How does AskVault decide when to call a skill vs answer from the knowledge base?

Each skill has a set of trigger intents. When the visitor's message matches a trigger intent with sufficient confidence, the skill fires. If no intent matches, the bot answers from the indexed knowledge base directly.

Can AskVault skills run silently without showing a message to the visitor?

Yes. Silent skills like sentiment_router operate in the background: they tag the conversation or route it without sending a visible message to the visitor. Toggle this per skill.

Which AskVault plans include skills?

Skills need the Growth plan or higher. Free and Starter workspaces can see the skills list, but no skill is handed to the AI on those plans. Discount, subscription and custom-webhook skills need Business. Every plan answers from your knowledge base without any skill.

AskVault skills, an overview

Written by Joseph Russo, Principal Architect, Networks and Tech, AskVault · Reviewed by Allison Chan

Last updated: June 18, 2026 · 6 min read

What a skill is

A skill is a structured capability the AI agent can invoke during a conversation. Different from a simple knowledge-retrieval answer in three ways:

Skills can take actions. Hand a conversation to your team, share a Calendly booking link, look up an order, capture a lead. Knowledge retrieval just produces text.
Skills have policy bounds. Each skill has hard limits the LLM cannot override (per-visitor caps, per-workspace caps, rate limits, audience restrictions).
Skills are configured per workspace. You decide which skills are on and set each one's options, such as endpoints or dry-run mode.

A conversation might invoke zero, one, or several skills. Most conversations use zero (just knowledge retrieval). Sales-intent conversations may use 3 to 5 (collect_lead, sdr_lead_qualifier, demo_scheduler, then maybe escalate_to_human).

The built-in skills

Categorized by purpose. Your workspace's own list is the authority: open Automation › Skills to see exactly which of these are available to you and which are switched on:

Skill	Purpose	Plan
knowledge_search	Let the agent run an extra search of your knowledge mid-answer (normal answers use your knowledge without it)	Growth
collect_lead	Capture name, email, phone with conversation context	Growth
collect_info	Show an editable contact form the visitor checks and confirms before it is sent	Growth
escalate_to_human	Hand off to live chat with full context	Growth
ticketing_router	Meant to create Zendesk tickets from chat; Zendesk can't be connected for this today, so it creates none. Use escalate_to_human instead	Growth
sentiment_router	Detect frustration and route or de-escalate	Growth
sdr_lead_qualifier	Send qualified prospects to HubSpot	Growth
demo_scheduler	Share Calendly booking links in chat	Growth
wismo_order_status	Live order lookup via Shopify	Growth
wc_order_status	Live order lookup via WooCommerce	Growth
cart_recovery	In-chat, policy-approved concession on a Shopify cart (cart lookup not wired yet)	Growth
search_products, get_cart, update_cart	Live Shopify catalog search with product cards, and the shopper's cart	Growth
discount_negotiator	Apply pre-approved codes within bounds	Business
subscription_manager	Stripe upgrade, downgrade and refund actions (plan changes run after the visitor confirms; refunds are still declined)	Business
custom_webhook	Fire your own webhook with conversation context	Business
recommend_product	Answer product, price and stock questions from your synced catalog Rolling out	Growth

How the bot decides to use a skill

Skills are tools the AI can call. On each message it reads the conversation and the descriptions of the skills you have turned on, and decides whether one would help: an order question may call an order lookup, a request for a person may call escalate_to_human. Most messages need no skill, and the bot simply answers from your knowledge base.

There are no trigger-phrase lists or intent classifiers to configure. To steer when a skill is used, turn it on or off, or add a line to your system prompt such as "Only offer a demo after the visitor shares a work email."

Policy bounds the LLM can't override

Each skill has bounds enforced at the AskVault layer, not in the LLM prompt. This matters because LLM-only guardrails are bypassable; AskVault's policy layer isn't.

Examples:

discount_negotiator has a maximum discount (15% by default). Even if the AI "decides" to offer 50% off, the policy layer rejects it.
collect_lead never creates a second lead for the same email in a workspace.
custom_webhook gives up on a call after 15 seconds and never retries a 4xx rejection, so a slow or failing endpoint cannot stall the conversation for long.

Set each skill's limits in its settings on the Skills page.

Customizing skill behavior

Each skill has a configuration box (JSON) on the Skills page for options such as require_phone for collect_lead, min_company_size for sdr_lead_qualifier, or dry_run for skills that call outside services. What the bot says around a skill follows your system prompt and persona.

Enabling and disabling skills

Each skill has an on/off toggle. A disabled skill is not offered to the bot. A server-side check that also refuses any skill the bot asks for by name when it was not offered in that conversation is rolling out behind a flag and is not switched on yet. Disable skills you don't need:

Free-tier customers don't need subscription_manager if they're not on Business plan.
Internal helpdesk bots don't need cart_recovery or discount_negotiator.
B2B SaaS without e-commerce doesn't need wismo_order_status.

Reducing skill surface area improves accuracy (fewer false-positive trigger matches) and audit clarity.

Skill execution latency

Turning skills on adds a step to each answer: the AI first decides whether a skill is needed, and a skill that calls another service waits for that service to reply. That can add a few seconds to an answer, and slow external systems (for example a store on shared hosting) add more. Switch on only the skills a workspace actually uses.

Audit logging per skill

AskVault records each skill run on the server, with its inputs and result, so a run can be traced afterwards.

Multi-skill conversations

Conversations can invoke multiple skills in sequence. Typical sales flow:

Visitor asks pricing. Bot answers via knowledge_search.
Visitor shows buying intent. collect_lead captures contact info.
sdr_lead_qualifier records company size, stage and budget, and updates HubSpot.
Lead syncs to HubSpot.
Visitor asks for a demo. demo_scheduler offers a Calendly booking link.
Booking confirmed. Confirmation emails fire.

Each skill maintains state in the conversation context. The bot's reasoning carries across skill invocations.

Limits

Plan availability. Skills start on Growth ($99/month, 6,000 messages, 5 workspaces); discount, subscription and custom-webhook skills need Business ($249/month, 15 workspaces). See the table above.
Custom skills. Today through custom_webhook (Business) and custom actions.

Common pitfalls

A skill runs when it shouldn't. Turn off skills you do not need, and say in your system prompt when a skill should not be used.

A skill never runs. Check it is turned on, that any service it needs (Shopify, HubSpot, Calendly) is connected, and that your plan is Growth or above.

The wording around a skill feels off. Adjust your system prompt and persona; the bot's tone comes from there.

Was this page helpful?