AI agent guardrails decide when an agent answers, when it hands off to your team and when it stays quiet. Pacing decides where it works and how fast it replies. Together they are what makes it safe to move an agent from Copilot, where a teammate sends every reply, to Review or Autopilot. This guide walks through the Guardrails and Behaviour sections of the agent editor, explains what happens in the inbox when the agent hands off and shows which analytics tell you whether the settings are right. It assumes you already have an agent; if not, start with Set Up Your First AI Agent.
Guardrails apply in every mode, including Copilot and the Playground, so you can tune them before the agent ever sends anything on its own. Pacing only takes effect in Autopilot.
Pick a carefulness preset
The Guardrails section opens with three presets. Each one sets the three controls underneath it, and the badge in the section header shows which preset is active.
Balanced is the default. Your own rules, the never discuss list and the reply limit stay in place whatever the preset.
| Preset | Confidence threshold | Hand-off style | Require sources |
|---|---|---|---|
| Cautious | 80% | Hands off readily | On |
| Balanced (default) | 60% | Balanced | Off |
| Confident | 40% | Only on explicit rules | Off |
Cautious only answers with sources, needs high confidence and hands off to a human readily. Balanced answers when reasonably confident and hands off when the question is outside its knowledge. Confident answers more often and hands off only when one of your Hand-off rules applies.
Change any of the three controls by hand and the badge switches to Custom. Clicking a preset again asks Replace custom settings? before it overwrites what you changed.
The three controls behind the presets
Confidence threshold is a slider from 0 to 100. With every decision the agent reports how sure it is that the reply is correct and complete. Replies below the threshold make the agent hand off instead, or stay quiet when Hand-off style is Only on explicit rules. In the Playground the readout shows this number as, for example, 78% confidence, which is the quickest way to find a threshold that fits your help center.
Hand-off style is how readily the agent passes to a human:
- Hands off readily: whenever the knowledge does not clearly cover the question, the agent prefers handing off over guessing.
- Balanced: answers when the knowledge supports it, hands off when it does not or when the customer explicitly asks for a person.
- Only on explicit rules: answers whenever it reasonably can, and hands off only when one of your Hand-off rules applies or the customer explicitly asks for a person.
Require sources means the agent only replies when the answer is backed by an article or data source. With it on, a question your help center does not cover is handed off instead of answered with a guess. It is the single most effective setting against made up answers, at the cost of handing off more often until your articles catch up. AI Question Insights shows you which topics those are.
Rules you write yourself
Below the presets are the controls that encode your own policy.
Hand-off rules is a list, one per line. When any applies, the agent hands off instead of answering. Write them as situations, for example "the customer asks to cancel a paid plan" or "the customer threatens legal action". These fire in every preset, including Confident.
Never discuss is a list of topics, one per line, such as "pricing negotiations" or "competitor comparisons". The agent declines and offers a human. Use it for anything where a wrong answer is costly or where only a person should speak for the company.
Max replies per conversation caps how many times the agent replies in one thread. After this many replies the agent hands off. Leave it empty for no limit. A cap of three or four is a sensible default for Autopilot: if a conversation is still going after four replies, a human is usually faster.
Admits it is an AI when asked is fixed in every preset. The agent never claims to be human, even when the AI label in the Identity section is hidden.
Pacing: where the agent works and how fast it replies
The Behaviour section controls which channels the agent covers and how quickly it answers. It opens with the same three personality presets you saw when creating the agent: Human-like, Transparent (the default) and Instant. Each card spells out what it means in practice, for example "Live chat: replies after ~6 s · typing bubble · waits while the customer types".
Pacing is split by channel because a chat visitor is waiting and an email sender is not.
A note under the presets reminds you how pacing applies to the current mode: Copilot suggestions are always instant; Review drafts are prepared right away and wait for approval, so pacing does not apply; Autopilot enforces this pacing, seconds on live chat and minutes on tickets and email.
Live chat
The switch in the Live chat header turns the agent on or off for chat. The customer is waiting, so the delays here are short.
