Stabilize, Coach, and Automate: Preparing the Service Desk for AI
AI agents are moving out of the testing phase and onto the service desk floor. The question is no longer whether we can build them, but how we manage and govern them without breaking our operations.
I have managed service desk teams for about 15 years. Over that time, I have watched trends come and go, but the core mechanics of a successful support team never really change. As we prepare to bring artificial intelligence agents into our daily workflows, managers need to rely on the same habits that make a traditional service desk work. We need to stabilize our queues, anchor our feedback to specific tickets, separate a final result from the behavior that caused it, and require a human check on anything risky.
Part 1, Get the desk ready first (the Queue Stabilizer)
Before you can introduce any new technology or process, you have to clean up the environment. You need to quarantine your old backlog. Stop the habit of cherry-picking easy tickets from the queue, and give your technicians some actual breathing room. When I stepped into a new role, I inherited a service desk with nearly 150 aged tickets. The team of four technicians had low morale, and aging requests were simply left to wither. Knowing end users were so frustrated that they had given up responding, I generated reports on the backlog and established a "soft close" protocol. Working together, we updated and resolved every ticket older than 14 days. That effort completely transformed our operations: today, open tickets consistently stay under 30, the backlog remains under 10 days old, and our CSAT score averages 98.6%.
Coaching and new tools simply do not stick on an unstable queue. If a technician is taking forty calls a day while staring at a backlog of two hundred aging requests, they are in survival mode. They will not absorb a new coaching framework, and they certainly will not have the patience to interact with a new AI agent properly. You have to stabilize the daily flow of work first. Put the aging tickets into a special project status, assign a dedicated team member to clear them out, and let the rest of the desk focus on today's intake.
Part 2, Coach from the ticket, not the scorecard (Ticket-Level CSAT Loop)
In my current role I rolled out a CSAT survey platform integrated with our ticketing platform, where our team had neither before, and I pull survey comments into the ticket as private notes so I can coach from them.
Do not read average CSAT scores aloud as a team scoreboard. It invites defensiveness and encourages people to game the system. Atlassian documentation for Jira Service Management notes that individual agent satisfaction can be compared with team scores to spot who may need training. My approach adjusts this common practice. I keep the comparisons private and tied strictly to the work done inside a specific ticket. A service desk CSAT benchmarking article from Myra argues the single average score is a headline number that is good for trends but weak for diagnosis. The real value for managers lives inside the written comments on the one and two star ratings.
You have to separate the rating from the behavior. Consider the false negative. This is a one star survey that arrives during a major network outage, even though the technician stayed calm, escalated correctly, and documented everything perfectly. You need to reinforce that positive behavior despite the bad score. Summary: An upgrade to the company's SaaS accounting system led to posting issues and a influx of service desk tickets due to limited prior notification. Because resolution depended on the vendor, user frustration grew, resulting in several 1-star CSAT survey ratings from the incident..
Then consider the false positive. This happens when a technician gets a glowing five star review after they broke a major security policy (like granting temporary local admin rights) just to please the user. You have to coach that behavior immediately. The end result was a happy user, but the behavior behind it put the entire company at risk.
Part 3, The 3-Question Diagnostic Script for a 10-minute one-on-one
When you sit down to review these tagged tickets in a short private weekly one-on-one, you need a reliable method to figure out what went wrong. I use a simple diagnostic script.
First, I ask: "Walk me through what happened on this ticket from your side of the screen." This opens the conversation without making it an attack.
Second, I ask: "What was the ripple impact, on the user, on the SLA, or on the teammate who caught the escalation?" This forces the technician to look past their own silo and see how their actions affect the larger organization.
Third, I ask: "If this exact problem hits your queue tomorrow at 9:00 AM, what is the one move you'd try differently?"
The answers to these three questions will tell you exactly what kind of problem you have. They separate a skill gap (which is highly teachable) from a will gap (which requires an accountability conversation). If they do not know what to do differently, you have a training opportunity. If they know exactly what they should have done but chose not to do it, you have an accountability issue.
Part 4, The Knowledge Flywheel
Once you identify a skill gap and fix it, you need to make sure the rest of the team learns from it. You have to turn each lesson into a five-bullet micro-SOP (Standard Operating Procedure) and share it.
