Platform · Guardrails
The safety layer that makes AI support safe to deploy.
Destructive actions require confirmation before they execute. Frustrated customers escalate before they churn. Knowledge boundaries surface honestly. Guardrails aren't an afterthought — they're how Velor was designed.
How it works.
- 01
Classify
Every turn, the agent classifies: Is this a routine request? A regulated or off-limit topic? A destructive action? A question at the KB boundary? A sentiment signal worth escalating?
- 02
Confirm before acting
Any action flagged as destructive waits for a human before the webhook fires: the customer sees the exact parameters it will run with — order, amount, currency — and presses Confirm on that card in the chat widget, or, with the approval queue on, a teammate approves it instead.
- 03
Detect and escalate
On the chat widget, a visitor whose messages score as frustrated or angry past a threshold you set is escalated — with full conversation context passed to the human agent.
- 04
Surface the boundary
When the agent can't answer confidently from the KB, it says so clearly — "I don't have that in my knowledge base" — and offers to connect the customer with a human.
What each guardrail does.
Regulated-topic gate
Off-limit topics aren't just discouraged in the prompt — a deterministic gate checks every visitor message before the model responds. If it matches, the agent refuses and hands off to a human, so a regulated ask can't be answered even if the model would have. Financial, legal, medical, and tax advice are recognized automatically; add your own match patterns for the rest.
Destructive-action confirmation
Any action marked destructive:true triggers a confirmation step. The agent describes what will happen in plain language — 'This will cancel your order, which cannot be undone' — and waits for explicit confirmation.
Sentiment-based escalation
On the chat widget, each visitor message is scored for sentiment; when the score falls to the workspace threshold, the conversation escalates (never on the visitor's very first message). The threshold is a single workspace setting — disabled, severe, moderate, mild or a custom value — in the routing settings. It is not a list of signals you define.
KB boundary handling
When retrieval doesn't surface a confident answer, the agent explicitly acknowledges the gap rather than guessing. Customers get an honest response, not a confident hallucination.
Configurable per action and channel
Whether an action needs confirmation is set per action (destructive or not). The sentiment escalation threshold is set per workspace. Destructive actions are confirmed by the customer only in the chat widget; on voice, SMS, WhatsApp, email and Slack they wait for a teammate's approval instead.
What triggers an escalation.
Every signal is configurable. You can adjust thresholds, add custom keywords, and define where escalations route — by team, skill, or availability.
Run it on your own tickets.
Every workspace starts with a 14-day free trial. No credit card. If you want a guided walkthrough against your data, we're happy to run one with you.