Accelerator 15 · IT operations & customer support
Agentic IT support desk with grounded answers and human hand-off
Answers employees' IT questions only from your approved knowledge base, with citations, checks every answer before it is shown, and hands risky cases to a person with a ready ticket.
The problem
- Most service-desk contacts are repeats: passwords, MFA, VPN, Outlook, printers, access requests.
- Chatbots that answer from general knowledge give confident but wrong steps, and the desk cleans up afterwards.
- Escalations arrive with little context, so engineers start again by asking what the user already tried.
- Security incidents reported in chat must reach the SOC immediately, not get a generic reply.
What it does
- Understands the request (intent plus a standalone search query, so follow-up questions work) and applies deterministic hand-off rules for security incidents, access approvals, hardware damage and requests for a person.
- Retrieves the top articles from your curated knowledge base. Retrieved text is treated as data, never as instructions.
- Drafts an answer only from the retrieved articles, with numbered citations.
- Checks the draft BEFORE showing it: citation and word-overlap checks in code, plus a checker model that sees the same passages. Anything not supported is not shown.
- Hands off with a structured ticket draft: category, priority from your SLA matrix, summary, steps tried and user impact.
Ask the demo service desk of a fictional bank a question, or pick a starter. Watch each agent step: intent, retrieval, draft, grounding check, then an answer with citations or a hand-off ticket.
Live demo: Agentic IT support desk
Synthetic data only · nothing you type is stored
Northbridge Financial (fictional) · IT Service Desk
Answers come only from the 53 articles in the demo knowledge base, with citations, and are checked before they are shown. Security incidents, access approvals, hardware damage and requests for a person go to a human with a drafted ticket.
The demo first calls the Surge in-house AI server and falls back to a hosted open-source model if it is busy. The badge on the output shows which one answered. Limited to a few runs per visitor per day.
How it integrates
In your deployment the accelerator runs inside your environment, behind your single sign-on, and is connected to the systems below. You choose the hosting and the model.
Systems and APIs
- ServiceNow or Jira Service Management: create the incident or request, attach the transcript and cited articles
- Microsoft Teams (or Slack) as the employee channel, with live hand-off to the desk queue
- Knowledge sources: ServiceNow KB, Confluence, SharePoint, with permission-aware retrieval
- Identity provider (Entra ID / Okta) for SSO and to tailor answers to the user's device and groups
- Security Operations: route suspected incidents to the SOC queue or SOAR playbook
Data we need for a pilot
- Your knowledge base articles and runbooks (we help clean and tag them)
- 6 months of resolved tickets to measure coverage and tune hand-off rules
- Your priority/SLA matrix, assignment groups and escalation rules
Where it runs
- On-prem or private cloud; the knowledge base and conversations stay in your network
- Open-source model behind your own OpenAI-compatible gateway; embeddings optional (BM25 works out of the box)
- Every turn logged with retrieved articles, model, checker verdict and hand-off reason for audit
Open-source models we start with
- Qwen3-class 14B instruct model for intent, drafting and checking (JSON outputs)
- Optional bge-m3 / e5 embeddings for hybrid retrieval
Walkthrough
A short recorded walkthrough of the agentic it support desk with grounded answers and human hand-off: the input, the output and how a reviewer uses it.
Next accelerator: Loan document underwriting agent →Walkthrough video coming soon
We are recording a short walkthrough of this accelerator. Meanwhile, run the live demo above or book a working session.
Run the it operations & customer support accelerator on your data
In a working session we map it to your process, systems and rules, and agree what a pilot must prove before it goes live.