AI Studio

Skills

Skills are how you give your AI agent abilities beyond answering from its knowledge. Each skill bundles a short description of when it applies with a set of tools the agent can use, such as searching your content, sending an email, or taking an action in a connected app.

You manage skills in AI Studio, inside the dashboard. AI Studio has two sides you switch between: Build, where you see and shape what your agent can do, and Test, where you talk to your live agent like a visitor. The most important idea on this page is also the simplest one: a skill is a capability you hand to the agent, not a fixed script. You describe what the skill is for and which tools it can use, and the agent decides for itself, turn by turn, whether the skill is relevant and which tool to reach for. There is no flowchart to draw and no rigid order of questions to maintain.

You build skills by instructing the build agent. On the Build side you describe what you want in plain language, and the build agent creates or updates the skill for you. There is no manual form to fill in. Every change, whether you start from scratch, pick a template, or refine an existing skill, flows through the same build conversation.

Skills build on Apps. Several tools take real action through services you connect, such as Calendly or HubSpot. Connect those first on the Apps page, then any skill can use them.

What a skill is

A skill is made of three plain-language parts. There are no input boxes to wire up and no steps to sequence. The agent fills in the conversational reality on its own.

A name. A short label such as Book a meeting or Members area access. It shows on the skill card and helps you keep skills organized.
A when-to-use description. One or two sentences describing the situation the skill is for, for example The visitor wants to book a demo or a call. This is the hint the agent uses to recognise when the skill applies. It is a natural description, not a list of keywords.
Instructions and tools. Optional free-text guidance on how to behave, plus the set of tools the skill can call. The tools are the actual capabilities, and any values a tool needs (an email address, a meeting type) are gathered by the agent as it talks to the visitor.

Inputs are part of a tool, not a separate step. The agent asks the visitor for a value because a tool needs it, and it only asks for what is still missing. If the visitor volunteers several details at once, or out of order, the agent picks them all up.

The tools a skill can use

A skill is built from tools, and each tool is one of a few types. When the build agent adds a tool, it picks the type that fits what you asked for. Each type gives the agent a different kind of ability:

Tool	What it lets the agent do
Knowledge search	Search this agent's knowledge base. You can scope it to certain sources and pick smart (meaning-based) search, exact keyword match, or both. The agent writes the search query live from the conversation, so it matches what the visitor actually asked.
API call	Call an external API you configure (method, URL, headers) to fetch data or take an action in your own system. Any secret is stored by reference and never shown to the agent.
Validator	Check a value the visitor gives, such as a password or access code, against a list you set. On a match it can unlock another tool. It is an exact match only, not a format check.
Send email	Send an email from build@ultimo-bots.com to one or more recipients you set, with a custom subject and body. Useful for owner notifications and lead alerts. Recipient counts are capped, so it is not a mailing tool.
App action	Use a connected app to do something real for the visitor, such as create a booking link or add a newsletter subscriber. You pick the app and the action; the available actions come from the app itself.
Decision guide	Give the agent a branching conversation map for triage, troubleshooting, or intake: a set of questions and where each answer leads. The agent follows it as advice, not as a locked path, so it can still answer side questions and adapt. For a rule that must always hold, use a Validator instead.
Lookup table	Map an exact key to a precise value, such as a product code to its price or a city to a shipping link. The match is exact and the value is returned as written, so the agent reads back the right answer instead of guessing it. It is the dependable counterpart to Knowledge search and the Decision guide.

Decision guide and Lookup table sit at two ends of a range. A Decision guide shapes a conversation the agent runs flexibly, which suits triage or a guided troubleshooter. A Lookup table is exact and predictable, which suits facts that must be right every time, like prices or links. Reach for the one that matches how much the answer is allowed to vary.

Built-in capabilities

Some abilities are available without building a skill yourself, because they are wired into the product:

Human handover. When you turn on live chat, the agent can hand a conversation to a person on your team. The agent recognises when a visitor wants a human, flips the conversation to a live request, and notifies you so an operator can step in. You switch this on in Live chat settings, where you also choose whether visitors see a button to ask for a human directly.
Lead capture. Any skill action can be set to mark the conversation as a lead when it runs. The contact details the agent gathered are saved as the lead, and your new-lead notifications fire. This turns a normal skill, like booking a call or sending an enquiry email, into a lead at the same moment it does its job.

Handover is part of Live Chat. The agent can only bring in a human if live chat is enabled for the agent. If it is off, the agent answers on its own and there is no one to hand the conversation to. See Live chat settings.

Create a skill

You create a skill by telling the build agent what you want, not by filling in a form. On the Build side of AI Studio, type what the skill should do into the message box at the bottom, for example Let visitors book a demo and email me when they do. The build agent assembles the skill, replies in the conversation, and shows you what it made.

Describe the skill
In Build mode, describe the skill in plain language, the way you would explain it to a colleague. Be clear about when it should apply and what it should do. You can also open the Add skill menu and choose Create with AI to start from a short description, or pick a template (see the next section).
Watch it appear on the blueprint
The build agent creates the skill and shows a read-only blueprint of it beneath its reply, along with a new skill card. The blueprint is a faithful picture of the tools the skill uses and how they connect. You do not edit it by hand.
Refine it by instructing the agent
Keep talking to the build agent to change anything: rename a tool, add a step, tighten when it applies. You can also open a skill card's menu and choose Adjust with AI to describe a change. Editing always happens through the conversation, so you never lose track of what changed.
Try it, then set it live
Switch to the Test side and talk to your agent as a visitor would. The test uses your real knowledge and real app actions, but nothing is saved and nothing is billed. When you are happy, set the skill card to Live.