- Reply delay: Instantly, A few seconds (3 to 8 seconds), Like a busy human (15 to 60 seconds, and longer replies take longer) or Custom range (seconds).
- Wait while the customer is typing holds the reply until they stop, so the agent does not answer half a question.
- Show typing indicator shows a typing bubble before the reply appears.
Tickets & email
Here replies are expected to take a while, so the delays are in minutes.
- Channels: tick any of Widget tickets, Ticket center, Contact form and Email. Untick a channel and the agent ignores conversations that arrive through it.
- Reply delay: Instantly, Within a few minutes (1 to 3 minutes), Like a busy human (4 to 15 minutes, growing with reply length) or Custom range (minutes).
Pacing is more than cosmetics. A reply that arrives a few seconds after the customer stops typing reads as attentive; one that arrives in 300 milliseconds reads as a bot, and customers answer bots differently. If a teammate replies during the delay, the agent's pending reply is cancelled rather than sent on top.
Availability, intro and daily cap
- When is either Always or Only when no operator is online. The second option makes the agent an out of hours cover: it answers only when nobody from your team is online, which pairs well with Hands off readily.
- Offer a human decides when the agent mentions that a teammate can take over: When unsure, In the first reply or Never.
- First reply is either Introduces itself as an assistant or No intro.
- Daily cap is the maximum number of automatic drafts per day. Leave it empty for no cap. When the cap is reached, Copilot shows This agent reached its daily suggestion cap. and an Autopilot agent hands the conversation off. It is a useful brake on token spend while you are still tuning.
If you run several agents on overlapping channels, remember the rule from the AI Agents list: when several agents match a conversation, the first one wins. Put the narrowest agent first.
What happens in the inbox when the agent hands off
When an Autopilot agent decides to hand off, four things happen at once:
- The agent stops working on that conversation and leaves a private note in the thread with the reason, for example "The customer asked to cancel a paid plan". Customers never see the note.
- The conversation is marked unresolved and unread, and shows a Needs human badge. The Needs human filter in the inbox sidebar lists all of them in one place; AI handled lists every conversation an agent is or was handling.
- The chip in the thread changes from Bella is handling to Bella handed off, and the agent panel shows the Hand-off reason together with the Last decision and the Sources used.
- Your team receives the new message email with the agent's name and its reason, even when Staff email notifications is set to Only on hand-off.
From there, reply to the customer as usual. If the agent can take over again, click Hand back. The same panel has a Pause button for the opposite situation, when you want to take a conversation away from the agent before it decides anything. Replying yourself also pauses it, and the chip reads Paused — a human replied.
In Review mode handing off works exactly the same way; there is simply no draft to approve, so the conversation goes straight to Needs human. In Copilot the card says Bella recommends a human handles this one. and leaves the reply to you. Read Working the Shared Inbox for the rest of the inbox workflow.
Checking the settings with analytics
Click Analytics on the agent's row in the AI Agents list and switch between 7 days, 30 days and 90 days. Four metrics tell you whether the guardrails and pacing are tuned well:
- Hand-off rate is the share of decisions that handed the conversation to your team instead of replying. A rate that stays high usually means Require sources is on and the help center has gaps, or the Confidence threshold is above what your articles can support.
- Resolved without a human counts conversations closed after the agent replied, with no operator reply in between. This is the number that should grow as you loosen the guardrails.
- Average confidence is the agent's own confidence across its decisions. If it sits well above your threshold, the threshold is not doing any work; if it hovers around it, small changes to the threshold will swing the hand-off rate a lot.
- Cancelled and Superseded show replies cancelled because an operator stepped in and drafts replaced because the customer wrote again before they were sent. Both rise when the reply delay is longer than your team's or your customers' patience.
A practical order: start with Cautious and Require sources on, let the handed off conversations show you where the help center is thin, write those articles, then move to Balanced and watch Resolved without a human climb. Whenever the agent hands off for a reason that should have been a reply, a thumbs down with a short reason turns into a suggested change on the Suggestions tab, as described in Train an AI Agent with Skills and a Markdown File.