How AI Is Changing IT Service Management (and What It Can't Do)

AI in IT service management is changing the parts of the work that are about language: understanding what a request is asking, routing it, finding the article that answers it, drafting replies, and letting people ask an assistant to do things ("request leave Friday") that it proposes and a human confirms. It isn't replacing work that needs access, judgment or accountability — fixing the laptop, approving the exception. Desks getting value treat AI as a helper around a system that still works without it.

What's actually changing

1. Intake: people describe problems in their own words

Traditional service portals ask requesters to choose from a catalogue of categories they don't understand. Language models let people just write "my screen flickers when I plug into the dock", then suggest the category and priority. The requester doesn't need to know your taxonomy; the desk still gets structured data.

How to do this reliably — schema-constrained JSON, validated, applied only above a confidence threshold — is in AI request classification with structured output.

2. Triage and routing

Classification feeds routing: the right team's queue, sooner. The biggest benefit isn't speed of the model; it's that fewer requests sit in a general inbox waiting for a human to sort them. The safe pattern is AI suggests, human (or a high-confidence rule) decides — covered in support ticket routing.

3. Knowledge that's findable

Semantic search matches meaning, not keywords, so "can't get on wifi" finds "Intermittent wireless connectivity". Shown during request creation, it answers some questions before they become tickets. See semantic search with Supabase pgvector and ticket deflection.

4. Drafting

First replies, summaries of long threads, knowledge articles from resolved requests. Drafts save typing; an agent still reads and sends. This is where models are strongest and risk is lowest.

5. Assistants that act — with confirmation

The newest shift: a chat assistant that can not only answer "how do I…?" but also read your workspace ("what's waiting on me today?") and propose actions ("create a task for Jordan due Friday"). LetRelay's assistant does this with a strict rule — every action is shown as an editable card and runs only when the user confirms, through the same server code and database permissions as the buttons on the screen. It can't do anything the user couldn't do themselves.

Anthropic's engineering guidance on agents makes a related point: the most successful implementations use simple, composable patterns rather than complex frameworks. For a service desk, one step — understand, propose, confirm — covers most of what people ask.

What AI can't do (yet, or safely)

• Physical and privileged work. Swapping hardware, granting production access, approving a policy exception. These need accountability.
• Be right without grounding. Asked about your organization without data, a model will guess. LetRelay's how-to answerer once invented a colleague's availability; it now has an explicit rule that it cannot see workspace data, and a separate, permission-checked path for questions that need data.
• Guarantee consistency. The same request can be classified differently on different days. Confidence thresholds, validation and evaluation sets manage this; they don't eliminate it.
• Stay available for free. Free tiers have daily limits — Groq documents 200,000 tokens per day per model on its free plan; Gemini's daily quotas reset at midnight Pacific. A desk that stops working when the quota runs out has made AI a dependency.

The risks to manage

Risk	Mitigation
Wrong classification	Confidence threshold; human triage below it; track corrections
Hallucinated answers	Ground in your knowledge base; say "I don't know"; evaluation sets
Actions nobody intended	Confirmation cards; same permissions as the UI; audit log
Data exposure	Keys server-side; model sees only data the user may see; no training on your data where the provider allows opting out
Outages and quotas	Save first; failover between providers; non-AI fallbacks
Cost creep	Rules and caches before model calls; small prompts; per-organization meters

The reliability side of this is the subject of the case for AI that degrades gracefully.

How to adopt AI on a service desk without regret

1. Instrument first. Capture created, first-response and resolved timestamps and categories for a month. Without a baseline, you can't tell whether AI helped. See help desk metrics that matter.
2. Start with suggestions. Category and priority suggestions shown to agents, not applied. Measure how often they're right.
3. Turn on auto-apply above a threshold once suggestions are right most of the time.
4. Add knowledge search during request creation. Track deflection as a trend.
5. Add drafting for agents.
6. Add an assistant with read-only tools first, then confirmable actions.
7. Keep a kill switch. Admins should be able to turn AI off for the organization without a deploy.

At each step, test the path with AI turned off.

How to tell whether AI is helping

Vendor demos show the best case. Your own numbers show the real one. Compare a month before and a month after each step, per category:

• Time to first human touch. If AI routing works, requests reach the right queue sooner, and the time until someone opens them should drop. If it doesn't drop, routing wasn't your bottleneck.
• Reassignments. Good classification means fewer requests change team. A rise means the model's categories don't match how your teams split work.
• Correction rate. How often an agent changes an auto-applied category or priority. Low and stable is the goal; a rise after a model or prompt change is a regression, whatever the release notes say.
• Deflection trend. Requests abandoned after viewing a suggested article, and answers to "did this help?".
• Assistant outcomes. Of the actions the assistant proposed, how many were confirmed unchanged, confirmed after editing, or dismissed. Many edits mean it understands intent but gets details wrong — usually dates, names or which task.

Keep a fixed evaluation set as well: fifty to a hundred real requests with the right category, and a few dozen assistant questions with required and forbidden content. LetRelay runs a 51-case router evaluation and requires every case to pass before switching models or prompts, because a change that looks better on five examples can be worse on fifty.

Cost, honestly

Model prices fall quickly, but costs come from volume and prompt size, and both grow quietly. The habits that keep AI affordable — on free tiers especially — are the same ones that make it reliable:

• Route obvious inputs with rules and never call a model for them.
• Cache answers to repeated product questions (never organization data).
• Keep prompts small; send only the guide pages or data the question needs.
• Meter per organization so one tenant can't spend everyone's quota.

On LetRelay's free keys, trimming the how-to prompt from 2,200–2,600 tokens to about 1,900, and skipping the model for obvious messages, measurably increased how many questions a day the assistant could answer before hitting provider limits.

What it means for the people on the desk

AI changes the shape of service desk work more than its size. Less time sorting and retyping; more time on the requests that genuinely need a person, and on writing the knowledge that makes the AI useful. The desks that benefit most are the ones that already had clean categories, a knowledge base and honest metrics — AI amplifies good structure and exposes missing structure.

FAQ

Will AI replace IT service desk agents?

It automates language tasks — classifying, searching, drafting — and some simple self-service. Work that needs access, judgment and accountability remains with people. Expect the job to shift toward harder requests and knowledge work.

What's the safest first AI feature for a help desk?

Category and priority suggestions shown to agents but not applied automatically. It's low-risk, easy to measure, and builds the data you need to decide whether to automate further.

Should an AI assistant be allowed to take actions?

Yes, if every action is proposed with editable details, executed only on explicit confirmation, and runs with exactly the user's own permissions.

Can a small team use AI in its service desk for free?

For modest volume, yes — free tiers from providers like Gemini and Groq are usable if you keep prompts small, fail over between providers, and design for the day the quota runs out.

Sources

• Anthropic — Building effective agents
• Atlassian — What is ITSM?
• Lewis et al. — Retrieval-Augmented Generation
• Groq — Rate limits
• Google — Gemini API rate limits