What Is Chatbase? A Practical Guide to Its AI Customer Support Platform

Chatbase is a hosted platform for creating AI agents for customer support and customer experience. A team can give an agent business knowledge, define its behaviour, connect actions and channels, test it, and route conversations to people when automation should stop.

That makes Chatbase more than an FAQ widget. Its current product surface includes website deployment, integrations, an internal helpdesk, APIs, actions, procedures, and analytics. It is not a finished support operation by itself: accurate answers still depend on maintained sources, clear boundaries, safe permissions, testing, and human ownership.

This guide explains how the pieces fit together and what to validate before choosing it.

How a Chatbase agent is built

The basic setup is approachable. Chatbase's first-agent guide starts with a new agent, adds business information, tests responses, adjusts instructions and model settings, and deploys the result to a website.

The current dashboard organizes the main configuration under Build. According to the Build documentation, instructions define the agent's role and response style, guardrails hold rules that should always apply, and model configuration controls the selected model and temperature. A Compare view can send the same test message to different configurations side by side.

That creates a useful separation:

• Knowledge supplies facts the agent can reference.
• Instructions define its purpose, tone, and expected behaviour.
• Guardrails set non-negotiable boundaries and fallback rules.
• Actions let it retrieve data, complete tasks, or hand off.
• Procedures give important multi-step jobs a more controlled path.
• Channels determine where customers can reach it.

A dependable deployment takes longer than the initial setup. Teams must decide which sources are authoritative, which requests can be automated, and how failures will be reviewed.

Knowledge sources and answer quality

Chatbase's data-source documentation lists files, websites, text snippets, question-and-answer pairs, Notion pages, and tickets from supported helpdesk connections. Sources can be reviewed and retrained after changes; automatic retraining is plan-dependent.

This flexibility combines documentation, policies, and curated answers, but does not resolve contradictions automatically. If a return policy says 14 days in one document and 30 days on the website, the source of truth is ambiguous.

A practical content process should therefore:

1. include only customer-safe material;
2. remove outdated and duplicate policy versions;
3. assign an owner to each source;
4. define what the agent should do when evidence is missing;
5. retrain and retest after meaningful updates.

Treat “training” as an operational content workflow, not a one-time upload. Tests should cover paraphrases, incomplete requests, policy exceptions, conflicts, and questions the material cannot answer. Our guide to training AI support on knowledge and brand voice includes a deeper checklist.

Instructions, guardrails, and model choice

Instructions can define the business context, the agent's job, response style, goals, and extra guidance. Guardrails are intended for rules such as topic restrictions, sensitive-information handling, safety requirements, and fallback behaviour.

Put durable facts in sources rather than a long prompt. Use instructions for behaviour: ask for an order number before a lookup, never promise an unapproved refund, cite policy, or offer a person when confidence is low. Put hard prohibitions and privacy rules in guardrails, then test adversarial wording.

Chatbase currently offers model selection and charges different message-credit amounts for different models. Avoid choosing from a single polished example. Use the Compare view with a fixed evaluation set and score correctness, appropriate refusal, action choice, latency, tone, and cost. A faster or cheaper model may be better for a bounded support job; a more capable model may be justified when the requests and tool decisions are harder.

Actions turn answers into outcomes

The AI Actions overview includes support escalation, Chatbase live chat, commerce tasks, lead and data collection, scheduling, Slack notifications, web search, and custom actions that can call an API or run client-side code. Availability and behaviour vary by action, integration, channel, and plan.

Actions are where a conversational answer can become a business event: look up an order, book a meeting, collect a field, create a ticket, or call an internal endpoint. They also increase risk. A wrong policy answer is harmful; a wrong cancellation or refund is both harmful and transactional.

Chatbase provides an Only use in procedures setting for actions that should not run on the model's initiative. Its procedure runtime documentation says a procedure can call designated actions in a defined flow, while procedure-only actions stay unavailable outside that flow. This is valuable for sensitive work that needs prerequisites, explicit confirmation, or a deliberate sequence.

Start important actions read-only where possible. For write actions, validate identifiers, confirm the change, log the result, and design recovery. Test every intended channel because a procedure is skipped when one referenced action is unavailable there.

Channels, website deployment, and APIs

For websites, Chatbase supports a chat bubble and other deployment formats. Its developer overview describes a JavaScript embed for custom events and browser-side actions, a REST API for custom interfaces and server integrations, and API v2 for programmatic agent and data management.

