Skip to main content
A skill is a job the agent is allowed to do. Switching one on adds a paragraph to the agent’s instructions and, for most of them, gives it a tool — a real lookup against your own data rather than a promise in a prompt. That is why Order status can answer truthfully and a hand-written instruction cannot: the skill grants the lookup.

The Skills tab

The tab opens with a count — 2 of 15 on. — and then the skills in six groups, by job rather than by integration. Each group has a heading, one line on what it covers, and its own count, such as 1/3 on. Each skill is a row with its name, a one-line description and a switch. A row that is on carries a tinted left edge. A skill that only works with something connected wears a badge saying so: The badge is a label, not a check — it does not say whether your workspace has the connection. The readiness checklist at the top of the studio is what checks. While a skill is on, a Settings link (or What it can do, for a skill with no settings) opens it: the tools it can call, listed as Can call:, and its settings. Save the tab with Save skills; Unsaved changes appears beside the button until you do. Two skills are on for a new agent: FAQ & support and Human handoff. Neither calls anything. Everything that reaches outside Watx is off until you switch it on and set it up.

The catalogue

Answering & support

What it says when somebody writes in.

Leads & CRM

What it records about the people it talks to.

Selling

Looking things up, recommending, and taking money.

Appointments

Offering times and booking them. Appointments carries no badge on purpose: without Google it still works, by sharing your booking link.

Messaging & automation

Reaching for the rest of the product mid-conversation.

Connected apps

Writing into tools you already use.
A skill whose integration is missing is not an error — the agent simply never sees that tool and answers without it. The readiness checklist is what tells you it is happening.

Pickers, not text boxes

A setting with a fixed set of answers is a picker. A pipeline id, a template name or a Cal.com event type cannot be typed correctly from memory, and a wrong one would save, look right and fail only when the agent tried to use it.
  • The choices are your workspace’s own — your pipelines and their stages, your approved WhatsApp templates, your published booking forms, your Cal.com event types — and they are read live when the tab opens.
  • When there is nothing to pick from, the picker is replaced by a sentence saying which screen fixes it, such as No pipelines yet. Create one under Pipelines and the agent will have somewhere to file a deal.
  • A saved value your workspace no longer offers is kept, not dropped. A deleted pipeline, a template that was removed, a form that was unpublished — it stays selected and is labelled Saved earlier — no longer in this workspace, so opening the tab and saving never quietly changes a setting. Choose something else, or remove it.
  • A field that depends on another is cleared when the other changes. Change the pipeline and the stage is cleared, because a stage belongs to one pipeline.
  • A few pickers also take your own value. Where any answer is legitimate — a booking link on another site, an unusual return window, a detail you want collected — the list ends with Something else… or an Or add your own… box. Where the agent can only use what is on the list, there is no such option.

