Skip to content
AskVault Docs
Try Free →

The AI Agents page

Written by , Founder, AskVault · Reviewed by Aashiq
Last updated: · 4 min read

What you see on the AI Agents page

A grid view of all available skills, grouped by category:

  • Core. knowledge_search, escalate_to_human, collect_lead.
  • Support. ticketing_router, sentiment_router.
  • Sales. sdr_lead_qualifier, demo_scheduler.
  • Marketing. content_recommender.
  • E-commerce. wismo_order_status, wc_order_status, cart_recovery.
  • Billing. discount_negotiator, subscription_manager.
  • Custom. custom_webhook.

Each skill card carries:

  • Skill name and category icon.
  • Description. One-line explanation.
  • Plan badge. Minimum plan tier required.
  • Integration status. "Connected" if integrations are wired up, "Not connected" otherwise.
  • Enabled/disabled toggle.

Click any card to open detailed configuration.

Per-skill configuration

Inside a skill's detail view:

  • Triggers. What customer phrases or signals fire the skill. Editable list.
  • Messages. What the bot says when the skill fires. Customer-facing wording.
  • Bounds. Policy limits the LLM can't override (per-visitor caps, per-workspace caps, rate limits).
  • Required integrations. Links to set up Shopify, HubSpot, Stripe, etc.
  • Test runner. Run the skill against a sample input. See output without affecting production.
  • Audit log. Recent invocations with reasoning.

Each section is collapsible. Open only the ones you care about.

Enable or disable a skill

Toggle at the top of the skill card. Changes apply within seconds.

When you disable a skill mid-conversation, in-flight invocations finish but no new ones trigger. The bot falls back to standard knowledge retrieval for affected queries.

Disabled skills don't count toward your plan's skill quota. Useful for keeping irrelevant skills off (e.g., disable cart_recovery if you're not e-commerce).

Integration setup

Many skills depend on OAuth integrations:

The skill card shows the integration status. Click "Connect" to open the OAuth flow. Typical setup: 3 to 10 minutes depending on the provider.

Bounds and policy

Each skill has hard limits the LLM can't override. Examples:

  • discount_negotiator has per-visitor (default 15%), per-workspace ($5,000 USD per month), and per-code caps.
  • collect_lead runs once per conversation.
  • custom_webhook has rate limits per workspace and per conversation.

Configure bounds under the skill's Bounds section. Changes apply immediately to new conversations; in-flight conversations honor the previous bounds until the next turn.

Per-channel skill settings

Some skills behave differently per channel. Configure per-channel overrides under the skill's Per-Channel Settings:

  • escalate_to_human. Different routing for web widget vs WhatsApp vs Slack.
  • collect_lead. Different prompts in different channels.
  • sentiment_router. Tighter thresholds in chat-heavy channels.

Per-channel overrides inherit from the workspace default. Override only what differs.

Test runner

Before pushing a skill change to production, test it:

  1. Open Test Runner in the skill detail view.
  2. Enter sample input. A customer message that should trigger the skill.
  3. Click Run.
  4. See the output. What the bot would say, which trigger fired, what policy checks passed.

Test runs don't affect production. They don't count against your query quota. Use them liberally when adjusting trigger phrases or bounds.

Agent insights panel

The Agent Insights panel surfaces patterns across all skills:

  • Skill firing frequency. Which skills run most often.
  • Skill success rate. How often each skill's invocation actually completes successfully.
  • Failure breakdown. When a skill fails, what's the cause (integration down, policy bounds hit, customer dropped off).
  • Customer satisfaction per skill. Thumbs up/down rates filtered by which skill fired.

Use insights to refine skill bounds and trigger phrases. Skills with low success rates usually need either tighter triggers or relaxed bounds.

Webhook subscriptions per skill

Each skill can fire workspace-level webhooks on certain events:

  • collect_lead fires lead.captured.
  • escalate_to_human fires conversation.escalated.
  • sentiment_router fires sentiment.threshold_crossed.
  • custom_webhook fires custom_webhook.triggered.

Subscribe under Dashboard > Webhooks to integrate with external systems like CRMs, alerting tools, or your own backend.

Limits

  • Plan availability. Skill availability varies by skill. See the skills overview for the per-plan matrix.
  • Concurrent skill firings. Up to 3 skills can run in parallel per bot turn.
  • Custom skills. Via custom_webhook only today. Native SDK on the roadmap.

Common pitfalls

Skill enabled but never fires. Trigger phrases don't match. Test under Test Runner with sample messages. Add synonyms if customers say it differently than the default list.

Skill fires but action fails. Required integration isn't connected. Check the skill card's integration status.

Policy bounds too tight. Skill keeps refusing. Loosen the bound under the skill's Bounds section.

Customer-facing messages off-brand. Default messages are generic. Customize under each skill's Messages section.

FAQ

Can I create custom skills?

Today via custom_webhook. Configure trigger phrases and a webhook URL; your endpoint implements the skill logic. Native custom-skill SDK on the roadmap.

Do skill invocations count against my query quota?

Yes. Each skill firing counts as 1 query. A conversation that fires 3 skills consumes 3 queries plus the standard answer query.

Can I see which skill fired on a specific conversation?

Yes. Open the conversation in Live Chat. Each bot message tagged with the skill that produced it.

Can I A/B test skill configurations?

Yes via separate workspaces. Configure one workspace with config A, another with config B, split traffic at the widget level. Compare in Analytics.

Does the bot need permission to invoke skills?

Some skills require identity verification (e.g., subscription_manager won't reveal account details to anonymous visitors). Others (knowledge_search, collect_lead) work for anonymous visitors. Configure per skill.

Was this page helpful?