Safety and alerts shows you every moment your bot's safety rules stepped in, and tells the right people when it matters. You find it under Messaging > Safety and alerts.

What is logged

Each time a safety rule acts, Velaro saves one event with the time, the channel, a short reason, and a short excerpt of the conversation. You see these event types:

• Emergency: someone described an emergency and the bot sent them to emergency help.
• Possible emergency: the bot was not sure and asked the person to confirm.
• Setting change request: someone asked the bot to change a therapy or device setting. The bot sent them to their provider.
• Bot could not answer safely: the bot could not check its own answer, so it held back and handed the chat on.
• Answer blocked: a reply broke a safety rule and was stopped before the visitor saw it.
• Visitor report: a visitor reported a problem with an answer.
• Off-topic attempt: someone tried to steer the bot away from your business and the bot declined.

The top of the page shows how many events happened in the last 7 or 30 days, by type and by day. Below that, the events table lets you filter by type and by whether an event still needs review. Select Mark as reviewed once someone on your team has looked at it.

Safety and alerts is part of the Troubleshooting Knowledge add-on. Without it, the page shows how to add it. To add it, contact your Velaro account manager or email support@velaro.com.

Safety rules every site has

Every site has a baseline output filter that stops a bot reply giving medical, legal, or financial advice, and the bot declines abusive visitors and attempts to steer it off topic. Sites with a troubleshooting collection also get the emergency and therapy setting rules, worded for each channel. Each time a rule acts, the event appears on this page as one of the types above.

Who gets alerts

Under Alert settings you choose:

1. Email addresses and phone numbers for text alerts. Add one per line. Use the full number with country code for text alerts, for example +15551234567.

2. Alert me right away for, which types of events send an alert the moment they happen. Emergencies are a good place to start.

3. Summary email, a daily or weekly rundown of everything that happened, or off.

4. Alert when events reach, a count and a time window. For example, 5 events within 60 minutes sends an alert even if none of them are on your right-away list. Set the count to 0 to turn this off.

Turn on Send safety alerts and select Save settings. Then select Send test alert. Everyone on your list should get a test message within a minute. If someone does not, check the address or number for typos.

What "off the rails" means

A bot is off the rails when it says something it should not: wrong instructions, advice outside your business, or a reply that could put a person at risk. Your bot's safety rules catch most of this before the visitor sees it, and each catch shows up here. A sudden jump in events, or any event where a visitor saw a bad answer, is worth a look right away.

How to ask for help

If you think your bot is giving bad answers, open Messaging > Safety and alerts. At the top of the page is a red box titled Is your bot going off the rails? Select We need help in that box, describe what is happening under What is happening?, and select Send to Velaro. Our team is alerted right away and will contact you. Include what the visitor asked, what the bot said, and about when it happened.

Safety controls and turning them off

Velaro's safety controls (emergency and self-harm detection, abuse handling, threat detection, input jailbreak defense and the baseline output filter) are on by default for every account. An administrator manages them under Compliance > Safety Controls. Each control has a switch for AI bot conversations and one for human agent messages, and can be set for the whole account or one channel.

Turning a control off opens an acknowledgement. The administrator must read it and accept it before the control goes off. Velaro keeps a record of the acceptance, and a warning shows on the page while anything is off. Turning a control back on needs no acknowledgement. While a control is off, Safety and alerts still notes when it would have acted.

Two protections are always on and cannot be turned off: the prompt leak check on replies and the block on therapy setting changes.

Abuse handling can be tuned: how many abusive messages bring a warning and how many end the conversation (1 to 10), plus the wording for each channel. Only the Voice wording can mention hanging up.

Related

• Troubleshooting Knowledge Guide