AI for IT Support: What It Fixes and What It Doesn't

AI for IT support works best on the boring 40% of your queue: password resets, access requests, VPN issues, software installs, "how do I" questions with a documented answer. It works poorly on anything requiring diagnosis across systems, and it fails badly on tickets where the user's description of the problem is wrong. Knowing which pile a ticket lands in is most of the implementation.

The short version:

• Deflection is real but the headline numbers are marketing. Vendors publish figures from their best deployments. Plan for a fraction of the advertised rate in year one.
• The value shows up in resolution time before it shows up in headcount. Triage and routing improve first because they don't require the AI to be right, only to be close.
• Your knowledge base is the constraint. An AI support agent is a retrieval system with a personality. If the answer isn't written down anywhere, it will invent one.

The four things AI actually does in IT support

"AI for IT support" covers four distinct capabilities that get sold as one product. They have very different risk profiles.

Start at the top of that table and work down. Triage is where most teams see a return in weeks with almost no downside, because a mis-routed ticket was already a normal occurrence. Autonomous resolution is where the integration work, the permission model, and the audit trail all have to be right before you turn it on.

This is the same escalation ladder we describe for IT process automation generally — the difference with AI is that the middle rungs became viable, not that the ladder changed.

Ticket types, sorted by how well AI handles them

Handles well — documented, single-system, verifiable:

• Password resets and MFA re-enrolment
• Group and application access requests that follow an approval rule
• Software installation and licence assignment
• VPN and Wi-Fi connectivity walkthroughs
• Policy questions with a written answer ("what's the laptop refresh cycle")

Handles partially — useful draft, needs a human:

• Hardware failure triage, where the AI narrows it and a technician confirms
• Onboarding and offboarding checklists spanning several systems
• Recurring incident summaries and post-incident write-ups

Handles badly — don't start here:

• "It's slow" — no defined symptom, no single system
• Anything involving a user who has already tried three fixes and changed state
• Security incidents, where being confidently wrong is the worst outcome
• Anything where the resolution requires physical access

The dividing line is the same one that governs agents generally: can the outcome be checked cheaply? A password reset either worked or it didn't. "Your network is fine now" is unverifiable, and that's why it's a bad candidate.

What the deflection numbers actually mean

Vendors across the ITSM market publish deflection rates — the share of tickets resolved without a human. Freshservice and Jira Service Management both ship AI agents with published customer outcomes, and the figures reach impressive levels in mature deployments.

Read them with three caveats:

1. Denominator games. Some counts include chatbot sessions that never would have become tickets. A user asking "what's the wifi password" in a chat widget was not a ticket you were paying to resolve.
2. Best-case customers. Published numbers come from reference accounts with clean documentation and deep integrations — not from a median deployment in month two.
3. Deflection isn't resolution. A deflected ticket that comes back tomorrow with an angrier user is worse than a ticket handled once.

The number worth tracking internally is different: the share of tickets closed on first contact without escalation, measured 7 days later. That catches the boomerang cases, and it's the metric that actually correlates with support cost.

Getting it working without a six-month project

The failure pattern is predictable — teams buy a platform, discover their documentation can't support it, and end up with an expensive chatbot that apologises a lot. A sequence that avoids that:

1. Pull your top 20 ticket categories by volume. The long tail is a distraction; the top 20 usually covers most of the queue.
2. Check which ones have a correct, current, written answer. Usually fewer than half. Fixing that is the highest-leverage work in the entire project, and it pays off even if you never deploy AI.
3. Turn on triage first. Let it categorise and route for a few weeks while you measure accuracy against what humans did.
4. Add draft assistance next. Your agents will tell you quickly whether the drafts are usable.
5. Enable autonomous resolution on exactly one workflow. Password reset is the standard first choice because it's high volume, verifiable, and reversible.
6. Instrument the boomerang rate before you expand.

Steps 1 and 2 are unglamorous and they determine the outcome. Every deployment that disappoints skipped them.

For the automation layer underneath — the part that actually resets the password or provisions the account — the tooling overlaps heavily with general IT automation tools. AI decides what to do; something else has to do it.

Where this fits alongside your existing stack

Most teams already run some automation: Ansible for configuration, a ticketing system with rules, scripts that nobody has touched in two years. AI doesn't replace that layer. It sits in front of it, turning an unstructured user request into a structured call against automation you already have.

That framing matters for buying decisions. If your automation layer is thin, AI has nothing to invoke and you get a well-spoken FAQ. Practices from ITIL — a defined service catalogue, categorised requests, documented resolution paths — turn out to be the prerequisite, which is an unexpected result for a technology sold as replacing process work.

If you're evaluating vendors, our guide to what to ask in an AI virtual agent demo covers the questions that separate a real deployment from a scripted one.

For smaller teams without an ITSM platform at all, the practical path is different: assemble the pieces from AI apps and workflows rather than buying a suite. Taku mirrors a working AI setup into a desktop workspace and runs it there, so you can put a proven support workflow to work without building the environment around it first. Taku is in Beta, and the Mac app is available now.

Key points

• AI handles documented, single-system, verifiable requests well and diagnostic work poorly.
• Deploy in order: triage, then draft assistance, then self-service, then autonomous resolution.
• Published deflection rates come from best-case deployments; measure first-contact resolution at 7 days instead.
• Your knowledge base quality sets the ceiling — fix that before buying anything.
• AI decides; your existing automation executes. A thin automation layer limits what AI can deliver.

FAQ

How much of an IT support queue can AI realistically handle?

Vendors publish high deflection figures from mature deployments. A more useful planning assumption is the share of your top 20 ticket categories that are documented, single-system, and verifiable — for most teams that's a meaningful slice, not a majority, and it grows as documentation improves.

Does AI for IT support replace service desk staff?

It changes the mix more than the count. Tier-1 volume drops while escalations, root-cause work, and the effort of maintaining knowledge and integrations grow. Teams that redeploy staff toward the second category get more out of it than teams that cut.

What's the difference between an AI chatbot and an AI agent for IT support?

A chatbot retrieves an answer. An agent takes an action — resetting a password, provisioning access, restarting a service — which requires integrations, permissions, and an audit trail. The second is considerably more valuable and considerably more work to deploy safely.

What should we fix before deploying AI for IT support?

Documentation for your highest-volume ticket types, and a defined service catalogue. An AI support agent answers from what's written down; if the answer doesn't exist, it will produce something plausible instead of admitting the gap.

Can small teams use AI for IT support without an ITSM platform?

Yes, though the path is different. Rather than buying a suite, small teams generally get further composing existing AI apps and automation around their current helpdesk or shared inbox.