In-Portal AI Copilot for Helpdesk Ticket Handling

Executive Summary

Consultants at a technology services firm spend their day inside a single ageing helpdesk portal — reading ticket history, drafting customer-facing updates, logging time, and moving work through a queue. The portal has no usable API and a fragmented multi-frame layout, so every AI tool the consultants wanted to use meant copying context out to a separate chat window and pasting results back in. That round-trip was slow, error-prone, and — because the tickets contain client data — a governance problem the moment anything left the firm’s own systems.

We built a browser extension that brings a private, ticket-aware AI copilot into the portal itself. It reads live ticket context, drafts and summarises through the firm’s own AI platform, transcribes dictation on-device, and layers workflow guardrails over the existing screens — all without a single change to the portal or its backend. Nothing leaves the firm’s environment: the AI panel talks only to the company’s private AI endpoint, and voice dictation is transcribed entirely on the consultant’s machine.

The Challenge

The firm was already sold on AI for consulting work; the problem was getting it safely to the point of use.

• AI lived outside the workflow. Drafting a reply or summarising a ticket for handover meant leaving the portal, reconstructing context in a general chat tool, and pasting the result back. The context-switch cost more time than the AI saved.
• Client data couldn’t go to public APIs. Tickets contain customer detail. Pasting them into a public, multi-tenant AI service was not acceptable, so consultants either self-censored the context (worse output) or avoided AI on real tickets entirely.
• The portal was a hostile integration target. A legacy, multi-frame/iframe layout with no API meant there was no clean programmatic surface to build against — and no appetite to modify a business-critical system.
• Death by a thousand small tasks. Around the AI gap sat a long tail of manual overhead: re-reading tickets to rebuild context, hand-typing repetitive comments, spotting SLA risks by eye, and reconstructing the week’s hours at timesheet time.

The Solution

We delivered a Chrome/Edge (Manifest V3) extension that overlays the portal and hosts a Copilot-style AI side panel. It is non-invasive by design — it reads the portal and augments its UI, but changes nothing server-side — and everything it does sits behind a master enable switch that defaults to off, so rollout is opt-in and reversible per user.

Architecturally, the extension is a thin bridge between two systems the firm already trusts — its ticket portal on one side and its own private AI platform on the other — with a completely separate, fully local path for voice that never touches the network:

Key Architectural Decisions

1. A private AI panel, grounded on the live ticket The side panel is a chat copilot wired exclusively to the firm’s own AI platform — never a public multi-tenant API. A “Use current ticket” action injects a structured snapshot of the ticket (status, priority, dates, time totals, quote, comment history) as grounded context, so answers are specific rather than generic. Per-environment access tokens are held in session storage only and are never synced or logged.

2. Structured extraction instead of scraping Rather than scraping the rendered page, a content script reads the real ticket edit iframe and assembles a single typed ticketData object shared across the portal’s frames. That clean data shape is what makes both the AI grounding and the workflow guardrails reliable on a messy legacy layout.

3. AI tasks as editable prompts, not hard-coded behaviour Common jobs — Draft Reply, Summarise for Handover, Infer Time, and dashboard-slice analysis — run as headless, ephemeral tasks driven by versioned Markdown prompts. The firm can tune tone and policy by editing a prompt file; no code change or redeploy is required, and no conversation is retained after the task completes.

4. Fully on-device voice dictation Push-to-talk dictation is transcribed by a speech-to-text model running entirely on the consultant’s machine. Audio never leaves the device — there is no cloud transcription service in the path — and transcribed text is smart-spaced straight into the comment composer, so consultants can compose updates hands-free without a privacy trade-off.

5. Workflow guardrails layered over the existing screens On top of the AI, the extension adds the operational scaffolding the portal lacks: a warning panel that flags overdue contact, past delivery dates, unassigned or over-quote tickets; a comment template picker with @-mention resolution and state-aware suggestions; one-click claim for unowned queue tickets; and a passive time-tracking engine that estimates per-ticket hours across the week to make timesheet completion nearly automatic.

The Outcomes

• AI is now usable on real tickets, safely. Because the copilot only ever talks to the firm’s private AI platform and voice stays on-device, consultants can ground the AI on actual client tickets without a governance exception. The tool that was previously avoided on sensitive work is now the default.
• No more copy-paste round-trip. Drafting a reply or a handover summary happens in the panel next to the ticket, already grounded on its context — eliminating the leave-the-portal, rebuild-context, paste-back loop that made AI feel slower than doing it by hand.
• Hands-free composition. On-device dictation lets consultants talk through an update while reviewing the ticket, with the text landing cleanly in the composer — useful for longer customer-facing notes without cloud-transcription privacy concerns.
• Fewer SLA surprises. The warning panel surfaces overdue-contact, over-quote and stale-update risk at the moment a ticket is opened, shifting these from “found at review” to “caught before acting.”
• Timesheets that fill themselves. Passive per-ticket time capture turns end-of-week timesheet reconstruction — previously a guess against memory and calendars — into a review-and-confirm step.
• Zero backend footprint. The whole capability ships as a browser extension: no portal modification, no API request to the vendor, no server to run. It works with the existing system exactly as-is and can be switched off per user at any time.

Measurement & Limitations

• Baseline: The firm’s own view, captured during discovery, was that AI on real tickets was effectively unusable in-workflow — the copy-paste round-trip and the public-API data concern meant consultants either avoided it or fed it de-scoped context. Routine ticket overhead (rebuilding context, repetitive comments, timesheet reconstruction) was a recurring daily drain.
• How the outcomes are framed: The results above are qualitative workflow improvements and the firm’s own before/after impressions from daily use — not an independently audited time-and-motion study. The headline is a governance outcome (private AI, zero external data egress), which is a property of the architecture rather than an estimated percentage.
• Scope & caveats: Figures and impressions come from a single firm and reflect its portal, ticket mix, and its own AI platform. They are indicative of what was achieved here, not a guaranteed result elsewhere. The AI drafts and infers; a consultant reviews and sends every customer-facing comment, and time estimates are suggestions the consultant confirms before booking.
• Provenance: The client, product name, and portal are anonymised at the firm’s request. The extension architecture, prompt set, and data-handling approach can be reviewed under NDA. No third-party security audit of these claims has been performed.

Technology Stack

• Surface: Chrome / Edge extension, Manifest V3, built with WXT (TypeScript)
• Integration model: Non-invasive content-script overlay + background service worker — no portal or backend changes, gated behind an opt-in master switch
• AI: Copilot-style side panel bound to the firm’s private AI platform (no public multi-tenant API); ticket-grounded context injection; headless tasks driven by editable, versioned Markdown prompts
• Voice: Fully on-device push-to-talk speech-to-text — audio never leaves the machine
• Data handling: Structured extraction from the real ticket iframe into a shared typed object; AI tokens in session storage (never synced or logged); time-tracking data local-only and auto-purged weekly
• Workflow layer: SLA warning panel, comment template picker with @-mention resolution, one-click claim, and a passive weekly time-tracking engine