Official guides also cover WhatsApp, Instagram, Messenger, Slack, email, phone, Shopify, and helpdesks. Each has provider-specific requirements. Instagram requires a professional account connected to a Facebook Page; WhatsApp requires a suitable business number and Meta setup.

Do not treat a channel name as proof of complete parity. Confirm the exact message types, attachments, identity model, actions, escalation path, rate limits, and provider rules you need. Test a deployed channel rather than relying only on the playground.

Human handoff and helpdesk work

Chatbase supports two related handoff models. The Escalations action can create a ticket in Chatbase or a connected helpdesk such as Zendesk, Salesforce, Intercom, Zoho Desk, Freshdesk, HubSpot, Help Scout, or Gorgias. Teams define when the action should run and on which channels it is available.

Chatbase's own Helpdesk also supports direct takeover. The takeover documentation says a support associate can stop AI replies in an ongoing conversation, create a linked ticket with an AI-generated subject and summary, and continue with the customer. It documents takeover for the chat bubble, email, WhatsApp, Messenger, and Instagram, plus a way to return the conversation to AI replies later.

Test the full handoff, not just ticket creation. Check notification, context, customer identity, ownership, and the waiting experience. Include explicit requests for a person, unknown answers, sensitive issues, repeated failures, and urgent language.

Analytics and the improvement loop

Chatbase's analytics documentation covers chat and message volume, feedback, active users, channels, action calls, topics, sentiment, and Helpdesk measures such as ticket volume and median response times. Some data and features depend on the plan, and the documentation notes that analytics update with a delay.

Dashboards show where to investigate, but volume is not resolution. Pair analytics with weekly sample review. Measure correctness, unsupported claims, successful actions, appropriate escalation, repeat contacts, and useful handoff context. Track cost per successful outcome rather than only fewer human tickets.

How Chatbase pricing works

Chatbase's live pricing page uses subscription tiers with included message credits and feature limits. Requests to the selected AI model consume credits, and an answer that requires actions can involve more than one model request. The page also presents add-ons and plan differences for items such as agent count, actions, workspace seats, retraining, and branding.

Because prices and allowances can change, model a representative month from the live page instead of copying an old quote. Include traffic, model choice, multi-step actions, agents, teammates, knowledge maintenance, integrations, human review, and peak usage. Then compare the total with the value of correctly resolved demand.

Strengths, tradeoffs, and likely fit

Chatbase's main strength is that one hosted product combines knowledge-grounded answers, configurable behaviour, actions, controlled procedures, multiple channels, human takeover, external helpdesk connections, APIs, and analytics. It can fit teams that want more operational depth than a simple embedded FAQ bot without starting from a developer framework.

The tradeoff is still operational complexity. More sources can conflict, more actions create permission and failure risks, and more channels create testing work. Helpdesk and analytics features do not remove the need for source owners, quality review, escalation staffing, and change control.

Fit should follow the dominant support job. A company that wants a configurable agent across a website, messaging channels, actions, and helpdesk workflows has a clear reason to evaluate Chatbase. A team whose narrower priority is answering public social comments and DMs from shared business knowledge might also compare a social-first product such as Luni Chat. That is a difference in focus, not a claim that either product replaces the other in every environment.

A focused proof-of-concept checklist

1. Choose one frequent support job. Avoid testing every department at once.
2. Prepare clean source material. Resolve conflicts before the agent sees them.
3. Define instructions and guardrails. State the role, tone, boundaries, and fallback.
4. Connect one real channel. Include its identity, attachment, and provider constraints.
5. Add one useful action. Start read-only or add confirmation and recovery.
6. Build handoff first. Test explicit and automatic escalation with real staffing.
7. Run a fixed test set. Include ambiguity, missing facts, exceptions, unsafe requests, and broken dependencies.
8. Measure outcomes. Score correct resolution, action accuracy, escalation quality, latency, and cost.
9. Name the operator. Assign weekly ownership for sources, failures, and changes.

The bottom line

Chatbase is a hosted AI customer-support platform, not merely a chat-bubble generator. Its current scope connects business knowledge, configurable agent behaviour, actions and procedures, website and messaging channels, human support, APIs, and analytics.

That breadth makes it worth evaluating when a team wants to build a tailored support agent without assembling the entire platform itself. The responsible buying test is an end-to-end proof of concept: one real channel, one real action, a clean knowledge set, difficult questions, failed dependencies, human takeover, measurable outcomes, and a named owner after launch.