Chatbase

Breadth with authenticated write action and an escape hatch for everything else.

AI Actions ship with native connectors to Stripe, Shopify, Zendesk, Salesforce, Calendly, Cal.com and Slack, with WhatsApp and Messenger alongside, and a Custom Action endpoint accepts any REST API, so coverage is bounded by what the customer can call rather than by a fixed catalog.

Real-time connectors sync to CRMs, order systems and help desks so replies reflect live business records rather than stale snapshots, and the model decides when to trigger an action from conversation context with guardrails preventing unintended operations.

Sourcechatbase.co/blog/chatbase-vs-custom-chatbotread 2026-09-05

AI Actions execute transactional work end to end rather than answering about it: managing orders through Shopify, pulling subscription and invoice details from Stripe for verified users, creating and auto-assigning Zendesk tickets with drafted responses, scheduling through Calendly and Cal.com, sending Slack alerts and updating records, with the model deciding when to trigger an action from conversation context and guardrails preventing unintended operations.

Escalation is part of the same chain, handing to a live agent through Salesforce Omni Channel with real-time presence and full conversation context carried across. Identity and Contacts authorize actions for known users, so a transaction is bound to a verified customer.

Sourcechatbase.co/blog/the-ai-agent-playbookread 2026-09-05

A maintained retrieval structure the customer owns and the platform keeps current. Chatbase ingests documents, web pages, help center articles and raw text into a searchable knowledge base retrieved at query time by a retrieval-augmented architecture, and can additionally ingest Zendesk and Salesforce support tickets as training data.

Source Suggestions automatically identify knowledge gaps from live conversations, and auto-retraining keeps the index in sync as source content changes, so the structure is maintained rather than loaded once. AI-powered guardrails ground responses in the configured sources.

Sourcechatbase.co/blog/the-ai-agent-playbookread 2026-09-05

Constraint sits ahead of the response and correction sits behind it. A proprietary prompt architecture prevents unconstrained generation by holding the model inside the configured knowledge scope, so the agent refuses to answer outside its domain, and guardrails prevent unintended operations when the model decides to trigger an action.

Behind that, the Improve Answer feature lets operators correct responses directly from chat logs and the correction persists, live agent handoff carries full context through Salesforce Omni Channel, and administrative control is granular: role-based access with custom roles across fourteen permission areas covering agents, sources, chat logs, contacts, integrations, billing, subscription, members, webhooks, workspace, API keys, actions, analytics and leads, with SSO on Enterprise. Chatbase states plainly that no guardrail system is perfect and that hallucination prevention is probabilistic.

Sourcechatbase.co/blog/chatbase-vs-custom-chatbotread 2026-09-05

SOC 2 Type II and GDPR compliant with AES 256 encryption at rest and in transit, SSO on the Enterprise plan added January 2026, RBAC with custom roles across 14 permission areas, per user rate limiting, domain allowlisting controlling where the agent can be embedded, and customer data never used to train models. A trust center at trust.chatbase.co publishes the subprocessor list, and the DPA commits to a subscribe mechanism notifying customers of new subprocessors.

Sourcetrust.chatbase.co/subprocessorsread 2026-09-05

Inspection tooling is named and step-level reconstruction is not.

Chatbase publishes Backstage as a workspace to debug, inspect and improve an agent, alongside analytics covering conversation logs, sentiment, resolution metrics and dashboards, with topic clustering, confidence scoring and geographic distribution, exportable to JSON, PDF and CSV, and reviewable chat logs behind them.

No step-level decision trace, tool-call log or retrieval record for an individual run is documented, and no configuration audit trail appears, so a team can read what was said and how it performed but not reconstruct which action fired and why.

Sourcechatbase.co/llm-inforead 2026-09-05

A customer record exists and an agent memory layer does not. Chatbase publishes Identity and Contacts as a platform component that personalizes and authorizes actions for known users, holding customer records alongside persisted chat logs, and full conversation context transfers when a conversation hands to a live agent. No cross-session agent memory with a stated scope or lifetime is documented. What persists is a contact record and a transcript, and a store is not a memory layer whoever writes it.

Sourcechatbase.co/blog/the-ai-agent-playbookread 2026-09-05

Hosting is fixed in the United States. Chatbase's privacy policy states customer personal information is stored in databases hosted by third parties located in the United States, with no alternative region offered, and the DPA covers restricted transfers through Standard Contractual Clauses and points to a subprocessor list rather than a residency option. Legal transfer cover with a single fixed jurisdiction is not residency, and no VPC, single-tenant or on-premises option is documented.

Sourcechatbase.co/legal/privacyread 2026-09-05

Packaged components exist below the level of a packaged agent. Pre-built AI Actions ship for the common support and commerce jobs, covering order management, ticket creation, subscription lookup, scheduling and lead capture, and an existing agent can be duplicated as a starting point for a new one. No agent template library, marketplace or named catalog of prebuilt agents a buyer browses and puts into service is documented; the unit a customer adopts is an action, not an agent.

Sourcechatbase.co/blog/the-ai-agent-playbookread 2026-09-05

One agent configuration reaches customers across six named channels and the phone: a customizable website widget or standalone hosted Agent Page, email where the agent reads, drafts and responds to incoming support mail around the clock, WhatsApp, Facebook Messenger, Instagram and Slack, plus telephony connecting the agent to phone lines to handle inbound calls with natural voice. Full API access sits alongside for custom surfaces, and the chat interface localizes into more than forty languages including right-to-left support.

Sourcechatbase.co/llm-inforead 2026-09-05

The customer chooses the model. Chatbase's no-code builder lets an operator configure the AI model alongside system instructions when creating an agent, and model comparison and selection is published as a platform differentiator, so an operator can directly compare models for the same agent rather than accepting a fixed provider.

Sourcechatbase.co/blog/chatbase-vs-custom-chatbotread 2026-09-05

The platform is callable from outside through documented interfaces: REST API v1 and v2 in beta give programmatic access to agent management, conversations and data, covering chatting with agents, creating and configuring them, and retrieving conversations, leads and analytics, with a JavaScript embed supporting identity verification through JWT and HMAC for authenticated sessions. Full API access is listed alongside the native channel deployments as a deployment surface in its own right, with capability availability varying by pricing tier.

Sourcechatbase.co/blog/chatbase-vs-custom-chatbotread 2026-09-05

A test surface exists on both sides of deployment and its verdict is not described.

Chatbase names Testing as a platform capability validating agent behavior before and after deployment, and Backstage as a workspace to debug, inspect and improve an agent, described as the control room for it; confidence scoring flags weak answers in production and the Improve Answer feature lets operators correct responses from chat logs so the correction persists.

No scored evaluation, scenario set, regression suite or readable comparable verdict on a proposed change is documented, so a team can exercise an agent but not compare two versions of it. Testing and Backstage are named on Chatbase's structured page for AI assistants, which asserts them rather than documenting how they work.

Sourcechatbase.co/llm-inforead 2026-09-05

No browser control, hosted session or computer use appears anywhere on Chatbase's product surface. Instead, the agent acts through configured AI Actions against native connectors and a Custom Action endpoint for any REST API, and reaches customers through its named channels and the public API rather than by operating an interface.

Sourcechatbase.co/llm-inforead 2026-09-05