5 ServiceNow CSM workflows AI can take off your team’s plate

Customer Service Management is where most of the AI demos in ServiceNow land first, and for a good reason: a case is a tidy package of text, history and outcome. That makes it easy to summarize, classify and draft from. It also makes it easy to automate the wrong thing. These are the five workflows I would look at first, in roughly the order I would try them, with what each one needs inside the instance and where a person should stay in charge.

A quick note on where this comes from. I hold the ServiceNow CSA, CAD and CIS-CSM certifications, and I built SlickNow.Learning, a mock-exam app with 180 CIS-CSM practice questions, so the CSM data model is something I have taken apart line by line. What follows is a practitioner’s reading of ServiceNow’s own documentation plus how I would scope each piece. It is not a list of things that “transform” anything. Every product detail below was checked against the ServiceNow docs in October 2026, and names change between releases, so check yours.

Before you start: three things to check

Every workflow below depends on the same three things, so check them once before you look at any of them.

• What your contract includes. Now Assist for CSM, Predictive Intelligence and Virtual Agent are separate capabilities, and which of them you can switch on depends on the packages you have bought. Ask your ServiceNow account team before you plan around one. I do not quote licence prices because they are negotiated.
• Which release you are on. The Now Assist feature list grows every release, and the naming moves too: in the newest documentation some features appear under the ServiceNow Otto name and in the CRM Workspace. Even small details shift. The minimum case length for generating resolution notes is not the same in every release’s docs I read. Read the docs for your release, not a blog post (including this one).
• Whether your case data is any good. AI on top of messy cases gives you confident, messy output. If short descriptions are “help pls” and categories are left on the default, start there.

That third point is the one people skip. A two-minute background script tells you a lot:

// How many active cases have no assignment group after four hours?
var gr = new GlideRecord('sn_customerservice_case');
gr.addActiveQuery();
gr.addNullQuery('assignment_group');
gr.addQuery('opened_at', '<', gs.hoursAgoStart(4));
gr.query();
gs.info('Unassigned after 4h: ' + gr.getRowCount());

Run something like that in a sub-production instance first, and you have a baseline before you change anything. Without a baseline you cannot tell whether the AI helped.

1. Case summaries and handoffs

The manual workflow. A case bounces between tiers, or a customer calls back, and the agent who picks it up scrolls through work notes, emails and comments to work out what has already happened. On a long case that is several minutes of reading before any work starts.

What AI does. Now Assist for CSM can summarize a case: the issue and the actions taken so far, drawn from fields like the short description, description, work notes, comments and email. The agent can refresh the summary and post it to the work notes. It can also summarize chats at the point a Virtual Agent conversation is handed to a live agent, and at wrap-up.

What it needs. The Now Assist for CSM plugin and its skills switched on, agents working in the workspace, and cases with enough written in them to summarize. A case with one line of description gets a one-line summary.

Where the human stays. The summary is a reading aid. The agent still decides what to do, and anything posted to work notes should be checked, because it becomes part of the record the next person relies on.

Why it goes first. Nothing reaches the customer, the risk is low, and the time saved is easy to see. If you only try one thing, try this.

2. Classification and routing

The manual workflow. A case arrives by email or portal, and someone triages it: picks a category, picks an assignment group, maybe sets a priority. When that someone is busy, cases wait in a queue that nobody owns.

What AI does. This one is older than generative AI. Predictive Intelligence, ServiceNow’s machine-learning framework, ships with classification solution definitions for CSM: one predicts the assignment group from the case’s short description, another predicts the category. They are templates. You train them on your own instance’s history and can create your own definitions for other fields.

What it needs. Past cases with the target field filled in correctly. This is the whole game. The model learns from what your team did, so if your history is full of cases that were reassigned three times, it learns the first, wrong group. Look at your reassignment counts before training.

Where the human stays. Start with predictions shown as suggestions, or applied only above a confidence threshold, with the agent able to change them. Then watch how often a predicted group reassigns the case. That number is your quality measure.

3. Drafted replies and resolution notes

The manual workflow. An agent writes the same kind of email ten times a day, then writes resolution notes at the end of each case, often in a hurry, which is why so many resolution notes say “resolved”.

