- Home
- AI agents
- Enterprise AI agent examples
- IT service desk
Plan · Enterprise AI agent examples
AI agents for the IT service desk: what they do and what companies report
An IT service desk agent answers employee IT questions, runs a defined set of fixes such as password resets, account unlocks and cache clears, opens and updates tickets, and hands the rest to a specialist with the context attached. It is one of the most documented enterprise agents in production because the work is repetitive, ticketed and measured.
In this guide: 83 pages
Key takeaways
- The published deployments with figures all run inside an existing channel (Microsoft Teams or a portal) and an existing ticketing system. None replaced the service desk; each sits in front of it.
- Measured results come from fixes the agent runs itself rather than from answers. CommBank reports a 2 minute fix against a typical 17 minute call, and Broadcom reports 88% of issues resolved without a person. Each figure is the company’s or the vendor’s own claim.
- Containment, deflection and resolution are three different numbers. Ask how each is defined before you set a target from someone else’s page.
- The controls that matter are an allow list of fixes, an identity that acts for the user on reads and as itself on fixes, approval for access changes, and a handoff that carries the context.
What do AI agents do on an IT service desk?
They take the tickets that follow a known path: answer a how-to question, reset a password, unlock an account, clear a cache, request access, open a ticket and report its status. The agent reads the knowledge base and the ticketing system, runs a fix from a list you approved, and hands anything else to a person.
Service desk tasks by what the agent is allowed to do. "Answers" means it replies from content; "assists" means it drafts or routes and a person acts; "acts" means it changes something itself.
| Task | Level | What it touches, and the control we add | Seen in |
|---|---|---|---|
| How-to and policy questions | Answers | Knowledge base articles, outage notices | Broadcom (3,000+ articles), Microsoft, Standard Bank |
| Password reset, account unlock | Acts | The identity directory; we verify the user first | Broadcom |
| Slow device: clear caches, reset a session | Acts | Endpoint tooling on the user’s device | CommBank (ChatIT) |
| Ticket creation and status | Acts | The ticketing system; we run this as the user | Broadcom, Grupo Falabella |
| Access and email group changes | Acts | Directory groups; we add approval by the resource owner | Broadcom |
| Link an issue to a known outage | Answers | Incident records | Microsoft Employee Self-Service Agent |
| Day-one onboarding support | Assists | Onboarding checklists, the HR system | Broadcom (during an acquisition) |
| Resolve an assigned L1 incident end to end | Acts | The ticketing system and runbooks | ServiceNow’s own help desk |
The level matters more than the task list. An agent that only answers needs read access and a good knowledge base. An agent that runs fixes needs its own identity, an allow list and a way to prove afterwards what it changed and for whom.
What have companies published about IT service desk agents?
Five organizations have published figures from live deployments, and every one is a vendor story or self-reported, so read them as claims rather than audited results. The table gives each figure as the source states it, who published it and how to read it.
Publicly reported IT service desk agents, as published by each source and opened on 2 November 2026. None of these companies is a QueryNow client.
| Company | What the agent does | Published figure | Source and date | How to read it |
|---|---|---|---|---|
| Commonwealth Bank of Australia (ChatIT) | Answers IT questions in Microsoft Teams and runs automated fixes such as clearing caches or resetting a session; hands off to a specialist when it cannot resolve the issue | More than 2.3 million messages; more than 12,000 automated fixes in six months; average fix time 2 minutes against a typical 17 minute call; satisfaction 4.63 out of 5 | Microsoft Source Asia, 1 July 2025 | Measured from internal logs and surveys; a vendor customer story on Copilot Studio and Azure |
| Broadcom | Moveworks assistant for about 15,000 employees: password resets, account unlocks, tickets and status, access and email group changes, a 3,000+ article knowledge base | 88% of support issues resolved without a person, up from about 10% at launch; more than 75,000 issues resolved | Moveworks customer story, undated, read 2 November 2026 | Vendor claim; the page does not date the rise from 10% to 88% |
| Microsoft (own IT) | Employee Self-Service Agent as the single front door for IT, HR and campus support: self-help, outage links, some fixes, handoff to a live agent | 30% reduction in tickets created; 50% of employees start with the agent, up from 27%; rolled out to 300,000+ people in 103 countries | Microsoft Inside Track, 8 October 2026 and 7 May 2026 | Vendor self-use; the 40% ticket reduction in the earlier article is a stated ambition rather than a result |
| ServiceNow (own help desk) | An L1 IT Service Desk AI Specialist takes assigned incidents and resolves them without a person | Resolves assigned IT cases "99% faster than human agents" | ServiceNow press release, 5 May 2026 | Vendor self-use; no case volume or resolution rate given |
| Grupo Falabella | A conversational agent for the support team that takes incident reports and creates tickets around the clock | Resolution times from 2 hours to 30 minutes; ticket creation from 16 minutes to 4 | Google Cloud use-case list, updated 22 April 2026 | A one-line vendor roundup entry; the depth of system actions is not described |
Standard Bank also appears in Microsoft’s July 2025 roundup with a Copilot Studio help desk bot that "now resolves 99% of all employee queries". The line gives no volume, period or definition of resolves, so it stays out of the table.
