# AI IT Support Agents: Capabilities and Rollout

> An AI agent in internal IT support triages tickets, answers routine questions from your knowledge base, runs approved runbooks and prepares access requests. It stays safe through narrow permissions: broad read access, changes only through specific tested tools, and human sign-off for account or security changes.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-it-support-agent  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-11-15

Mid-sized companies often run their entire IT with a handful of people. The same small team handles laptops, accounts, printers, VPN, business applications and, somewhere in between, actual projects. Much of the incoming work repeats: a locked account, a missing share, a question about delegating a mailbox. An AI agent can take that first contact, sort it, answer what has a known answer and run routine steps, so the team spends its time on the problems that need expertise.

This guide is about **internal** IT support, where employees ask and your own IT answers. It covers tickets, self-service, user accounts and runbooks, with one question running through all of it: what is this agent allowed to touch? Unlike an agent answering customer emails, an IT support agent works right next to the keys of the company.

## What is an AI IT support agent?

An AI IT support agent is a language model that takes requests from employees via chat, email or a service portal, looks up information in your service desk, directory and knowledge base through APIs, and turns that into answers, tickets or prepared actions. It does not replace the IT team; it handles first contact and the routine work.

Two things set it apart from a scripted helpdesk chatbot. It understands requests in plain language, such as "Teams stopped working this morning but Outlook is fine", and works out what it needs to know. And it can call tools: check an account's status, open a ticket, start a runbook. Typical jobs:

- **Ticket intake and triage:** work out category, urgency and the affected system, ask for missing details like device name, error text and when it started, then route the ticket to the right person or queue.
- **Answers to routine questions:** setting up VPN, adding a printer, changing an email signature. The answer comes from your own solution articles, not from the model's general knowledge.
- **Steering people to self-service:** for a forgotten password, explain and link the existing self-service reset rather than changing the password itself.
- **Preparing access requests:** who needs what, which group grants it, who approves. The agent fills in the request; a person approves it.
- **Running runbooks:** fixed, tested procedures such as clearing a print queue, applying an approved mailbox permission or collecting diagnostics from a device.
- **Spotting incidents:** when ten people report the same thing within an hour, the agent opens one incident and alerts the team instead of creating ten separate tickets.

## How does it work under the hood?

The agent combines a language model, a set of narrowly scoped tools, a knowledge base of your solution articles and a permission layer that lives outside the model. The model chooses which tool to call; the target system decides whether that call is allowed.

The building blocks:

1. **Channel:** Microsoft Teams or Slack, a shared mailbox, or your service desk portal. The channel also provides the requester's identity. This matters: the agent must know who is asking from the sign-in, never from what the message claims.
2. **Model and instructions:** role, tone, limits, and the rules for handing over to a human.
3. **Knowledge base:** your solution articles, how-tos and operating procedures, indexed for search. Without it the agent gives generic advice that rarely fits your environment.
4. **Read tools:** search tickets, check account status, list group memberships, look up a device in inventory, check known incidents.
5. **Write tools:** create and comment on tickets, start a runbook, raise an access request. Each is a single, tightly defined tool with fixed parameters, never a general-purpose shell.
6. **Approval step:** anything that changes permissions, accounts or security settings becomes a proposal a person confirms.
7. **Audit log:** every request, tool call and approval is recorded so you can trace who triggered what.

The OWASP list of risks for LLM applications describes this shape as the defence against "excessive agency": few extensions, granular tools instead of open-ended ones, minimal permissions, running actions in the user's own context and human approval for high-impact actions, as set out in [OWASP LLM06:2025 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/). The key idea there is "complete mediation": authorisation is enforced in the downstream system, not left to the model's judgement.

## Where can it help in IT support?

The use cases differ mainly in how much the agent is allowed to change. Answering questions, triaging tickets and running runbooks sit on a rising scale of risk, and the demands on permissions, approvals and logging rise with them.

**Intake.** The agent receives requests, asks follow-up questions and opens complete tickets. It changes nothing in any system. This is the lowest-risk starting point, and it already gives the team tickets with device names, error messages and sensible categories.

**Knowledge base answers.** Where a solution article exists, the agent answers and links it. Where none exists, it says so and opens a ticket. The broader pattern of making internal knowledge searchable for an agent is covered on the [knowledge management agent](https://www.ruebstudio.de/en/ai-agents/knowledge-management) page.