What AI does. Now Assist for CSM can generate suggested email replies and chat replies, and draft resolution notes from the case record. In the resolution flow the agent picks a resolution code and reviews the generated text before proposing the fix to the customer. Newer releases add suggested work notes and comments as well.

What it needs. Cases with enough detail to draft from (the resolution notes skill has a minimum amount of case text in the docs, and the figure differs by release), and a tone your team agrees on. If your best agents write very differently from each other, settle that first.

Where the human stays. Always on send. A drafted email is a draft. This is the first workflow on the list where AI output reaches a customer, so the agent reads every one before it goes. I would also keep a short list of topics where drafts are switched off: refunds, legal complaints, anything regulated.

The quiet win. Better resolution notes are worth more than they look. They are the raw material for workflow 5, and they make the next case summary better too.

4. Deflection before a case exists

The manual workflow. A customer has a question the knowledge base already answers, cannot find the article, and opens a case. An agent then pastes the article link.

What AI does. Virtual Agent on the portal, backed by AI Search over your knowledge base, answers the question in the conversation. Now Assist in Virtual Agent adds generative answers drawn from knowledge articles, and can use a language model to work out which topic the customer wants instead of matching keywords. Virtual Agent has a deflection topic block that records whether a conversation was a confirmed deflection, a potential one (results were shown, but nobody confirmed they helped) or none, so you can measure it honestly.

What it needs. A knowledge base that is current and written for customers, not for agents. This is the workflow most likely to fail for a non-AI reason: the AI answers confidently from a 2021 article. If you want to understand how answering from your own documents works and where it goes wrong, read RAG explained for business owners.

Where the human stays. A clear, always-visible route to a person. Customers forgive a bot that hands over quickly. They do not forgive one that loops. For sensitive subjects, Now Assist Guardian has sensitive topic filters for Virtual Agent (with CSM or HR Service Delivery installed) that route matching conversations to a fallback topic instead of an AI answer.

How to measure it. Count confirmed deflections, not conversations. A thousand conversations where nobody confirmed anything is not a thousand saved cases.

5. Knowledge articles from resolved cases

The manual workflow. The fix for a new problem lives in one agent’s resolution notes and in their head. Writing a knowledge article is the step everyone agrees matters and nobody has time for.

What AI does. Now Assist for CSM can draft a knowledge article from the case data once a resolution has been proposed or the case is closed.

What it needs. Good resolution notes (see workflow 3) and a knowledge workflow with a reviewer. Knowledge Management already has review and publishing states. Use them.

Where the human stays. The draft goes through your normal knowledge review before it is published. An article with a wrong step in it does more damage than no article, because workflow 4 will now serve that wrong step to customers at scale.

These two workflows feed each other. Better articles mean more deflection, and more deflection means fewer cases. That loop is the real prize in CSM, and it only works if a person checks each article on the way in.

What I would do first

If you asked me where to start on a CSM instance I had not seen, I would do this, in order:

1. Run a handful of data checks like the script above: unassigned time, reassignment counts, how many cases have real descriptions and real resolution notes.
2. Switch on case summarization for one team in sub-production, then production, and ask the team after two weeks whether they still read the full history.
3. Train the assignment classification on your history and run it as a suggestion only. Track reassignment rate.
4. Only then look at drafted replies, and only with a person on every send.
5. Leave deflection until the knowledge base has been cleaned, because it amplifies whatever is in there.

That order is deliberate. It goes from “AI helps an agent read” to “AI helps an agent write” to “AI talks to a customer”, raising the risk only as you gain evidence. I wrote the general version of this method in How to scope an AI automation project without wasting money.

When the built-in skills are not enough

Sometimes Now Assist is not in the contract, or the workflow crosses into systems ServiceNow cannot see, or you want a model you choose. Then the answer is a custom integration: a scoped integration user with only the roles it needs, ServiceNow’s REST APIs (or a scripted REST API you write), and an outside AI that reads cases and writes back suggestions as work notes. An MCP server is a clean way to give an assistant like Claude that typed, limited access. I compare the native and custom routes, and the guardrails each needs, in ServiceNow AI integration: what actually works.

If you would like a second pair of eyes on your instance before you buy or build anything, that is what the AI Readiness Audit is for. For the build itself, see ServiceNow AI integration, or tell me what your team is stuck on. If you just want to know whether your workflows are ready, the free 2-minute audit is a decent first filter.

Questions people ask