CUSTOMER AI · CHATBOT

A customer-facing AI assistant
grounded in your knowledge.

Build your Knowledge Base. Deploy the chatbot on your site. Test it before launch, measure its quality, and hand off to a human when needed.

KNOWLEDGE BASE

The chatbot answers from what you put in it.

Before the chatbot can answer a single question, you build the Knowledge Base. Upload documents, connect your help centre, add policies, and link structured data. The chatbot draws only from this approved content.

When you update a document, the chatbot reflects the change at the next sync. No retraining. No manual rebuilds. The source of truth is always what you last approved.

Help articles and policies

Upload documents, paste content, or sync from a URL.

Website and help centre

Connect your public-facing docs and product pages.

Structured records

Sync product data, FAQs, and business records.

Chatbot scope control

Restrict the chatbot to specific knowledge sets per deployment target.

DEPLOY

From Knowledge Base to live chatbot in five steps.

The chatbot deploys as a lightweight widget. One script tag on your site is all it takes.

STEP 1
Add the script tag

Copy one line of JavaScript into your site's <head>. The widget loads asynchronously and does not affect page performance.

STEP 2
Configure the widget

Set the chatbot name, colour, position, and initial greeting. Apply source filters if you want to scope answers to a specific knowledge set.

STEP 3
Run Test Chat

Use the built-in sandbox to send sample questions and review answers before the chatbot is visible to customers. Fix knowledge gaps before launch.

STEP 4
Pass AI Readiness

The readiness scorecard runs automatically. Review the results, address any flagged issues, and publish once all thresholds are met.

STEP 5
Go live

The chatbot is live on your site. Analytics, gap logging, and live support escalation are active from the first conversation.

EMBED CODE
<script src="https://cdn.composedai.com/widget.js"
data-workspace="your-workspace-id">
</script>
Test Chat — SandboxTESTING MODE
What is your refund policy for annual plans?
Annual plans are eligible for a pro-rated refund within 30 days of the billing date. After 30 days, the plan is non-refundable.
✓ GroundedSource: Refund Policy v3
TEST CHAT

Test the chatbot before any customer sees it.

The Test Chat sandbox lets you send questions directly to the chatbot and review every answer before going live. Each test response shows the source document it was drawn from, the grounding score, and any flags raised.

Use Test Chat to identify gaps in your Knowledge Base, tune the chatbot tone, and confirm edge cases are handled correctly. Nothing is published until you are satisfied.

ANALYTICS

See what your chatbot is doing in production.

Analytics run from the first conversation. No third-party tracking required.

Deflection rate

The percentage of conversations resolved by the chatbot without requiring a human handoff. Measured per day, week, and knowledge category.

Conversation volume

Total conversations, unique users, and average session length. Segmented by page, language, and time of day.

Gap logs

Questions the chatbot could not answer from your Knowledge Base. Reviewed by your team and used to identify missing content.

Answer quality trends

Grounding scores and hallucination flags over time. Shows whether chatbot quality is improving or degrading as your Knowledge Base changes.

Escalation patterns

Which questions most frequently trigger a handoff to a human agent. Use this to prioritise new Knowledge Base content.

LIVE SUPPORT

When the chatbot cannot help, a human can.

The chatbot recognises when a question is outside its knowledge scope, when a customer is frustrated, or when a situation requires a human judgment call. At that point it initiates a clean handoff.

The human agent receives the full conversation transcript, the questions that were asked, the answers that were given, and the point at which escalation was triggered — so they do not need to start from scratch.

Trigger-based escalation

Define the conditions that cause a handoff — unanswered questions, escalation keywords, or customer-initiated requests.

Full conversation context

The agent sees the entire thread: questions asked, answers given, and the escalation point.

Multilingual response widget

Interact with customers in their preferred language. Supports all 40+ global and regional Indian languages (Hindi, Tamil, Bengali, Spanish, French).

Escalation analytics

Track escalation volume and recurring triggers to identify Knowledge Base gaps.

CHATBOT QUALITY SYSTEM · AI READINESS

Know whether your chatbot is ready before customers do.

AI Readiness is not a separate product. It is the quality system built into the chatbot deployment workflow. Before your chatbot goes live — and on every Knowledge Base update — it runs a scored evaluation across seven dimensions.

Accuracy

Compares each answer against the source document to verify factual correctness. Flags answers that diverge from what your knowledge says.

Grounding

Measures whether the answer is derived only from your connected knowledge. Ungrounded responses — generated from model memory rather than your documents — are flagged.

Hallucinations

Detects claims in the answer that are not supported by any source record. Hallucinated content is blocked or annotated before it reaches the customer.

Coverage

Tracks which customer questions your Knowledge Base can and cannot answer. Questions outside coverage become logged gaps for your team to review.

Unknown Handling

Validates that the chatbot responds correctly when it does not know the answer — declining gracefully rather than guessing or producing a low-confidence reply.

Security

Checks for prompt injection attempts, data exfiltration patterns, and unauthorised content in both inputs and outputs before they complete.

Deployment Readiness

Produces a scorecard across all dimensions before each deployment. The chatbot does not publish until it passes the readiness threshold you configure.

DEPLOYMENT READINESS SCORECARD

Each dimension produces a score from 0 to 100. You set the minimum threshold required before publication. Composed AI enforces strict grounding guardrails: if the chatbot detects that the answer is not present in the Knowledge Base, it is blocked from guessing. The chatbot does not go live until all safety thresholds are met.

AccuracyGroundingHallucinationsCoverageUnknown HandlingSecurity→ Go / No-Go

Build your Knowledge Base. Deploy your chatbot.

Start with your existing documents. Test the chatbot before it goes live. Measure quality with every update.