How to Build an AI Chatbot For Customer Support that Can Cut Support Costs By Up to 30%.

An AI customer-support chatbot can be deployed quickly by grounding answers in a company’s own help-center pages, with built-in citations and a feedback loop that helps teams reduce support costs—claims in the demo peg savings at up to 30%. The workflow centers on ingesting documentation, turning it into searchable knowledge (via chunking and embeddings), and then serving responses through a website widget or an API, so customers get fast, policy-accurate answers instead of generic model output.

The demonstration uses Jablo’s help center as the training source. After signing up, the user creates a Mendable project and connects a data source through a website crawler. The crawler pulls help-center content from the site map, filters down to 74 relevant help articles, and processes them in parallel. The ingestion step handles the “hard stuff” automatically—text splitting, chunking, metadata extraction, and embeddings—so the team doesn’t need to build a separate retrieval pipeline. Despite expectations that training might take days, the demo shows the process completing in minutes, with slower pages attributed to JavaScript-heavy or blocking content.

Once ingestion finishes, the chatbot is testable immediately. A user asks a specific question—such as what happens if a flight is cancelled—and receives an answer tailored to the help-center policy. Crucially, the response includes references that link back to the exact source page(s), addressing skepticism about hallucinations. The demo also highlights a retrieve-augmented generation approach: the model generates responses using retrieved context from the company’s documentation, which makes it more likely to answer precisely than a generic assistant.

Deployment is presented as “three buttons” using Mendable’s embeddable components. The user obtains an API key from the dashboard, swaps it into a JavaScript snippet, and the chat widget appears on the Jablo website with customizable branding (logo, colors, and UI elements). The dashboard also includes a “Workshop” area for customizing behavior: selecting underlying models (including GPT 3.5 turbo and Claude from Anthropic), adjusting prompts, controlling response length, and setting a “creativity” mode to “precise” so answers stay within the documentation.

To improve over time, the system logs conversations and user feedback (helpful/not helpful). If an answer is rated poorly, the “teach model” feature lets teams correct the bot without editing the help center itself. The correction process associates the new answer with the relevant question using embeddings and metadata, so similar future queries get prioritized toward the corrected response.

The demo further supports hybrid support flows. If the bot can’t answer, it can direct users to a support-agent link embedded in the widget, and the custom prompt can instruct the assistant to route unresolved questions to that contact path. Mendable also supports multiple chatbots for different data sources within the same account (e.g., customer support vs. sales).

Finally, the pricing discussion frames costs around message credits: a free tier includes 500 message credits per month, while higher-volume usage, fine-tuning, and enterprise features (like SSO and SLA support) require a custom plan. The overall pitch ties speed of implementation, measurable cost reduction, and continuous documentation improvement into a single operational loop for customer support teams.