What controls does a service desk agent need?
Five, and they follow the level of the task. An agent that answers needs an identity and defined data access. An agent that runs fixes also needs an allow list, an approval step for anything that changes access, an evaluation set before release and monitoring after it.
- Identity. The agent reads the knowledge base and the ticket queue as the signed-in user, so it can never show a ticket the person could not open. Fixes run under the agent’s own identity with rights limited to the runbooks on the allow list. On Microsoft that is a Microsoft Entra Agent ID, which Copilot Studio creates for every new agent since July 2026; on AWS an AgentCore Identity workload identity; on Google a SPIFFE-based Agent Identity, whose API was still in preview on 2 November 2026.
- Data access. Name the systems: knowledge base, ticketing, endpoint management, the directory. The agent gets a connector to each with the narrowest scope that does the job, and an HTTP connector only to approved endpoints.
- Approval. Password resets and cache clears run without a person once the user is verified. Access and group changes, device wipes and anything irreversible wait for the owner of the resource. Microsoft’s own action taxonomy for agent testing puts irreversible actions in a "disclosure plus confirmation" tier.
- Evaluation. Before release, replay 100 to 200 real tickets from the categories in scope and score three things: did the agent pick the right path (answer, fix, ticket), did the fix work, and did it hand off when it should. A fast fix time means little if the reopen rate goes up.
- Monitoring. Trace every run with the ticket ID, the fix run and the result, and watch four numbers weekly: the share of tickets handled without a person, the reopen rate within seven days, satisfaction, and the handoff rate per category. Microsoft’s own team reports the share of employees who start with the agent, because a front door only pays when people use it.
Write the handoff as carefully as the fix. CommBank’s agent and Microsoft’s both pass the conversation to a person when they cannot resolve it, and the person gets the context, so the employee does not start again.
Where should you start?
Start with the ten ticket categories that make up most of your volume, answers first and fixes second, inside the channel people already use. Measure before the agent goes live, or the published figures will be the only ones you have.
- Pull a year of tickets and rank the categories by volume and by how often the resolution was the same runbook. Password, access and "slow device" usually top the list.
- Put the agent in Teams or your portal as an answer-only front door for those categories, with the knowledge base cleaned first. Record the baseline: time to resolve, tickets per 100 employees, satisfaction.
- Add the first automated fixes from an allow list of three to five runbooks, each with a verified user and a logged result. Replay last month’s tickets against them before release.
- Add the handoff with context to the ticketing system, and the outage check, so the agent stops answering questions an incident already explains.
- Review the four weekly numbers for a quarter before you add access changes, which need approval, or a second function such as HR.
How do Microsoft, AWS and Google implement it?