We do this so one technician's learning prevents repeat tickets across the entire desk. Nobody is going to read a ten page manual while a frustrated user is on hold. A five bullet summary is fast, actionable, and easy to reference. This builds a knowledge flywheel where the team constantly gets smarter based on real interactions, which creates the perfect foundation for automation.
Part 5, What this looks like when AI agents join the desk
The strongest bridge between traditional management and modern automation is this: a technician who grants temporary local admin rights to earn a five star rating is the exact same risk as an AI agent that is allowed to take high risk actions on its own. Both need strict guardrails, and both need a person reviewing the risky decisions.
We are seeing this reality take shape in the market right now. ServiceNow's Zurich release documentation describes connecting Now Assist in Virtual Agent to Microsoft Copilot through the Teams conversational integration. According to this documentation, users can get answers and take actions from Copilot, and Copilot can hand requests to Now Assist in Teams. Employees can reach knowledge, catalog requests, and ticket status from inside Microsoft Teams, with a seamless handoff to a live agent when the AI cannot resolve it. The documentation lists a Now Assist license as a prerequisite, and a third-party analysis reported by S2 Labs states that using Now Assist inside Teams needs both a Now Assist license and a Microsoft 365 Copilot license. Readers should confirm all licensing requirements directly with both vendors.
As these tools connect, we have to split our workflows. Low risk actions (like a password reset or a software version check) can run on their own. High risk actions (like granting local admin rights or approving major hardware purchases) should pause for human approval. This mirrors the five star false positive from Part 2.
Governance tools exist to track which agents exist, what they can access, and what they did. According to a ServiceNow press release via Business Wire, ServiceNow announced new integrations with Microsoft at Microsoft Ignite on Nov. 18, 2025. This included an integration with Microsoft Agent 365, noting that AI Control Tower will connect with Microsoft Foundry and Copilot Studio to govern agents across platforms. Microsoft describes Agent 365 as a control plane that gives admins a central registry of agents and extends Entra, Purview, and Defender to agents. Some of these integrations will leverage MCP, a common standard that lets AI tools connect to business systems, to share data securely.
You should test an agent with a set of realistic cases and review its conversation traces before it goes live, the exact same way I review tickets before trusting a new human process. Clean knowledge articles and accurate CMDB data matter immensely because these agents will answer directly from them.
What I would do this quarter
If you are preparing for these tools, here are five plain steps you can take right now.
Stabilize the queue by quarantining your old backlog.
Start adding ticket-level CSAT notes for private review.
Run the three diagnostic questions weekly with your team.
Write and share five-bullet micro-SOPs from those coaching sessions.
List which service desk actions you would never let automation do without human approval.
Closing
The honest takeaway here is that AI will not fix a broken service desk process. It will only make a broken process execute much faster. If you build strong human habits now, governing an AI agent later becomes a familiar task rather than a frightening one.
How do you handle coaching on your team, especially when a good CSAT score hides a bad technical decision?
Sources
ServiceNow advances enterprise AI through seamless integrations with Microsoft (Business Wire): https://seekingalpha.com/pr/20312287-servicenow-advances-enterprise-ai-through-seamless-integrations-with-microsoft-enabling
ServiceNow Docs (Zurich Release): https://www.servicenow.com/docs/r/BfFhnVBqpVWNfzvxxnpNzw/w0PKuUX3pNksLjJDmI6R3w
ServiceNow Now Assist vs Microsoft Copilot (S2 Labs): https://s2-labs.com/blog/servicenow-now-assist-vs-microsoft-copilot/
Microsoft Agent 365 Overview (Microsoft Learn): https://learn.microsoft.com/en-ca/microsoft-agent-365/overview
Collecting customer satisfaction feedback (Atlassian): https://confluence.atlassian.com/servicedeskcloud/collecting-customer-satisfaction-feedback-790791332.html
CSAT Benchmarks for IT Service Desks (Myra): https://myra.meta-inf.hu/blog/csat-benchmarks-it-service-desks
(Note: Product details, integrations, and licensing requirements change quickly. Readers should confirm current specifications directly with their vendors. Certification paths are also changing in 2026, so please check Microsoft Learn before choosing an exam).
