From Development to Production: Ask ServiceNow's AI Experts Anything

[ryanyong]

Hey everyone, I'm Ryan, an engineer on ServiceNow's Forward Deployed Engineering team. We're a team of experts and builders who've gained deep AI deployment expertise at tier 1 technology companies and research labs. We're certified experts on the ServiceNow AI Platform, with direct access to Product and the ServiceNow C-suite.

What we do
We partner directly with customers facing the most complex challenges to co-innovate agentic AI use cases and ship them to production on the ServiceNow AI Platform.

Why this thread
If you're a ServiceNow customer sitting on AI entitlement you haven't fully put to work, this is your chance to ask the Forward Deployed Engineers building these solutions in the field. Bring your hardest questions about moving from unused credits to AI running in production, and getting the most value out of your existing ServiceNow investment.

Ask us anything. We are the subject matter experts, and we will get you an answer.

We'll be answering live from August 3rd to 7th. Post your question anytime between now and 8/7, upvote the ones you want answered first.

[GreysonS]

The premise is right, and the useful part is that provider selection is already a configuration point rather than something baked into your build. You can set it at the instance, skill group, or skill level in the Now Assist Admin Console, with AI Control Tower governing what's allowed, and Now LLM Service stays available if you'd rather remain there. So portability is less an architecture problem than a question of whether you're in a position to change providers with any confidence.

That reframes where to spend effort:

Build an eval set now, before you need one. This is the thing that makes model portability real. A configuration knob doesn't help if you can't tell whether flipping it improved or degraded anything. Take fifty or a hundred representative HR scenarios, define what a good answer looks like for each, and run them whenever something changes. Most teams that feel locked into a model are really just locked out of measuring.

Put your effort into knowledge, not prompts. HRSD answers come from retrieval over your knowledge base. A stronger model reading stale, badly structured, or wrongly permissioned HR content still gives bad answers. Article ownership, retirement discipline, and user criteria segmentation pay off under every model and are the least likely work to be obsoleted. Prompt text tuned to one model's quirks is the most likely.

Keep deterministic work deterministic. Push business logic into tools, flows, and script includes, and let the model orchestrate. The narrower the model's job, the less its identity matters and the easier the whole thing is to test.

Use the governance layer instead of building one. AI Control Tower now governs customer-configured providers alongside ServiceNow's own from a single page, with data routing controls and approval enforcement at the Skill Kit level so builders can only reach approved providers and models. For HR specifically, the output screening for PII leakage deserves more attention than it usually gets, given what sits in HR cases.

[LyndseyD]

Hi,

We recently installed CSM Pro so now have all the AI features, but we're unsure where to start. We are waiting on information back from our account representative but this is taking some time. Which plugins would you recommend we start with to start seeing some immediate benefits, such as reduced resolution times for the Service Desk?

Thanks

[GreysonS]

Dustin's answer above on where to start with CSM and ITSM applies almost directly here, so I'd begin in the same place.

AI Search first. It's the foundational RAG layer for the platform and most of the Now Assist skills lean on it, so getting your search profiles right pays off across everything you turn on afterward. If they're wrong, the skills sitting on top will underperform and you'll spend weeks blaming the wrong component.

Then install at the product level. Request Now Assist for Customer Service Management from the Store rather than picking individual child plugins off the list. That brings the dependencies and entitlements down with it and saves you a whole category of version mismatch problems later.

On your specific goal, the fastest path to lower handle time is the single turn summarization skills: case summarization and resolution notes generation. Minimal configuration, and the benefit shows up immediately in agent time. Dustin's insurance example is the right mental model. Five to seven minutes of reading case history before a call turns into seconds, and at any real volume that adds up quickly.

I'd hold off on Suggested Steps for now, even though it sounds like the obvious resolution time play. It runs on GAF clustering, which needs AI Search active and a decent volume of resolved cases before it produces anything useful. Bank the summarization wins first.

One thing you can do without waiting on your rep: Now Assist Admin has a plugin page under Settings showing what's installed and what you're entitled to. Worth a look before you plan, since the CSM SKU name and the GenAI entitlement aren't always the same thing and it's better to find that out now than three weeks in.

Last thing. Capture your current average handle time and resolution time before you switch anything on. You'll want that baseline when someone asks what this bought you.

[JannikW]

How does ServiceNow support customers in complying with the EU AI Act? Are there official guidelines, best practices, or recommendations covering:

• AI use case classification and risk assessment
• Transparency, traceability, and documentation
• Human oversight, monitoring, and auditability
• Data quality, privacy, and governance
• Responsibilities of ServiceNow and its customers
• Available templates, technical features, or compliance tools
• Planned product updates related to the EU AI Act

[AlpDoganA]

We already have Now Assist licensed within IT Operations Management Advanced. We mainly work with Incident and Task management, backed by our CMDB. What would a realistic first deployment look like here? What are the common pitfalls to watch out for?

Thanks