How to use ServiceNow AI Agents for productivity

Last quarter, our IT service desk was drowning. We had a backlog of over 200 routine incidents—password resets, software provisioning requests, access approvals—that were technically simple but required a human to click through four different systems to resolve. My team was spending hours acting as bridge traffic between ServiceNow, our Active Directory, and our licensing portals. I kept thinking, "There has to be a way to make the platform just do this itself."

I'd played around with ServiceNow's Virtual Agent chatbots before, but they always felt like glorified decision trees. If a user went off-script, the bot fell apart. Then I got access to the Yokohama release with Now Assist and started experimenting with AI Agents. The difference was night and day. Instead of rigid conversation flows, these agents can actually plan, reason, and take actions across the platform autonomously.

Here's how I got my first AI Agent up and running, the mistakes I made along the way, and what I actually saw it do.

The Prerequisites: Getting Your Instance Ready

Before we build anything, you need the right foundation. I learned this the hard way—my first attempt failed silently because I skipped a dependency. Here's what you actually need:

A Now Assist License (Pro Plus or Enterprise Plus)
Instance version: Zurich Patch 2+ or Yokohama Patch 8+ (always grab the latest patch recommended for your release)
Store Apps installed and updated:
- AI Agents store app
- The relevant "Now Assist for..." apps for your use case (e.g., Now Assist for ITSM)
- Critical step: When installing, make sure you select "Load demo data." This pre-configures sample agents and skills that make learning infinitely easier.

If you run into plugin version conflicts (I did, and it was frustrating), try repairing these specific plugins in order:

Flow Designer (ServiceNow Product - 20 plugins)
Microsoft Azure OpenAI Generative AI Spoke (sn_azure_openai)
Generative AI Controller (sn_generative_ai)
Now Assist AI Agent (sn_aia)
Now Assist for [Your Product] (e.g., sn_itsm_gen_ai for ITSM)

Don't skip the demo data. I tried to build everything from scratch initially and spent two hours reinventing the wheel before realizing the demo data gave me perfectly functional templates to study.

Understanding How AI Agents Actually Work

Here's the mental shift that took me a while to internalize: AI Agents are not chatbots. A traditional chatbot follows a script—you say X, it says Y. An AI Agent is more like a new hire who has access to your tools and can figure out how to use them.

On the Now Platform, an AI Agent operates through three core behaviors:

Planning: It breaks down a complex request into logical steps
Reasoning: It evaluates context and decides what to do next based on the situation
Acting: It executes tasks using "Skills"—which are essentially tools that connect to your workflows, integrations, and data

The magic is that the agent decides which skills to use and when. You don't hardcode the sequence. You give it tools and a goal, and it figures out the path.

Building My First Agent in AI Agents Studio

Alright, let's get hands-on. I wanted to build an agent that handles the most common complaint in our org: "I can't access the software I need."

Step 1: Open AI Agents Studio
Navigate to Now Assist > AI Agents Studio in your application navigator. The studio gives you a clean interface to create agents and assign skills to them.

Step 2: Create a New Agent
I clicked "Create Agent" and filled in the basics:

Name: Software Access Assistant
Description: Helps employees request and get provisioned for standard software applications
Conversation mode: I chose "Autonomous with human approval for sensitive actions" — more on why in a minute

Step 3: Assign Skills
This is where the real work happens. Skills are the actions your agent can take. Out of the box with the demo data, ServiceNow provides pre-built skills for common ITSM, HRSD, and CSM scenarios.

For my software access agent, I assigned these skills:

Check User Access: Queries the CMDB to see what software the user already has
Create Catalog Request: Submits a request via the service catalog
Check Approval Rules: Determines if the software needs manager approval
Notify Requester: Sends a notification to the employee about the status

The key insight: I didn't tell the agent in what order to use these skills. I just gave it the toolkit. The agent reasons through the situation. If a user already has the software, it tells them. If they need approval, it routes for approval. If it's auto-approved, it provisions directly.

