All work
042023SCAD

AI Conversational Chatbot (18K+ Monthly Queries)

One assistant absorbing 18,000+ support questions a month — so 500+ people a day stop waiting in a human queue.

Before

Support agents were drowning in repetitive, high-volume queries about publications, data definitions, and admin processes. Response times stretched, and the same questions were answered over and over.

After

A context-aware assistant now resolves 90% of questions on first contact, handles 18,000+ queries a month for 500+ daily users, and cut support costs by 43% — freeing agents for the genuinely hard cases.

Challenge

High-volume support queries overwhelming human agents.

Approach

Deployed conversational AI chatbot with context-aware multi-turn dialogue and intent routing.

How it was built

  1. 1

    Query mining

    Phase 1

    Analysed historical support tickets to find the highest-volume intents. A small set of question types accounted for the majority of load — the obvious first targets for automation.

  2. 2

    Intent routing

    Phase 2

    Built an intent classifier in front of the model so each question is routed to the right knowledge and tone, rather than sending everything to a single generic prompt.

  3. 3

    Multi-turn dialogue

    Phase 3

    Added context-aware multi-turn handling so follow-up questions keep their thread — the difference between a demo and something people actually rely on.

  4. 4

    Escalation & rollout

    Phase 4 — Present

    Confidence-gated hand-off to human agents for anything the assistant is unsure about, then a staged rollout to staff and the public with ongoing review of missed answers.

Key architecture decisions

Intent routing before generation

Why · Routing to the right knowledge slice and tone per intent is what pushed first-contact resolution to 90% — far more than prompt-tuning a single catch-all prompt.

Confidence-gated human escalation

Why · A support bot that guesses erodes trust fast. Handing low-confidence questions to a human keeps quality high while still deflecting the bulk of volume.

Grounded answers over free generation

Why · Retrieval augmentation keeps answers tied to real publications and definitions, which matters when the audience includes the public.

Impact

  • Handling 18K+ monthly queries with 90% first-contact resolution
  • Cut support costs by 43%
  • Serving 500+ users daily
18K+/mo
queries
90%
resolution
-43%
cost

What I'd tell someone building this

  • 01 · Most support volume is a handful of intents — automate those first, not everything.
  • 02 · First-contact resolution is the metric that actually reflects user experience.
  • 03 · An escalation path is a feature, not a fallback — it's what makes automation safe to ship.
  • 04 · Multi-turn context is what turns a novelty bot into a daily-use tool.

Tech stack

Azure Bot FrameworkGPT-4Azure OpenAILangChain.NET Core

Ask anything about AI Conversational Chatbot (18K+ Monthly Queries)

AI scoped to this project · GPT-OSS 120B