AIWeb, WhatsApp18 weeks SprintConceptual Case Study - Proposed Solution

AnswerDesk

An AI Support Copilot That Answers From Your Help Content and Knows When to Hand Over

A concept study: we worked up this product, its architecture and its screens as an example of how we would approach it. The screens use fictional demo data and are design concepts, not screenshots of a live product, and this is not a client project.

AnswerDesk - Engineering Case Study Cover
Concept
A design study, not a live product
Fictional
Demo data on every screen
18 weeks
Estimated build time

About the Project

AnswerDesk is a design concept for an AI customer support platform for product companies and online retailers that get the same questions every day in English and Hindi. The bot answers only from the company's approved help articles and shows the passages it used. It never answers from general knowledge.

Before any answer is sent, a second step splits the draft into claims and checks that each one is backed by a passage. If a claim has no source, the bot does not send it. It asks a question or passes the chat to a person. Asking for a person always works, in one message, and the agent opens the chat with the transcript, the sources the bot used, the customer's recent orders and a short summary.

Support leads review sampled chats and mark answers right or wrong. Each marked chat becomes a saved test, and a change to a prompt, model or article goes live only after those tests pass. The knowledge owner sees a queue of questions that no article answers, so the help centre improves where customers actually ask.

Why This Matters Now

Market, rule and technology context for this kind of product, checked on 16 November 2026. Each item names its source.

  • Arize ran a test where a web page told a support agent to look up another customer: it leaked on 28 of 100 runs, and binding tools to the session in code stopped the data leaving.Arize (vendor blog), 26 Oct 2026
  • DesignRush, summarising Gartner's survey of 3,566 customers, reports that 87 percent say AI customer service must offer access to a human agent (trade press, citing Gartner releases).DesignRush, 21 Oct 2026
  • Law firm Lewis Silkin reports the EU's AI Omnibus took effect on 27 Jul 2026, and duties to tell users they are dealing with AI under Article 50 still apply from 2 Aug 2026.Lewis Silkin, 27 Jul 2026
  • DIGIT reports a Gartner survey in which customers were about three times more likely to use third-party generative AI than a company's own chatbot for service issues.DIGIT, 9 Jul 2026
  • Salesforce announced a Help Agent grounded in Salesforce Knowledge, priced per autonomous resolution, with no charge when a customer asks for a human.Salesforce, 25 Jun 2026
  • Law firm MME reports that on 12 May 2026 a German appeal court (Hamm) held a clinic responsible for false statements by its website chatbot, even if the bot was trained on accurate data.MME (law firm magazine), 12 May 2026

Requirements & Objectives

  • Answer customer questions in English and Hindi only from approved help articles, and show a link to each passage used.
  • Check every claim in a draft answer against the retrieved passages, and block the send when a claim has no support.
  • Hand a chat to a person when the customer asks, when the topic is on a sensitive list or when the answer check fails.
  • Give the agent the chat, the sources used, recent orders and a short summary, so the customer never repeats the problem.
  • Read order data through server-side tools tied to the signed-in customer, with a confirmation step before any change.
  • Group unanswered questions by topic, with an owner and a status, so the content team knows which article to write or fix.
  • Save every answer marked wrong in review as a test, and run the test set before any prompt, model or article change goes live.
  • Tell customers they are talking to an AI, keep consent records and support retention settings and deletion by request.

The Engineered Solution

1Answers built from approved content

The bot retrieves passages from the knowledge base and may use only those passages. Each answer carries source chips with the article name and its update date, and expired articles are left out.

2A check before every send

A second step splits the draft into claims and tests each one against the passages. One unsupported claim stops the send, and the chat goes to a person or the bot asks a clarifying question.

3Handover as a normal path

Handoff rules are a readable list that a support lead edits. When a rule fires, the agent receives the full chat, the sources, recent orders and a summary, and the customer is told who is joining.

4Mistakes that become tests

Leads mark sampled chats right or wrong. Each wrong answer is saved as a test, and a release candidate shows which tests changed from pass to fail before anything goes live.

Design Concept Screens

Design concepts with fictional demo data. They show how the product could look and work, not a live release.

AnswerDesk agent workspace with queue counts for waiting, handed over and mine, an inbox list, a conversation with a cited bot answer and a handoff, a customer profile, recent orders, a drafted reply with sources and a bot hand-over note

01 / 05 Agent workspace - take a handed-over chat, read the summary and sources, edit the drafted reply and send it

Engineering Process

  • Collect the top 50 customer questions from the support team and mark which help articles answer them and which do not.
  • Model the data: tenants, articles with owners and review dates, passages, conversations, answer checks, handoffs, review marks and tests.
  • Prototype the chat widget and the agent workspace and walk a few agents through a handed-over chat before building the pipeline.
  • Build retrieval and the draft-and-check pipeline, and score it on a first set of 80 test questions in English and Hindi.
  • Build the session-bound order tool and handoff rules, and try to break them with prompt-injection tests.
  • Build the gap queue, review marking and the CI test run, so a prompt or content change cannot go live with a failing test.
  • Load test the chat gateway, add the provider fallback and retention settings, then run a pilot with one support team.

Production Tech Stack

Layer / ComponentProduction Technologies
Web apps
Next.js (App Router)TypeScriptTailwind CSSshadcn/uiZustand for workspace state
Chat widget
Preact web component in a shadow DOMso it does not clash with the host site's styles
Backend
Node.js with TypeScriptZod validation on every input and model outputone modular application at first
Data and search
PostgreSQL on Supabase with pgvector and full-text search for hybrid retrievalrow-level security by tenant
AI layer
Anthropic Claude or OpenAI for draftsa smaller model call for the claim checka rerankerpinned model versions and a provider fallback
Channels and hosting
WhatsApp Business Platformemail intakeVercela container service for the chat gatewayRedisGitHub Actions for CI

What the design aims for

AnswerDesk aims to ground every support reply in a company's own approved help content and to route the chat to a person, with full context, whenever no source supports the answer.

ARCHITECTURE & SCOPE CONSULTATION

Need a System Engineered Like AnswerDesk?

Deploy a dedicated engineering pod or initiate a fixed-scope milestone sprint with full IP ownership and guaranteed delivery.