Step 4: Configure Guardrails
This is where I made my second mistake. Initially, I set the agent to fully autonomous—no human in the loop. Within five minutes of testing, it tried to deprovision someone's Adobe license because it misread a request. Lesson learned.

I went back and configured the agent to require human approval before executing any skill that modifies access or creates a financial commitment. The platform makes this straightforward—you can mark individual skills as requiring confirmation. So "Check User Access" runs freely, but "Create Catalog Request" pauses and asks a human, "I'm about to request Photoshop for this user. Approve?"

This is what ServiceNow means by "AI agents work with you, not just for you." The governance controls are granular and practical.

Testing It Out: What Actually Happened

I opened the Virtual Agent chat in my test instance and typed: "I need Microsoft Project for a client engagement starting next week."

Here's what the agent did—on its own:

It checked my user record and saw I didn't currently have Microsoft Project
It checked the catalog and found the software item
It checked approval rules and determined Microsoft Project requires manager approval for my department
It created the catalog request and routed it to my manager
It sent me a message: "I've submitted a request for Microsoft Project. Since this requires manager approval, I've routed it to your manager. I'll notify you once it's approved and provisioned."

The whole interaction took about 15 seconds. Previously, this would have been a 20-minute back-and-forth between me, the service desk, and my manager.

What impressed me most was the reasoning. When I tested with a different user who already had Microsoft Project, the agent didn't just blindly submit another request. It said, "Looks like you already have Microsoft Project assigned. Are you having trouble accessing it, or do you need an additional license?" That's contextual problem-solving, not scripted response.

Practical Tips from the Trenches

After spending a few weeks with AI Agents, here's what I wish someone had told me on day one:

Start with the out-of-box agents. ServiceNow ships pre-configured agents for IT, Customer Service, and HR use cases. Clone one of these and modify it rather than building from zero. You'll learn the patterns faster.

Name your skills descriptively. The agent uses skill names and descriptions to decide which tool to pick. If you name a skill "Action 1," the agent has no context. Name it "Reset Active Directory Password for User" and the agent knows exactly when to deploy it.

Test with real-ish scenarios, not happy paths. My agent handled "I need software" beautifully. It struggled with "My boss said I should have gotten Project last month but it never showed up." Complex, multi-part requests reveal where your skill set has gaps.

Use the analytics. AI Agents Studio includes dashboards showing how often agents intervene, where they escalate to humans, and which skills get used most. This data is gold for figuring out where to add new skills or adjust guardrails.

Honest Limitations

Let me be real about what's not perfect:

You need clean data. The agent's reasoning is only as good as your CMDB and catalog data. If your software catalog is a mess of duplicates and stale entries, the agent will make confusing recommendations. Clean your data first.
Complex multi-system workflows still need Flow Designer. If a resolution requires orchestrating across five external APIs with error handling and retries, build that in Flow Designer and expose it as a single skill. Don't make the agent micromanage complex integrations.
There's a learning curve on skill design. Figuring out the right granularity—what's one skill versus two—took iteration. Too granular and the agent gets confused by options; too broad and it can't adapt. Expect to refine.
Patch requirements are strict. If you're on an older instance, the upgrade path to get AI Agents working is non-trivial. Plan for it.

The Bottom Line

ServiceNow AI Agents represent a genuine shift from scripted automation to dynamic problem-solving. After seeing my Software Access Assistant handle real requests—checking context, reasoning through approval rules, and taking appropriate action—I'm convinced this is the direction service management has to go. The days of building rigid decision trees for every scenario are numbered.

The key is starting small. Pick one painful, repetitive workflow. Build an agent for it. Give it a handful of well-named skills. Set guardrails that keep you comfortable. Then watch it work, learn from the analytics, and expand from there.

My backlog is down to 40 incidents. My team is finally working on problems that actually require human judgment. And the agents? They're handling the rest, 24/7, without complaint.