Chatbase builds customer agents that answer questions and act on account data

Chatbase is a platform for building AI agents that answer customer questions and perform connected tasks. A team supplies its business information, configures behaviour and actions, and deploys an agent into customer channels. Its appeal is the ability to assemble that experience without building the complete conversation infrastructure. The serious buying question is what happens when an answer needs private account data, an external system change or a person’s judgement. Those boundaries determine both the useful scope and the implementation work.

01 / ProductWhat does Chatbase actually provide?

The product overview¹ presents knowledge sources, agent instructions, integrations, deployment and conversation improvement as parts of one platform. This puts Chatbase between a simple FAQ widget and a custom support application. The interface may look like a chat bubble, but an agent that retrieves an invoice or changes an account has a different responsibility from a searchable help page. Buyers should evaluate the information service and the action service separately.

A useful distinction is between company knowledge and customer state. A cancellation policy is shared knowledge. Whether a particular subscription is eligible for cancellation is account state. A language model can explain the former; a trustworthy workflow must retrieve and validate the latter. Chatbase’s documented identity and contact mechanisms make that distinction relevant to implementation, rather than leaving every request as anonymous text. The product does not remove the need for your systems to decide what a customer is permitted to do.

02 / AudienceWhich teams are likely to get useful value?

Chatbase is a plausible starting point for a software business, online service or support team with repeated questions and maintained documentation. The clearest early workload is bounded: explain a feature, identify the relevant policy, collect information and pass an unresolved case to the right person. That can be useful even before allowing the agent to make account changes. It also provides a smaller set of answers for the team to inspect when refining the service.

It is a less straightforward fit when the underlying information is contradictory, each customer has individually negotiated terms, or the desired task depends on inaccessible legacy systems. A conversational front end does not settle those issues. An organisation that already runs a substantial support operation should also examine how agents fit its inbox, routing and reporting. Our Intercom profile explores a helpdesk-centred approach; the relevant comparison is the cost of connecting and operating the whole support process, not which demo answers a question more fluently.

Can a nontechnical team implement it?

A public-information agent can be a business-team project, provided someone owns content and testing. Authenticated actions need engineering involvement. Chatbase’s identity verification guide³ recommends signed JWT tokens generated on the server and documents resetting the widget’s user on logout. That is a meaningful boundary: the website’s existing authentication must remain authoritative. Pasting an email address into a conversation should not become permission to access the corresponding account.

03 / WorkflowA proposed workflow for subscription support

Consider a software company that wants to answer billing questions and help customers find invoices. This is an illustrative deployment plan, not a test we performed. Start by separating public policies from account-specific operations. Import the current billing help material, remove superseded plan descriptions and define which exceptions need a person. Give the agent an explicit fallback when the sources do not answer the question. A useful answer may be a handoff with context rather than an attempted resolution.

Chatbase’s source preparation guidance⁴ stresses readable text and notes that inaccessible website content can undermine answers. In practice, inspect what was ingested rather than assuming a successful crawl captured the right policy. A pricing page that mixes archived offers and current plans is a poor source even when every word is technically readable. Create a small set of unambiguous reference answers for the questions that cause expensive misunderstandings.

Next, connect authenticated customers to the right account record. Use the server’s trusted user identity and only the fields necessary for the task. For the invoice operation, the connected service should select invoices belonging to that account; it should not trust an arbitrary customer identifier supplied in the conversation. Test logout, account switching and an expired session. These cases are easy to overlook because they occur around the chat rather than inside the model’s answer.

The Playground documentation⁵ describes a place to configure and try agent behaviour before deployment. Use that environment for representative requests: a valid invoice question, a missing invoice, a request for another person’s details and a policy exception. Record the expected answer or escalation for each. When an action fails, the desired response should acknowledge the failure and offer a next step; it should not describe the attempted operation as completed.

After a limited release, inspect conversations alongside actual account outcomes. Count an invoice request as successful when the customer reaches the correct document, not merely when the agent sends a link. Assign somebody to correct source gaps and review unresolved cases. If the supporting workflow spans multiple applications, n8n’s workflow approach is relevant to the integration layer. It addresses how steps and failures are orchestrated, while Chatbase supplies the customer conversation.

04 / PricingHow Chatbase pricing works

The public pricing page² uses subscriptions with message-credit allowances. The following is its monthly-billing view checked on 15 September 2026; dollar amounts are USD. Credits are a platform billing unit, so do not treat the allowance as a guaranteed count of completed customer cases. Model choice and the shape of conversations need to be reflected in a measured trial before estimating operating cost.

The page also lists automatic recharge at $40 per 1,000 message credits, an extra agent at $25 per month and branding removal at $99 per month. For a production budget, identify mandatory capabilities before choosing the smallest subscription. A prototype can demonstrate answer quality without demonstrating the API, personalised experience or support workflow that the eventual deployment requires. Enterprise security and administration requirements can likewise move the buying decision beyond self-service plans.

An illustrative cost worksheet should separate the subscription, recharge, optional add-ons and your team’s implementation time. Avoid estimating annual savings from the number of incoming messages alone: repeat contacts, human follow-up and incorrect actions can change the economics. During the trial, group conversations by intent and compare credits consumed with successfully completed requests. That gives a more useful cost denominator than an average credit count across unrelated conversations.

05 / DistinctionsThe important distinction is the operating loop

The strongest architectural reason to consider Chatbase is that the agent can be maintained as a service rather than treated as a finished prompt. Sources, configured actions and conversation inspection are connected parts of its public product story. Our interpretation is that this makes the product particularly relevant when the business wants to own customer-facing behaviour but does not want to build every surrounding interface. The benefit still depends on someone regularly doing that maintenance.

The settings documentation⁶ describes agent credit limits, domain restrictions and request rate controls. These serve different purposes. A domain restriction helps control where a widget is embedded; it does not establish that a customer owns an account. A credit cap helps contain usage; it does not guarantee a graceful experience when the allowance runs out. Test the visible failure behaviour as part of the service, because customers experience limits as interrupted help.

For a founder studying the business, this is a useful packaging lesson. The customer buys a supported path from question to outcome, including setup, monitoring and escalation. Presenting only the generated answer hides much of the work required to deliver that outcome. Conversely, adding every possible integration can increase operational complexity before the team has proved that its first use case is useful.

06 / QuestionsWhat should buyers verify before committing?

The privacy policy⁷ describes US data storage, service-data retention and the company’s statement that customer data is not used to train AI models. Treat those statements as a starting point for reviewing the actual deployment, not as a substitute for the applicable agreement. Establish what the widget collects, which connected systems receive data, who can inspect conversations and how deletion requests flow through those systems.

There are also practical questions the public pages cannot answer for a particular company: whether its source documents are retrieved reliably, whether its account actions enforce policy, how often a person must intervene, and how customers react to the handoff. We did not use a private workspace or measure those outcomes. A useful pilot should include ambiguous requests, conflicting source material and unavailable downstream services, with expected behaviour agreed in advance.

07 / DecisionChoose the smallest complete service to evaluate

The right first deployment has a clear customer task, an authoritative information source and a named owner for exceptions. Decide in advance what evidence would justify extending its scope: correct answers, verified action results and manageable follow-up work. Chatbase deserves evaluation when those ingredients exist. If they do not, improving the support process itself is likely to be the more productive first investment.