What Happens When AI Handles 90% of Your Helpdesk Tickets

I started in IT support taking L1 tickets at a plywood manufacturer in Nairobi. Password resets. VPN access. Someone's mouse wasn't working. Someone else couldn't find a document they'd definitely saved to the wrong folder. The work was repetitive in a way that felt useless at the time and turned out to be foundational.

In February 2026, ServiceNow announced that their internal L1 Service Desk AI Specialist was handling more than 90% of employee IT requests autonomously — and doing it roughly twice as fast as human agents on the cases it touches. I've been sitting with that number ever since, because I know exactly what those tickets feel like, and I have a view on what the number actually means.

What ServiceNow Actually Shipped

The ServiceNow Autonomous Workforce is not a chatbot with escalation logic bolted on. It's a set of AI specialists — the first being the L1 Service Desk AI Specialist — designed to hold a defined job role inside a governed workflow, not just answer questions at an endpoint.

The L1 agent does what a human L1 agent does: it receives a request, diagnoses it using the enterprise's knowledge base and historical incident data, decides the right action, and executes — end to end, without a human in the loop. Password resets, software access provisioning, network troubleshooting where the resolution is lookup-and-restart. ServiceNow has been running it internally and reporting that it's handling the majority of employee cases. General availability for external customers was targeted for Q2 2026.

That's not marketing copy for a chatbot with canned responses. It's a structural shift in what a helpdesk team needs to staff — and IT departments that aren't planning for it are already behind.

The Numbers Worth Interrogating

90% handled autonomously. These are ServiceNow's own figures for their own internal deployment, and they deserve some scrutiny before you start making headcount decisions based on them.

"90% of employee requests handled" is a carefully chosen metric. It reflects volume — a large proportion of routine, well-defined requests that the org generates every month. It doesn't tell you what happens with the 10% that escalates, how escalations are handled, or how the system performs at an org that doesn't have clean, well-governed ServiceNow data to begin with. An org with messy, inconsistently documented infrastructure is going to see a very different number.

The speed claims are also context-dependent. A password reset that used to take eight hours to close — because of email queue backlog, not because a human spent eight hours on it — now closes in seconds. That's a real improvement. It's not the same as saying a human agent working actively on the same ticket would have taken eight hours.

I'm not dismissing the results. They're meaningful. But if you're a manager reading a vendor's internal benchmark, the right question is: what does "handled" mean, and how does the system fail? Those answers shape whether you're looking at a 90% displacement or a 40% one in your specific environment.

What It Actually Replaces

The honest answer: the well-documented, genuinely repetitive subset of L1 work. In my early support roles, I'd estimate roughly 40–50% of tickets by volume fell into that category — possibly higher at a large, mature enterprise with clean ITSM data, possibly 20–30% at a fast-growing company still sorting out its processes.

The AI is very good at the tickets that follow a script. It doesn't degrade at 3 AM. It doesn't get impatient on the eighth similar ticket of the day. On the requests that are well-defined and well-documented, it is genuinely faster and more consistent than a human doing the same thing for the fourth time that shift.

What it's not good at — yet — is the contextual layer. The user who's embarrassed about a basic question and is underselling the problem. The manager whose all-hands is in fifteen minutes and whose laptop suddenly won't connect to the projector. The onboarding call where you realise, halfway through, that the new hire has been provisioned into the wrong Salesforce instance and nobody caught it in the workflow.

These aren't exotic edge cases. They happen constantly. And they're exactly the situations that determine how IT is perceived inside an organisation.

What the AI Still Needs Humans For

There's a layer of work that expands specifically because AI agents are running L1: the configuration and maintenance work that makes them useful in the first place.

Someone has to write the knowledge base articles the agent draws from — specific enough that an AI can act on them, not the vague three-line entries most helpdesks actually have. Someone has to define what "resolved" means for each request type, build the approval flows for access provisioning, review the escalations the agent couldn't close and understand why. Someone has to catch it when the agent confidently does the wrong thing because the knowledge base was outdated.

That work is expanding, not contracting. It requires people who understand both the underlying IT systems and how the agent is reasoning about them — a skill set that barely existed as a job description two years ago.

The people who will do well in this shift are not the ones who can do what the AI does. They're the ones who can work a level above where the AI tops out — and who can diagnose when the AI is confidently wrong.

Where the Real Risk Lives

The risk isn't that AI displaces skilled IT professionals. The more concrete risk is that it removes the entry-level jobs that were historically the training ground for them.

L1 tickets are tedious by year three. In month three, they're educational. You learn what users actually struggle with. You develop an instinct for what "it's not working" really means when you can't see the screen. You build the judgment to know when something that looks like a password issue is actually an Active Directory replication problem that needs to go to infrastructure.

If L1 is fully automated, where does a junior IT professional develop that judgment now? Some will find it in the review and maintenance work described above. Some will need organisations that deliberately create learning paths that don't route through a ticket queue. Most organisations haven't thought this through yet, and in three to five years it will show.

The other real risk: IT managers who see "90% of L1 tickets handled" and immediately cut headcount without thinking about what the remaining 10% requires, or about the upstream configuration work that makes the whole system function. The ones who treat this as a headcount calculator rather than a skill-mix problem are going to have a worse helpdesk in two years, not a better one.

What to Do If You're in IT Support Right Now

Stop worrying about whether L1 automation will make you redundant in the next six months. Unless your entire role is password resets at an org that already has clean ITSM governance and an active ServiceNow deployment, your timeline is longer than the headlines suggest. The technology is genuinely new, implementation takes time, and the gap between "it works in ServiceNow's own environment" and "it works in your mid-market company with three overlapping ticketing systems" is significant.

What you should actively be doing:

• Build above the ticket. Learn to build and maintain the workflows the AI executes inside. Learn to write runbooks that are specific enough for an automated agent to follow — most existing documentation isn't.
• Document your contextual knowledge. The stuff in your head about which infrastructure quirks are recurring, where the undocumented dependencies live, which stakeholders need which communication style — write it down and own it. AI doesn't have it. Your org increasingly needs it in a form that survives beyond your memory.
• Get hands-on with agentic tools. If you haven't built a workflow in n8n or wired an AI tool into your ticket queue on a test environment, do it. Not because it will save your job, but because it will show you exactly how these systems fail — and understanding failure modes is where the human judgment still lives.
• Expand your scope if you can. The people at real risk are those whose entire role is reactive L1 work with no room to improve the systems around them. If that's your situation, make the case to expand it or find an org where that scope exists. Proactivity, not reactivity, is the durable skill.

The Actual Shape of the Shift

The ServiceNow announcement is real and the internal results are significant. But "AI handles 90% of L1 requests" is not the same as "AI replaces 90% of IT headcount." The task layer shrinks on well-defined, high-volume request types. The judgment, configuration, escalation, and relationship layers stay — and in most organisations, those layers were already understaffed relative to the size of the ticket queue.

The IT professionals most likely to struggle aren't the ones who close a hundred tickets a day. They're the ones who spent those years closing tickets without ever learning why the tickets happened. The ones who treated L1 as a waiting room for a promotion rather than an information-rich source of data about organisational dysfunction.

The ones who were always doing both — closing the ticket and asking what system failure produced it — are going to find that the AI handles the first part and their scope just got bigger.

Thinking Through What This Means for Your Team?

If you're figuring out where automation realistically fits in your support setup — and what the honest gaps are between vendor benchmarks and your actual environment — I'm happy to talk through it. The gap between the demo and the deployment is usually where the real decisions live.