Where the published service desk agents ran, and what each cloud offers for the next one, with statuses as the vendor pages showed on 2 November 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Where the examples ran | CommBank, Microsoft and Standard Bank on Copilot Studio, with Teams as the channel | No service desk example with figures in the research set; Amazon Q Business is closed to new customers, with Amazon Quick as the successor | Grupo Falabella on Google Cloud AI (roundup entry); ServiceNow and Moveworks run as SaaS on their own platforms |
| Build the agent | Copilot Studio (low code, agent flows, computer use GA May 2026) or Microsoft Foundry Agent Service (pro code) | Amazon Bedrock AgentCore (Runtime, Gateway, Identity GA 13 October 2025); Bedrock Agents Classic is closed to new customers since 30 July 2026 | Gemini Enterprise (no-code Agent Designer) or Gemini Enterprise Agent Platform with ADK and Agent Runtime |
| Identity for the agent | Microsoft Entra Agent ID (GA); created automatically for Copilot Studio agents since July 2026 | AgentCore Identity workload identity; on-behalf-of token exchange since April 2026 | Agent Identity (SPIFFE-based); API in preview since 18 June 2026 |
| Allow list and approval on fixes | Copilot Studio agent flows with approvals; Foundry tool call guardrails (preview); Agent Framework per-tool approval | AgentCore Policy, Cedar rules on every tool call (GA 3 March 2026) | Agent Gateway as the policy point (GA 18 June 2026); ADK tool confirmation (experimental) |
| Ticketing and endpoint connectors | Microsoft 365 Copilot connectors (ServiceNow among 100+ prebuilt) and Power Platform connectors | AgentCore Gateway wraps APIs, Lambda functions and MCP servers as tools | Gemini Enterprise connectors; Agent Registry holds MCP servers and tools |
| Evaluation and monitoring | Copilot Studio agent evaluation (no preview label on 5 October 2026); Foundry tracing GA for prompt and hosted agents | AgentCore Evaluations (GA 31 March 2026); CloudWatch generative AI observability (GA 13 October 2025) | Agent Observability (GA 18 June 2026); Gen AI evaluation service |
ServiceNow (AI Agent Orchestrator and AI Agent Studio GA 12 March 2025) and Moveworks are the SaaS route. Both sit on the ticketing system you already run, and they were the platform in two of the five examples.
Frequently asked questions
Is "resolved without a person" the same as deflected?
No. Deflection counts contacts that never became a ticket, resolution counts issues the agent fixed, and containment counts conversations that ended without a handoff. Broadcom’s 88% is described as resolved, Microsoft’s 30% is tickets not created. Ask each source which one it measured before you compare them.
Should the agent run fixes under the user’s identity or its own?
Both, by task. Reads and ticket updates run as the user, so the agent can never see more than the person could. Fixes that need elevated rights, such as a cache clear on a managed device, run under the agent’s own identity with rights limited to the allow list, and each run is logged with the user it was for.
How many tickets does a service desk need before an agent is worth it?
No published threshold exists. Microsoft’s deployment serves 300,000 people, CommBank’s runs across a large bank, and Grupo Falabella’s figures come from one support team. The better test is concentration: if ten categories make up most of your volume and the resolution is usually the same runbook, an agent has work to do.
Can one agent cover IT and HR?
Microsoft runs one front door for IT, HR and campus support, and its published figures are IT-specific. Start with one function so the allow list, the owner and the measurement stay clear, then add HR as a second connected agent rather than a longer prompt.
What happens when the agent cannot fix the issue?
It hands off to a specialist with the conversation and what it already tried. CommBank and Microsoft both describe this handoff, and in our experience the handoff rate per category is the earliest signal of where the knowledge base or the allow list is thin.
More in Enterprise AI agent examples
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- CommBank’s ChatIT helps employees fix IT issues in seconds using generative AI
- How Broadcom powers its knowledge base with AI to help employees help themselves
- Building trust in AI-powered support at Microsoft: our Employee Self-Service agent story
- Transforming IT support across Microsoft with the Employee Self-Service Agent
- ServiceNow brings Autonomous Workforce to every major business function
- 1,302 real-world gen AI use cases from the world’s leading organizations
- AI-powered success with more than 1,000 stories of customer transformation and innovation
- What’s new in Copilot Studio
- Evaluate agents in Copilot Studio
- AI Red Teaming Agent
- Policy for Amazon Bedrock AgentCore is now generally available
- Amazon Bedrock AgentCore is now generally available
- Amazon Q Business availability change
- Gemini Enterprise Agent Platform release notes
- ServiceNow’s latest platform release adds to thousands of AI agents across CRM, HR, IT and more
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner