Alliance corporate platform
Our own bilingual corporate platform: five practice sites, projects, insights, search and lead capture on an edge-hosted, near-zero-JavaScript architecture.
Client engagement · IT services
An AI agent that reads incoming tickets, classifies and routes them, and drafts first responses for a technician to approve.
01
An IT service provider wanted to spend less technician time on sorting tickets and more on solving them. Every new ticket had to be read, categorised, prioritised and routed before real work could begin, and first responses varied in quality.
We designed and built an AI agent that triages incoming tickets and drafts a first response for a technician to approve, with governance, security and measurement built in from the start.
02
Technicians spent the first part of every ticket working out what it was about and who should own it. Quality varied with who happened to be on shift.
Tickets arrived by email, portal and Teams, often with screenshots and incomplete information. Routing rules lived in people’s heads, so accuracy depended on who was on shift.
The organisation was open to AI but had clear conditions: no customer data used to train public models, a human decision before anything reaches a customer, and a full audit trail.
03
04
Analysis of historical tickets and routing rules with senior technicians, definition of categories and priorities, and a governance checklist agreed with leadership.
A narrow agent role: read tickets and the knowledge base, suggest category, priority and queue, and draft a reply. No ability to close tickets, change systems or email customers directly.
Retrieval over approved knowledge articles, prompt and tool design, redaction of personal information before model calls, and integration with the ticketing workflow and Teams.
A test set built from real, anonymised tickets with known correct answers, run on every change to prompts, models or tools, with accuracy and escalation tracked over time.
Shadow mode first, where suggestions were logged but not shown, then assisted mode for one team, then wider rollout as quality targets were met.
05
06
Every draft reply requires technician approval, and the agent’s suggestions are shown with the reasons behind them.
The agent runs under its own least-privilege identity, cannot reach systems outside its role, and treats ticket content as untrusted input to limit prompt-injection risk.
Personal information is redacted before processing, model data is not used for training, and every suggestion, approval and change is logged.
07
Tickets reach the right queue from the start, first responses are consistent, and every AI action is logged and reviewable.
08
The evaluation suite runs on every change, knowledge articles are refreshed as gaps are found, and new workflows are added only once the current ones meet their quality targets.
Our own bilingual corporate platform: five practice sites, projects, insights, search and lead capture on an edge-hosted, near-zero-JavaScript architecture.
One accountable IT team for several plants and a head office: monitoring, patching, identity security and a helpdesk that knows the shop floor.
A tenant secured and cleaned up before AI arrived, so Copilot could be switched on without surfacing what it should not.
Talk to us
A 30-minute consultation with a senior consultant. You leave with a clear next step — whether or not it involves us.