Understanding Agents
What an agent includes, where to teach it, and how to check it before customers hear it
Understanding Agents#
An agent is the customer-facing AI for one business. It answers from the sources you teach it, and it can use tools you turn on. It does not go live just because setup is finished.
Where to work#
| Section | What it is for |
|---|---|
| Playground | A private customer conversation. Nothing here is sent to customers |
| Agent Studio | Guidance, Data sources, Tools, Procedures, Workflows, Escalations, Test, and Source Health |
| Deploy | Channel setup. Confirm live replies separately from connecting an account |
| Activity | Real customer conversations after launch, including human takeover |
| CRM | People and companies. A conversation is not a person |
| Outbound | Approved broadcasts, templates, automations, and contact links |
| Growth | Content plans, social posts, brand assets, and paused ad preparation |
| Analytics | What customers asked, how replies covered them, and what still needs attention |
| AI Assistant | Your private assistant for setup, review, and approvals |
| Settings | Name, availability, calendar, team alerts, connections, and webhooks |
Outbound, Growth, and Tickets depend on project feature access and permissions. See Growth for its setup and publishing boundaries.
What the agent uses#
| Piece | Where you set it | What it changes |
|---|---|---|
| Business facts | Agent Studio → Data sources | What the agent is allowed to say |
| Voice and model | Agent Studio → Guidance | Tone, language, and the model shown with its credit cost |
| Website greeting | Deploy → Chat widget or Chat bar → Content | The initial message and suggested questions |
| Actions | Tools, Procedures, and Workflows | Checks and writes, such as a booking or a form |
| Live replies | Deploy | Whether customers can receive those replies. For website chat, the switch is AI Auto-Reply |
Response creativity is in Guidance. Leave it on Automatic unless you have a reason to set a custom level. A lower custom level stays closer to the same factual answer.
How an answer is chosen#
The agent looks through the knowledge you prepared and answers from what it finds. It should not invent prices, hours, or policies that are not there.
Uploading a file stores a source. Training, when the source type requires it, is what makes that content available for answers. Those are different steps. Training your agent explains the status you should see. How retrieval works explains how to improve a weak answer without extra settings.
Use the source type that matches the fact:
- Files, text, Q&A, websites, and Google Drive for documents and approved answers
- Structured data for rows such as a menu, catalog, or price list
- Time-sensitive content for promos, closures, and dated facts
- Source Health for gaps, conflicts, and duplicates
A booking policy in a document does not connect a calendar. Turn on the booking tool and test the saved appointment.
Before customers can hear it#
- Name the business and add current facts.
- Train sources that are still waiting.
- Ask a real question in Playground.
- Save the important questions in Agent Studio → Test so you can run them again.
- Open Deploy and confirm a live channel only after those checks.
A connected page, a finished setup screen, or a private reply you liked is not a launch.
Practices that hold up#
- One agent with complete knowledge is easier to keep consistent than several agents for the same job.
- Prefer Q&A for answers that must stay exact, and include the ways customers actually phrase the question.
- Retest after you change a source, the guidance, or a tool.
- Read Activity → Inbox after launch for questions the sources do not cover.
- Check Settings → Plans for agent, storage, and credit limits. Do not rely on an old price list.
Common mistakes#
- Treating an upload as trained knowledge
- Testing only in a customer channel
- Raising creativity for factual answers
- Writing vague guidance such as "be helpful" with no fallback when the answer is missing
- Expecting CRM to score every chat as a sales lead. CRM stores people; conversations stay in Activity