**Account self-service.** Locked accounts and forgotten passwords are classic helpdesk work, and most directories already offer self-service for them. Microsoft Entra ID's self-service password reset, for example, lets users reset their password or unlock their account without an administrator or the helpdesk, with administrators choosing whether one or two authentication methods are required, according to [Microsoft's SSPR documentation](https://learn.microsoft.com/en-us/entra/identity/authentication/concept-sspr-howitworks). The agent's job is to guide people into that flow, not to set passwords itself.

**Access requests.** "I need access to the purchasing folder" is quick to type, but someone has to own the decision. The agent finds the right group and the responsible approver and raises the request. The business owner approves; the technical change can then run automatically.

**Runbooks.** For repeatable technical procedures, the agent can trigger a script that someone wrote and tested beforehand. The agent picks the runbook and fills in parameters; the script validates its inputs and runs with exactly the rights it needs.

**How it differs from related agents.** A [customer service agent](https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-customer-service-agent-b2b) answers external questions about orders, deliveries and returns, mostly by reading ERP data. An [agent for technical field service](https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-agent-technical-field-service) works with fault patterns on equipment at customer sites. An IT support agent serves your own staff and reaches into identities, devices and infrastructure, which is why permissions dominate this topic more than in either of the others.

## What do you need before you start?

You need a service desk with an API, a directory service, current solution articles, clear ownership of decisions and a permission model for the agent itself. If one of these is missing, fix that first; an agent only exposes existing gaps faster.

**Systems.** A service desk with an API, such as Jira Service Management, TOPdesk or ServiceNow. A directory such as Active Directory or Microsoft Entra ID that the agent can read account status and groups from. Optionally, a device management system for inventory and device health.

**Content.** Solution articles for your most common requests, current and written as clear steps. A practical starting point is the last three months of tickets grouped by topic: anything that comes up often and has one clear fix gets an article.

**Ownership.** Who approves which access? Which requests must never be handled automatically? Who picks up when the agent hands over? In many teams these answers exist only in people's heads. For an agent they have to be written down.

**A permission model.** Give the agent its own service identity, never an administrator's account. For every tool, define the rights it needs. Read access can be broader; write access should be as narrow as possible. OWASP also recommends executing actions in the context of the requesting user, so the agent can never do more for someone than that person could do themselves.

**Employee representation and privacy.** Logging employee requests creates personal data about how people work. Involve your data protection lead early, and in countries with works councils, such as Germany, the works council as well. This is not legal advice, but it is where rollouts most often stall.

## How do you roll it out step by step?

Roll the agent out in stages, each with more rights than the last: read access and ticket creation first, then knowledge base answers, then runbooks with approval, and only at the end a few actions without approval. Let each stage run for several weeks before moving on.

1. **Analyse your tickets.** Export recent tickets and group them. You will see what is frequent, what has a clear fix and what always needs an expert.
2. **Write solution articles.** Cover the frequent topics with clear fixes. Keep them short and step-by-step, using the menu and button names your staff actually see.
3. **Stage 1: intake.** The agent takes requests, asks follow-up questions and opens tickets. It has read access plus "create ticket". The team checks whether category and routing are right.
4. **Stage 2: answers.** The agent answers questions covered by an article and links it. Spot-check whether answers are correct and whether it hands over honestly when no article exists.
5. **Stage 3: runbooks with approval.** Connect individual tested procedures as tools. The agent proposes, a person confirms with one click, the script runs.
6. **Stage 4: selected actions without approval.** Only for procedures that ran cleanly in stage 3 for a long time, and where a mistake is minor and reversible, such as clearing a print queue.
7. **Keep tuning.** Review every handover. Is an article missing? Is a category unclear? Does the agent ask too few questions?

In our own operations, it has paid off to write down before each stage what success looks like and what would trigger a step back. That way the decision rests on criteria rather than gut feeling.

## What are the limits and risks?

The main risks are overly broad permissions, manipulation through injected instructions and social engineering around account recovery, plus wrong answers from outdated articles. All four can be reduced, none can be removed entirely.

**Too much access.** An agent with admin rights is a target. Anyone who can trick it gets an admin's reach. The remedy is the permission model above: narrow tools, with checks enforced in the target system.

**Prompt injection.** The agent reads text written by others: ticket descriptions, emails, attached log files. Instructions can be hidden in any of them. OWASP calls this indirect prompt injection and recommends least privilege, human approval for privileged operations and clear marking of external content, while noting that no method is foolproof, in [OWASP LLM01:2025 Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/). For IT support, the rule is simple: nothing in a ticket's text should ever lead directly to a permission change.

