The AI agent & Knowledge Hub

The feature is two things that work as a pair. The AI agent is the part that talks: it reads a customer’s message and replies in your voice. The Knowledge Hub is the part it looks things up in: a shared library of your facts. The agent doesn’t have to invent an answer; with its Search Knowledge Base action on (it is on by default) it searches the Hub first and replies from what it finds.

Both live one level up from the workspace you work in every day, in your brand’s Admin Dashboard, under a group called AI & Knowledge: AI Agents for the agents, Knowledge Hub for the library, plus Internal Input (Beta) and AI Feedback for the human side. One Knowledge Hub is shared across every agent and every workspace in the brand, so a fact you fix once is fixed for all of them.

The most common first stumble isn’t setting an agent up, it’s finding the screen. If you’ve been hunting for the Knowledge Hub in your workspace sidebar, between Inbox and Workflows, it isn’t there and nothing is broken. Your workspace is one outlet’s day-to-day work; the Admin Dashboard is the brand-wide floor above it, and that’s where AI agents and the Knowledge Hub live. Two steps, from anywhere in the app.

You can stop after three steps and have a real, useful agent. Here’s the whole minimal path before we go deeper.

Start with knowledge, not the agent: an agent with nothing to look up can only fall back on generic answers. The Knowledge Hub has four tabs, each with a count beside its name:

• FAQs: short question-and-answer pairs grouped into folders (Pricing, Schedules, Policies). This is the backbone; it’s where most answers should come from.
• Websites: point us at a site and we crawl and index it, keep it linked to the live site, and you refresh it in one click.
• Documents: PDF, TXT and DOCX files. Read the warning below before you rely on these for the agent.
• Suggestions: questions customers keep asking that the Hub can’t answer yet, waiting for you to approve them as FAQs.

When a message arrives the agent doesn’t free-associate. It calls its Search Knowledge Base action, pulls the best-matching FAQs and website pages, then writes its reply from them. The customer just sees an answer:

You, the team, see more. Under the AI’s reply in the Inbox a small pill counts the sources it drew on.

If an answer looks wrong, open its sources and click the FAQ row: it opens that FAQ in the Knowledge Hub so you can fix it, and every future reply uses the corrected version.

Click an agent’s card to open its editor. It opens full screen with no sidebar (that is deliberate) and a tab for each part of how the agent behaves.

The tabs, in plain terms:

• Instructions: what the agent is and how it behaves. This is the single biggest lever on reply quality. Write is where you type; Preview shows what the agent will actually read. Press / to drop in a reference to a pipeline, stage, field or exit reason, and use the Arguments bar for named values a workflow can fill in.
• Actions: the things the agent may do (next section).
• Required Resources: the pipelines, stages and fields this agent needs to exist in any workspace it runs in. Whatever you declare here becomes something you can reference with /.
• Exit Conditions: the ways a conversation can end, and the reason it reports back.
• Output Data: structured fields you want the agent to return to the workflow when it ends, for later steps to use.
• Tests and Feedback: replay saved conversations after a change, and review flagged replies (a red number shows how many are waiting).
• Settings: conversation limits.

Actions are the capabilities an agent can use mid-conversation. They come in four groups: Data Access (read-only: Search Knowledge Base, Get Pipelines, Get Leads, Get Custom Fields; on by default), Write Actions (Update Contact, Create Lead, Update Lead), Communication (Request Internal Input) and Calendar (List Services, Find Available Times, Get Contact’s Appointments, Book a Service, Reschedule and Cancel Appointment). Tick a write action and the reads it depends on switch on and lock, labelled “Required by …”.

The Actions tab only says what the agent is permitted to touch. When to use an action, and how, is written once, in the Instructions. So a sales agent and a support agent can read different shelves of the same library: point each one at the folders and websites it should search.

Turn on Request Internal Input (Actions, under Communication) and the agent can pause and ask your team when it can’t answer. The customer waits, a rep answers, and the agent carries on. Its panel also has an optional Holding message: a note sent to the customer once, the moment the agent pauses, so they are not left in silence (the default text is “Thanks for your message! Let me check on this and get back to you shortly.”). It is off unless you switch it on.

For the rep it looks like this. In the conversation list the chat gets an amber robot, the chat shows Agent is waiting for your input, and the reply box is replaced by an Awaiting Your Input panel.

You don’t have to be in the conversation to answer. Every waiting request in the brand is also listed under Internal Input (Beta) in the Admin Dashboard: click Internal Input, the third row of AI & Knowledge.

Every agent turn ends with an exit reason, returned as exitReason so the next workflow step can branch on it. Three system defaults are always there and can’t be removed. Of those, the agent can only choose goal_achieved; the runtime writes timeout and max_turns_exceeded itself.

Add custom exit conditions for the outcomes that matter to your business, such as trial_booked, not_interested or needs_human. Keys are lowercase with underscores. Generate with AI needs instructions first, since it reads them to decide what the exits should be. A workflow can also see error, taken_over_by_human (a teammate took over) and agent_not_published, which the agent never picks itself.

Two limits in Settings stop a conversation running forever.

An agent does its real work inside a workflow. Add an AI Agent action (in the AI group), pick your published agent and a channel, then branch on how it ended.

After the node, three outputs are available to later steps: the agent’s Agent Status, its Exit Reason and its Exit Message, plus every field you defined under Output Data. Exit Reason is the join between the agent and the rest of your automation.

Anyone reading a conversation can rate an AI reply. A thumbs-up (Mark as good AI response) saves at once; a thumbs-down (Flag AI response - needs improvement) opens a short flow that captures what should have happened.

Four configurations, simplest first. Each is built from pieces above: adapt them, don’t copy exactly.

• It’s not in your workspace sidebar. The Knowledge Hub and AI Agents live in the brand’s Admin Dashboard: brand tile, top-left, then Admin Dashboard, then AI & Knowledge.
• Knowledge before agent. An agent is only as good as the Hub behind it. Seed your top FAQs first: pricing, schedules, location, policies.
• Publish to use. Workflows only see Published agents. A Draft is invisible in the AI Agent action’s dropdown.
• Documents are not searched by the agent. It reads FAQs and websites. Uploaded documents feed the Retrieve Knowledge workflow action instead.
• Only text blocks are searched. An FAQ’s images, videos and documents are sent to the customer as they are; the facts belong in a Text block.
• Fix facts in the Hub, behaviour in the prompt. Wrong figure: edit the FAQ. Wrong tone or structure: flag it and use Review & improve.
• There’s no built-in “human handoff” exit. Use a custom exit condition you branch on, or Request Internal Input. They’re different tools.
• Scope the knowledge source per agent. Point each agent at the folders and websites it should read, so a sales bot doesn’t quote support-only policy.
• Improving an agent is an admin decision. A thumbs-down queues the item; nothing changes until someone reviews the fix and clicks Apply the fix.