimed actions, what happens when unsure, decision logs and published hallucination rates, as of 21 October 2026.
Tool	Checks the answer against your content	Checks a claimed action really ran	When it isn't sure	Decision log	Published hallucination rate
Hay.chat	Yes. Stage 2 scores each company claim against your documents, tool results and playbook	Yes. Stage 0 compares claimed actions with the tools that actually ran	Rechecks once with more documents, then escalates to your team	Yes, per conversation, for Stages 1 and 2	None published
Fin	Yes. A validation step checks the answer is grounded in your knowledge and support content	Not stated	Asks the customer to clarify, or says it can't answer and escalates to your team	Not stated	None. Claims "industry-low hallucinations" without a figure
Zendesk AI agents	Says every generated answer is supported by your approved knowledge sources. No separate check described	Not stated	Falls back and hands off to a human agent	Not stated	None published
Ada	Yes. A final check against your approved knowledge sources. Non-compliant answers are blocked	Not stated	Asks clarifying questions, or hands off to a human agent	Yes. Every conversation is monitored and logged	None published

As of 21 October 2026, from each vendor's own pages: The Fin AI Engine (updated 6 March 2026), Zendesk's reliability architecture article (28 April 2026), and Ada's agent logic docs and trust page. "Not stated" means we couldn't find it there. Pricing, sources shown to customers and EU hosting are compared in the safe AI guide.

What do the standards say an AI agent should have?

US · NIST

The Generative AI Profile (NIST AI 600-1, July 2024) calls it confabulation: "confidently stated but erroneous or false content". Stages 1 and 2 are aimed at exactly that.

OWASP

Excessive Agency (LLM06:2025) asks for least privilege and "human-in-the-loop control to require a human to approve high-impact actions". In Hay.chat that's your playbook rules, with Stage 0 checking the claims.

UK · ICO and NCSC

The ICO's guidance on accuracy wants wrong outputs caught and fixed, and the NCSC's secure AI guidelines ask for logging. That's the hand-offs and the guardrail log.

EU · EDPB

The LLM privacy risks report written for the EDPB (April 2025) suggests post-processing filters to catch inaccurate output. Stages 1 and 2 are post-processing filters.

Find out what your documents can and can't back

Upload your policies, ask the questions your customers ask, and see which answers it sends and which it hands to you. Every check on this page is on from the first day of the trial.

Start Your Free Trial or book a demo

30-day free trial
No credit card required
Set up in under 10 minutes
Hosted in Europe

Frequently asked questions

Not completely, and anyone who says otherwise is selling something. Language models predict likely text, so they'll sometimes produce something confident and wrong. What you can do is catch it before the customer does. Hay.chat's guardrails check answers against your documents and escalate the unsure ones to your team.

Written by Damien Mulhall. Last reviewed 21 October 2026 against Hay.chat's guardrails documentation. See also security, how documents work and pricing.