**Social engineering around accounts.** Attackers call helpdesks pretending to be employees to get passwords or MFA reset. In [advisory AA23-320A](https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-320a), CISA and the FBI describe the Scattered Spider group persuading IT helpdesk staff by phone to reset passwords or transfer MFA tokens, and recommend phishing-resistant methods such as FIDO/WebAuthn. An agent that restores accounts on request would be an ideal target for exactly this. So identity comes from the sign-in, recovery goes through the directory's own verified flows, and MFA resets stay with people.

**Stale knowledge.** If the VPN client changes and the article does not, the agent will confidently hand out wrong instructions. Make article updates part of every system change, and watch for requests where people open a ticket anyway after getting an answer.

**Transparency.** In the EU, Article 50(1) of the AI Act requires that people be told when they are interacting with an AI system unless that is obvious, see [Article 50](https://artificialintelligenceact.eu/article/50/); under [Article 113](https://artificialintelligenceact.eu/article/113/) the regulation generally applies from 2 August 2026. Even where no such rule applies, a clear label in the chat window avoids anyone thinking they are writing to a colleague.

**When it does not fit.** If your IT support is mostly one-off problems and projects rather than repeat requests, an agent adds little. The same goes if there are no solution articles and nobody has time to write them.

## How should accounts and recovery be handled?

Accounts are the most sensitive area because this is where help turns into a security hole. The agent should read and explain account status, but hand recovery over to directory processes that verify identity technically.

Microsoft's ecosystem offers a concrete example. For first sign-in or lost sign-in methods, Entra ID has the Temporary Access Pass, a time-limited code that can be single-use or reusable within its lifetime. According to [Microsoft's Temporary Access Pass documentation](https://learn.microsoft.com/en-us/entra/identity/authentication/howto-authentication-temporary-access-pass), the default lifetime is one hour and only specific administrator roles can issue one. Whoever issues it must reliably verify the person first, and that step should not be left to an agent alone.

A sensible split of work:

| Request | Agent | Human |
|---|---|---|
| Forgotten password | explains and links self-service reset | only if self-service fails |
| Locked account | checks status, explains cause | unlocks if needed |
| New phone, MFA lost | opens ticket, explains process | verifies identity, issues access |
| Access to a folder | finds group, raises request | business owner approves |

For accounts holding administrator roles, Entra ID already applies a stricter reset policy, as the SSPR documentation explains. The agent should route anything about privileged accounts to people by default.

## How do you get started?

Start by analysing your tickets and running intake with no write access to accounts or systems. Within a few weeks you will see how much of the load repeats, which solution articles are missing and whether your service desk categories hold up.

For the first weeks:

- identify your 20 most common request topics from recent months,
- write or update articles for the topics with a clear fix,
- define requests that always go to a person (privileged accounts, MFA, security incidents, HR matters),
- give the agent its own identity with read access and "create ticket",
- after four weeks, review: how many tickets were complete, and where did the agent misroute?

Only once that base is solid does it make sense to connect runbooks. For an overview of agents that work well across departments and how they fit together, see [AI agents for business departments](https://www.ruebstudio.de/en/ai-agents).

## Frequently asked questions

### Should an AI agent reset passwords?

It can, but it should not be the one deciding. Point employees to your self-service reset flow, which verifies registered authentication methods, and keep admin account resets and MFA resets with people.

### Do we need a new ticketing system?

No. The agent works through the API of the service desk and directory you already run. Clean categories, clear ownership and current solution articles matter more than the tool.

### Do employees need to be told they are talking to an AI?

In the EU, Article 50(1) of the AI Act requires people to be informed when they interact with an AI system unless it is obvious, and that includes internal use. Outside the EU, a clear label is still good practice; check the rules that apply to you.

### How is this different from a customer service agent?

A customer service agent mostly reads order and delivery data for external customers. An IT support agent serves your own staff and sits close to identities, devices and infrastructure, so its permissions carry far more risk.

## Sources

- [OWASP Top 10 for LLM Applications 2025: LLM06 Excessive Agency](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)
- [OWASP Top 10 for LLM Applications 2025: LLM01 Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/)
- [CISA/FBI Advisory AA23-320A: Scattered Spider (updated 29 July 2025)](https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-320a)
- [Microsoft Learn: Self-service password reset deep dive (Microsoft Entra ID)](https://learn.microsoft.com/en-us/entra/identity/authentication/concept-sspr-howitworks)
- [Microsoft Learn: Configure a Temporary Access Pass in Microsoft Entra ID](https://learn.microsoft.com/en-us/entra/identity/authentication/howto-authentication-temporary-access-pass)
- [EU AI Act, Article 50: Transparency Obligations](https://artificialintelligenceact.eu/article/50/)
- [EU AI Act, Article 113: Entry into Force and Application](https://artificialintelligenceact.eu/article/113/)
