Now Assist Skills vs AI Agents: When to Use Which, and What It Costs in Assists (as of Sept 2026)

sachinbhasin11 · an hour ago

Now Assist Skills vs AI Agents on ServiceNow: Where to Use Each, and What It Really Costs in Assists

Updated September 2026. Reflects the Australia release, the April 2026 Foundation / Advanced / Prime packaging, and the Now Assist rate card effective 24 August 2026.

Skills are cheap. Agents cost more per run but own a whole outcome. On the official Now Assist rate card, most summarisation and resolution-note skills cost 1 assist. An agentic workflow costs 25, 50 or 150 assists depending on how many tools it calls.
Agents dominate consumption. In ServiceNow's own forecasting example, AI agents handled 20% of incident volume but drove 82% of assist consumption.
Hybrid designs win. Use flows and Predictive Intelligence for the predictable parts, skills for human-reviewed drafting, and agents only for ambiguous, multi-step work. Skills called directly as agent tools are not billed separately.
Rates are published; pool sizes are not. Negotiate pool size, overage rate and alerting before you sign, and switch on the runaway-consumption controls from day one.

Why this question matters now

Since ServiceNow launched AI Agents, architects keep asking the same thing: if an agent can summarise and update a record, why keep skills? Some even ask on the Community whether the framework will slowly move from skills to agents.

The answer is no. Skills and agents solve different problems and are metered very differently. Choosing the wrong one is one of the fastest ways to use up an assist entitlement.

A note on naming: at Knowledge 2026 ServiceNow brought Now Assist, Moveworks and AI Experience together as ServiceNow Otto. Otto is the experience layer, so the Now Assist Panel is now the Otto panel.

Consumption still follows the same rate card, and the skill-versus-agent distinction below applies either way.

The building blocks

Construct	What it is	Determinism	Assist profile
Flow / subflow / script	Rule-based automation	Fully deterministic	0 (unless it calls GenAI)
Predictive Intelligence	Classic ML for classification, similarity and routing	Deterministic once trained	0 (not on the rate card)
Now Assist skill (OOTB)	One bounded GenAI task: summarise, draft, extract or classify	Predictable, with a human reviewing	Typically 1–10
Custom skill (Now Assist Skill Kit)	Your own prompt and tools	Predictable	1 per LLM call
Virtual Agent topic / conversational catalog item	Guided or LLM-assisted conversation	Scripted to semi-scripted	10 per end-to-end topic
Otto (Now Assist) Panel conversation	Conversational assistant in the workspace	Semi-open	5 per conversation
AI agent	A reasoning worker that plans its steps and calls tools	Probabilistic	25 / 50 / 150 per run, by tool count
Agentic workflow	An orchestrator coordinating a team of agents	Probabilistic	Same tiers, counted across the run
AI specialist (Autonomous Workforce)	A packaged, role-scoped team of agents	Probabilistic, with guardrails	e.g. L1 IT Service Desk AI Specialist: 15 per assigned incident

ServiceNow's AI Centre of Excellence describes the choice as a spectrum from deterministic to probabilistic, with cost generally rising as work becomes more probabilistic.

ServiceNow's FSO guidance separates the two agent constructs simply: an orchestrator directing several agents is an agentic workflow, and a single agent acting alone is an AI agent.

Decision framework !!

Rule of thumb: if you can write the steps down in advance, use a flow or a skill. If the AI has to decide the steps, use an agent.

Ask these questions in order and stop at the first “yes”:

Can it be scripted? Use a flow, or Predictive Intelligence for classification. Cost: 0 assists.
Is it one generative step that a human reviews? Use an OOTB or Skill Kit skill. Cost: 1–10 assists.
Is it a known conversation with a known outcome? Use a VA topic or conversational catalog item (10 assists). For open questions, use AI Search Q&A Genius (0) or the Otto panel (5).
Are there several generative steps in a fixed order? Chain the skills in a flow. Cost: the sum of the skills.
Does each step depend on what the previous one found, across several tools? Use an AI agent, designed to stay in the Small tier.
Does the work span teams or modules? Use an agentic workflow, budgeting 50–150 assists per run.
Before building anything custom, check the AI Gallery for an OOTB agent or AI specialist. This is the first rule in ServiceNow's Australia build and governance guide.

