AI Support Automation · SaaS Platform
PagerGPT lets any team train a chatbot on their own knowledge, deploy it across web, Slack and Teams, and hand off to humans when it matters — turning scattered help docs into answers in seconds.
01 — Overview
Support teams are drowning in repetitive questions while the answers already exist — buried in help centres, PDFs and Slack threads. PagerGPT turns that scattered knowledge into a chatbot anyone can build in an afternoon, deploy anywhere, and trust to escalate to a human the moment it's out of its depth.
As Lead Product Designer I shaped the full experience — from the training pipeline and live agent inbox to the analytics dashboard and the component system tying it all together.
The core tension: make something powerful enough for support ops, yet simple enough that a non-technical admin never feels they're "configuring an AI." Every screen had to disappear behind the task.
02 — The Challenge
Answers live across help centres, PDFs, wikis and Slack threads. Agents re-type the same replies all day while users never find the docs.
Off-the-shelf chatbots are rigid and hard to train. They dead-end users instead of helping — so people learn to skip the bot and wait for a human.
Teams can't see which questions go unanswered, where the bot breaks down, or what knowledge to add next. The loop never closes.
03 — Design Goals
Point the bot at a URL, sitemap or file — no prompts, no code. Knowledge in, answers out.
The bot knows its limits and passes to a human — with full context — the instant confidence drops.
Every unanswered question surfaces in analytics, so the knowledge base improves itself over time.
Web widget, Slack, Teams and the agent inbox all speak one visual and interaction language.
04 — How It Works
Point it at URLs, sitemaps, files and apps.
Chat with the draft bot and refine its answers.
Ship to a web widget, Slack or Teams in one click.
It answers instantly, 24/7, with real context.
Unanswered questions surface — you top up knowledge.
05 — Research & Insights
"By the time I've explained the problem to the bot, I could've Googled it."
So training became a 3-step wizard: add a source, auto-fetch, train — zero prompt engineering.
"If the bot can't help, don't make me start over with a human."
So escalation carries the full transcript and user details straight into the agent inbox.
"I honestly don't know what my bot is bad at."
So the dashboard leads with unanswered questions and feedback — not vanity session counts.
06 — Design System
A geometric display face carries personality; a neutral UI sans keeps data-heavy screens quiet. Electric indigo does the pointing; everything else gets out of the way.
Aa
Plus Jakarta Sans
Display & headings — a geometric sans (standing in for the product's Euclid Circular B) that gives PagerGPT a modern, confident voice.
ABCDEFG · abcdefg · 0123456789
Aa
Inter
Body & UI — a neutral workhorse that stays legible in tables, logs and dense dashboards without ever competing for attention.
ABCDEFG · abcdefg · 0123456789
Palette
07 — The Product
Add a website, sitemap, files or connected apps — PagerGPT fetches and indexes everything, shows a live character count, and trains in the background. The three-step rail keeps "what's next" obvious the whole way.
A three-pane log pairs each transcript with rich context — channel, questions asked, bot answered vs. unanswered, feedback and escalation — so teams can audit quality at a glance and spot exactly where knowledge is missing.
08 — Deep Dive · Analytics
Instead of vanity metrics, the home dashboard foregrounds what a support team can act on — what got asked, what went unanswered, and how people felt about the answer.
Sessions, asked, answered and unanswered — the four numbers that matter, up top.
A yearly question curve shows demand spikes so staffing and content can keep pace.
The real backlog — the exact queries to turn into new knowledge next.
Helpful vs. not vs. escalated, so quality is a number, not a gut feeling.
09 — Test & Deploy
A live test chat runs the trained bot right inside the wizard — no staging, no guesswork.
Swap the underlying model (GPT-4o and others) from a dropdown — never a config file.
One button pushes the same bot to a web widget, Slack or Microsoft Teams.
10 — Omnichannel
An embeddable, fully themeable widget — the familiar bottom-right helper that matches your site out of the box.
Users ask in a DM or channel; the bot answers inline and escalates to your agents without leaving Slack.
The same trained bot inside Teams — so internal IT and HR support lives exactly where employees already work.
11 — Outcomes
Illustrative metrics — swap in real figures once you have them.
Questions answered end-to-end without a human agent.
Deflected by the bot across web, Slack and Teams.
From a user's question to the first useful answer.
"Helpful" feedback averaged across live sessions.
12 — Reflection
Restraint was the real design tool. The temptation with an LLM product is to expose every knob; the win was hiding almost all of them behind three verbs — train, test, deploy — so a non-technical admin ships a capable bot without ever feeling they're "configuring an AI."
Designing the failure path mattered as much as the happy path. The human-handoff flow and the unanswered-questions loop are what make the confident moments believable — they're the reason people keep trusting the bot instead of routing around it.