Settings, skill by skill

Details to collect — pick from Budget, Timeline, Company name, Team size, What they need it for, Quantity, Location, Email address, Who decides and What they use today, or type your own in Or add your own…. Up to eight. The agent asks at most two per message, conversationally, never as a numbered form.Treat as qualified when — free text, e.g. they have a budget over ₹50,000 and a timeline this quarter.Once it has the details it records them against the contact once, then confirms that somebody will follow up.
Return window — 7, 10, 14, 15, 21, 30, 45, 60 or 90 days, or Something else… to type an unusual number.Policy notes — the conditions, e.g. unopened items only, original packaging, customer pays return shipping.The agent explains the policy and never approves a refund, waives a fee or commits to a timeline itself. Anything outside the policy is handed to a person.
Pipeline for new deals — one of your pipelines. Unset uses the workspace’s first pipeline.Stage it starts in — the stages of the pipeline above, each shown with its pipeline’s name. Unset uses that pipeline’s first stage. Changing the pipeline clears this field, and a stage from another pipeline is ignored at run time.Create a deal when at least this confident — from 0.5 (anything that sounds like interest) to 0.9 (only when they have all but asked). 0.8 by default.If the customer changes their budget or requirements later in the same conversation, the agent updates the deal that conversation already has rather than opening a second one.
How to assign deals — free text, e.g. assign to the sales director for deals over 50k, to team leads otherwise.The agent names a teammate the way you would out loud (“Priya”, or their email) and Watx resolves who that is. It does not discuss your team structure with the customer.
No settings. When the customer has said what they want and how many, the agent sends a link that opens their basket on your store with that product and quantity. The store collects the delivery address, works out shipping and takes the payment, so the agent never asks for an address or card details and never quotes a delivery charge.It never chooses for the customer:
  • Several products match — it asks which one they mean.
  • The product comes in sizes, colours or other options — it lists them and asks which one. An answer that matches more than one option (“Large” against “Large” and “Extra Large”) gets the question again, and an option the product does not have is said to be unavailable, with the real options.
  • Every option is sold out — it says the product is out of stock rather than sending a link.
  • Nothing matches — it says so and offers a colleague, rather than suggesting something it invented.
When the catalogue shows the chosen option as out of stock, the agent says so before the customer pays. A product the store cannot link straight to a basket gets its product page instead. The tool reads the catalogue Watx synced from the store, so sync after you change products.
Most it may ever ask for — required. The agent refuses any amount above this and hands over instead. The ceiling is enforced by the tool, not by the instructions: a customer can talk a model into a number, so this is the part that is not up for discussion. Leave it unset and the tool refuses every time.What things cost — the price list it may quote, one per line, e.g. Consultation — 1500. Without it the agent has nothing to work from.It confirms what is being bought and the amount before sending a link, sends one link per agreed purchase, and never claims a payment has been received — it cannot see that.
Works with Shopify, WooCommerce and a custom store. The agent chooses the percentage; Watx makes up the code, creates it at your store, and only then hands it to the agent to send.Most it may ever give — required: 5% to 50%. The agent refuses anything higher and offers a colleague instead. The limit is enforced by the tool, not by the instructions. Leave it unset and the tool refuses every time.Code expires after — 1, 2, 3, 7, 14 or 30 days. 7 if unset.When to offer a code — e.g. Only for a first order, or when a delivery was late. Without it the agent offers one only when the customer asks for a discount.Minimum order — optional. The code works only on orders of at least this much, in your store’s currency.Code prefix — letters and numbers, up to 10. Codes look like WA-7K3Q9P. The rest is random, so a code cannot be guessed.Store that issues codes — only needed with more than one store connected. A store that cannot issue codes is listed with the reason.Each code works once, for one customer. Asked again, the agent gives the customer the code they already have rather than a new one. Every code appears on the store console’s Discount codes tab with who got it, which agent gave it, the reason it gave, and whether it was used.What each store needs:
  • Shopify — Watx’s access must include the write_discounts permission. A store connected by installing the Watx app before this permission was added grants it through More actions → Reinstall the app on the store console. A store connected with a pasted token: add write_discounts to your custom app’s Admin API scopes in Shopify, then choose More actions → Re-register webhooks.
  • WooCommerce — an API key with Read/Write permission. A read-only key is refused by the store, and the code is listed as Refused by store.
  • Custom store — a callback URL that creates the code when asked.
Booking link — your published booking forms are listed; choose Something else… to paste any other link (Calendly, a page on your site). Used when Google is not connected, and as a fallback the agent can always offer.Availability in words — e.g. Mon–Fri 10:00–18:00 IST, closed on public holidays. This matters even with Google connected: a calendar says what is busy, never what is open. Without it the agent will happily offer 3am, because nothing is booked then.Meeting length — 15, 20, 30, 45, 60, 90 or 120 minutes. 30 if unset.With Google connected, the agent checks availability before naming a time and books only after the customer has agreed to a specific one. Without it, it never states that a slot is confirmed.
What it books — one of your Cal.com event types. Each shows its length, and a note when it is hidden on your Cal.com profile or its booking page does not ask for a phone. Bookings are made on this event type, with its length, location and confirmation rules from Cal.com. With Cal.com not connected, the picker says Connect Cal.com in Integrations — its event types appear here.When to offer a booking — optional, e.g. When someone asks for a demo, a consultation or a call back. Without it the agent offers to book whenever a customer asks to meet or talk.How it works in a conversation:
  • It looks up open times — up to 14 days ahead, a week by default — and offers two or three of them, in the customer’s time zone when it knows it. It only offers times Cal.com returned.
  • It books only after the customer has agreed to one specific time. It needs the customer’s name, and asks for an email only if they want the calendar invite by email.
  • The customer’s phone number comes from the contact in this conversation, never from the chat.
  • After booking it gives the time, the meeting link when there is one, and Cal.com’s link to reschedule or cancel. For an event type that needs your confirmation, it says the booking is requested and not confirmed yet.
  • It never cancels or reschedules anything itself.
The booking fires the same appointment triggers as any other Cal.com booking. See Cal.com.
Templates the agent may send — a picker over your WhatsApp templates, up to 20. Nothing selected means the agent sends none, never “all of them”. There is no way to type a name: a name that does not match a sendable template would simply never be used.Sendable templates come first, each with its languages and the start of its approved wording. Templates that cannot be sent safely are listed after them, greyed out, as Cannot be sent — reason:A template with several languages is judged per language. If one language version cannot be sent, the template stays choosable but its line says which version and why — fix or remove that version, or the agent may fail to send it.The agent sees each allowed template’s approved wording, so it can pick one that genuinely fits. It cannot edit a word of it, and when none of them says what it means it writes an ordinary reply. A template it sends is the reply: it sends no text after it, and the template counts towards the conversation’s reply limit. It works only in WhatsApp conversations and never while drafting. See Message templates.
Automations this agent may trigger — pick from: Send an email, Send an SMS, Update the contact, Create a task, Notify the team, Add a tag, Add to a segment, Schedule a follow-up. Nothing selected means any of them; these eight are the only types the tool accepts.
Today this skill only records the request. When the agent uses it, Watx adds an internal note to the contact — [AI Automation] Triggered: notify_team, for example — and nothing else happens: no email or SMS is sent, nobody is notified, and none of your automations runs. To act on what customers say, build an automation on a Keyword match trigger instead.
It needs a customer attached to the conversation, so it does nothing in the test drawer.
Spreadsheet id — from the sheet’s URL, between /d/ and /edit. The agent can only append to this one sheet; it cannot read or change what is already there, and it cannot be talked into writing somewhere else.Tab name — blank uses the first tab. A name that does not exist fails every write.Columns, left to right — e.g. Date, Name, Phone, What they wanted. Without it the agent guesses what goes in each cell.These three are typed, because nothing in Watx lists your Google files. The agent logs a row or saves a contact once per conversation, not per message, and raises a task only when something genuinely needs a person later. None of it is narrated to the customer.
FAQ & support, Human handoff, Order status, Product recommendations and Sell from the chat have no settings.

When a tool is withheld

A tool is removed from the agent’s toolset rather than offered and allowed to fail. A failing tool costs a round trip and a credit to discover something that was knowable before the reply started — and, in front of a customer, an agent that tries and fails reads worse than one that never offered. request_payment with no Most it may ever ask for set is not withheld: it is offered and refuses every request, so the agent hands over instead. create_coupon with no Most it may ever give behaves the same way. An agent takes at most three tool rounds per reply — enough for “look up the order, then search the catalogue, then answer”.

Adding a skill to an existing agent

Skills are per agent. Switching Take payments on for your sales agent leaves your support agent alone. If you want the same set everywhere, set one agent up and duplicate it.