As of October 2026. Now Assist is being renamed ServiceNow Otto; I use "Otto (Now Assist)" below. "Flow" here means Flow Designer flows.
With Otto (Now Assist) skills, AI Agent Studio and Flow Designer all on the platform, almost every new use case raises the same question: should we build an AI agent for this?
In my experience, that's usually the wrong first question. A better one is older than AI:
Who decides what happens next: the developer, or the LLM?
Here's the framework I use to answer it.
The ServiceNow AI ladder
Five rungs, one rule: pick the lowest rung that works.
Rung Use it when… Built with Who's in control
| 1. Rule or flow | The logic is a clear if-then on structured fields (priority, category, assignment group) | Flow Designer, business rules | Developer |
| 2. Decision table | Many conditions, and the business wants to own the logic | Decision tables | Developer |
| 3. Classic ML | You need to classify or route from history, at volume | Predictive Intelligence | Developer |
| 4. GenAI skill in a flow (Pattern A) | One step must read or write language: summarise, extract, draft | Now Assist Skill Kit, called from a flow | Developer |
| 5. AI agent (Pattern B) | The path itself varies and needs judgement across several tools | AI Agent Studio | LLM |
Notice that the first three rungs don't use generative AI at all. And on rungs 1–4, the developer still owns the path. Only on rung 5 does the LLM decide the next step.
Every time you climb, ask: "Why not the rung below?" If there's no good answer, you've climbed too far.
Same idea, wider-AI vocabulary
If you follow AI outside ServiceNow, you'll have heard "workflows vs agents". It maps cleanly:
- Single LLM call (classify, summarise, extract) → a Skill Kit skill called as an action in a flow
- Chain (several LLM calls in a fixed order) → a skill with ordered steps, or a flow calling several skills
- Router (the LLM picks one of a few fixed paths) → before building one, check whether Predictive Intelligence or a decision table already does the job
- Agent (the LLM picks a tool, reads the result, decides what's next) → an AI Agent Studio agent whose tools are your flows, scripts and skills
Skills and agents aren't competitors
Two lines worth remembering:
- A skill is a workflow with an LLM inside it.
- An agent is an LLM with workflows inside it.
An agent calls your skills, and your flows can call the very same skills. Build once, use in both. So building good skills now is never wasted work: if you need an agent later, your skills become its tools.
The one-line decision test
Can you draw the flowchart before you start? → Skill inside a flow. Can you only state the goal ("resolve this", "investigate that")? → Agent.
Choose a skill in a flow when… Choose an agent when…
| The steps are known up front | The next step depends on what the last step found |
| You need auditable behaviour | The work is investigative, not procedural |
| Cost must be predictable | You can measure quality with evaluations |
Three things to plan for before choosing an agent:
- Cost: agent runs consume assists, and agents can loop through several steps.
- Observability: a flow is easy to trace; an agent needs step-by-step traces and evaluations.
- Licensing: check which Otto (Now Assist) tier you have before promising an agent. Creating or significantly modifying agents depends on your tier.
My verdict
- Flow here: structured data, known steps, clear rules (rungs 1–3).
- AI here: when one step reads or writes language, add one skill to the flow you already have. For most teams, that's the quickest way to real value.
- Agent: when the path truly varies and you can measure it with evaluations.
Start every use case on the lowest rung, and earn your way up.
Over to you
What use case are you weighing up right now: flow, skill or agent? Share it in the comments and I'll gladly think it through with you.
I've also walked through this ladder visually in a short video, if that format helps: https://youtu.be/HWl39KR_qaA?si=FKnNgHABjMnwsdEH
If this was useful, a 👍 helps others find it.