Need a value before something else can happen? Ask the build agent to gate a tool behind a check, for example a Validator that confirms a password before a Knowledge search that holds members-only content runs. The lock is enforced behind the scenes, so the protected tool cannot run until the check passes.

Start from a template

You do not have to start from a blank message. AI Studio offers ready-made starting points you can pick and then refine by talking to the build agent.

Skill templates. Common patterns, such as capturing leads, answering frequently asked questions, booking a meeting, or qualifying a visitor. Pick one and the build conversation starts already pointed at that goal.
Per-app starters. For an app you have connected, AI Studio lists ready-made skills built around it, for example a Calendly booking skill or a Mailchimp newsletter signup. Choose the ones you want and the build agent assembles a skill from them.

A template is only a head start. The build agent still assembles the real skill in the conversation, so you can adjust anything before you set it live.

The read-only blueprint

Every skill has a blueprint: a clear, read-only map of the tools it uses. You see it under the build agent's reply right after a change, and again when you open a skill card to view its details. Click a tool on the map to see plain-language facts about it, such as what it asks the visitor for, whether it needs verification, and whether it captures a lead.

The lines on the blueprint are not a hand-drawn sequence. They are derived from the rules the agent actually enforces: a tool that is locked until a check passes, or an action that needs another action's result before it can run. So the map reflects real order, not a suggested one. To change the blueprint, you instruct the build agent. There are no nodes to drag.

The blueprint is for understanding, not editing. It is there so you can see at a glance what a skill does and audit any gates or verification steps. All changes go through the build conversation.

Manage your skills

Your skills sit as cards on the Build home. Each card shows the skill name, a Live or Off pill, and a three-dots menu. The pill is the quickest control: a skill only affects your live agent when it is set to Live. Turn it Off to take a skill out of play without deleting it, for example while you are still building it or testing a connected app. The change applies to the agent on your site right away.

Open a card's three-dots menu for the rest of the actions:

Details opens the read-only view of the skill, with its plain-language summary and the blueprint.
Rename changes the skill's name.
Adjust with AI opens a short box where you describe a change, then hands it to the build agent. This is how you edit a skill.
Delete removes the skill for good. Deleting is permanent, so set a skill to Off first if you only want to pause it.

How the agent decides to use a skill

This is the heart of how skills work, and it is worth understanding clearly. Your agent runs as an agentic loop: on every turn it reads the conversation, weighs the skills you have given it, and decides what to do next. It is not following a flowchart and it is not locked into a fixed sequence of questions.

In practice that means:

The agent picks the skill. It compares what the visitor is asking against each skill's when-to-use description and brings in the relevant one on its own. You do not route messages to skills by hand.
The agent chooses the tools and the order. Within a skill, it decides which tool to call and when, gathering any values it needs from the conversation along the way.
Going off-script is handled for free. If a visitor interrupts a booking to ask about pricing, the agent answers the pricing question, then naturally returns to the booking. There is no brittle state to lose track of, because the agent re-reads the conversation each turn rather than following a saved step.

What you control versus what the agent decides. You declare the structure: the skill's purpose, its tools, and any hard rules (a value that must be unlocked first, an action that needs another action's result). The agent handles the conversation: when to ask which question, how to phrase it, whether the visitor already answered, and when the skill is done.

Apps and real actions

App actions are what let a skill do something in the outside world rather than just talk. You connect an app once on the Apps page using a secure sign-in, and from then on any skill can use that app's actions. The app acts under your own connected account, so a booking lands in your calendar and a contact lands in your CRM.

The apps you can connect today are:

Calendly and Cal.com for scheduling and bookings.
Mailchimp for newsletter audiences and subscribers.
HubSpot for contacts.
Stripe for products, payments, and customer self-service.
Zendesk for creating and looking up support tickets.

When an action returns data, the agent sees a clean, summarised version of the result rather than the raw response. This keeps it from inventing fields, links, or options that the app did not actually return.

Actions on private data require visitor verification. Some actions touch a specific person's private data, such as reading invoices, cancelling a subscription, or listing someone's bookings or tickets. Before any of these run, the visitor must prove control of their email with a one-time code sent to them, and the action is then bound to that verified email so a visitor can only ever act on their own data. This protection is automatic for those actions, and you do not configure it.

Skills guide the agent, they do not script it. A skill makes a capability available and biases the agent toward using it in the right situation. It is not a deterministic flow, so the exact wording and the order of soft questions can vary between conversations. The hard rules you set (a required unlock, a verification step, an action that needs another action's result) are always enforced; everything else is the agent's judgement.

Next steps

Round out your agent's abilities and decide who steps in when:

Connect appsLink Calendly, HubSpot, Stripe, and more so your skills can take real action.Shape behaviorSet the agent's persona, tone, and ground rules for how it responds.Set up live chatTurn on human handover so your team can take over a conversation when needed.