The hybrid pattern is usually the best answer. A deterministic flow decides what enters the agent, the agent reasons only over the ambiguous cases, and skills and topics are called directly as tools inside it.

Practical examples

Volumes are illustrative. Rates come from the August 2026 rate card.

Note: Key assist table where you can monitor your assist usage and can also generate interactive analysis report is "sys_gen_ai_usage_log"

#	Use case	Recommended approach	Why	Assists per unit	Illustrative monthly
1	ITSM: engineer reads incident history and writes closure notes	Skill (summarisation + resolution notes)	Bounded and human-reviewed	2	10,000 × 70% × 2 = 14,000
2	ITSM: categorise, set service/CI, link to major incident or problem	Hybrid: Predictive Intelligence first, OOTB triage agents only for low-confidence records	Most records are predictable	0 (PI) or 25–50	15% residual × 25 = 37,500 (vs 250,000 if every incident goes to agents)
3	ITSM: L1 password, VPN, access and software tickets resolved end to end	AI specialist (L1 IT Service Desk) where licensed; otherwise a lean custom agent	Packaged outcome, metered below a small agent run	15 (vs 25–50 custom)	3,000 × 15 = 45,000
4	Knowledge: turn resolved incidents into articles	Skill (KB generation); knowledge gaps detection quarterly	Drafts for a human editor	10 per article; 750 per gap scan	400 × 10 + 750 = 4,750
5	CSM: case intake triage and first response	Hybrid: triage agentic workflow, email-response skill as a tool, case summary for humans	Ambiguous intent and multiple lookups	25–50 + 1	2,400 × 25 + 8,000 × 1 = 68,000
6	Sentiment monitoring	Skill on a schedule, with a tuned filter	A classic source of runaway consumption	1	Filter to “updated in the last 24 hours, once daily”, not all active records every 15 minutes (source)
7	HR: policy questions in Employee Center or Teams	Search (Q&A Genius) + VA topics; an agent only for resolving HR cases	Answers cost 0; transactions are scripted	0 / 10 / 25+	3,000 topics × 10 = 30,000
8	ITOM / AIOps: alert triage	Skill (alert investigation) for humans; AI specialist for autonomous remediation	Investigation is cheap; remediation needs governance	1; agentic tiers	5,000 × 1 = 5,000
9	SecOps: closing security incidents	Agent, designed to make 4 or fewer tool calls	Repetitive, multi-step closure	25 (50 if it slips to medium)	600 × 25 = 15,000
10	SPM: status reports and portfolio insights	Skills	Draft-and-review content	10–20	200 projects × 10 × 4 = 8,000
11	Creator / App Engine	Skills for code (5); govern the generative builders tightly	Few actions, but very expensive ones	Flow generation 500; app generation 2,500	10 app generations = 25,000
12	FSO: ACH dispute resolution	Agentic workflow + dispute summary skill	Regulated, multi-agent evidence gathering	Likely 50–150	1,000 × 50 = 50,000
13	External agent (e.g. Claude) summarising incidents via Action Fabric	MCP Server tools	Headless and governed by AI Control Tower	Skill rate + 1 (FAQ)	20,000 × 2 = 40,000

How assists are consumed

An assist is ServiceNow's unit of AI consumption. One is used each time a Now Assist skill or agentic workflow runs. ServiceNow meters these actions, not tokens.

Key rates (rate card effective 24 August 2026)

Action	Assists	Notes
Summarisation (incident, case, change, etc.); resolution notes; sentiment; alert investigation	1	Per request
Skill Kit / custom call; MCP server call; AI gateway call	1	Per LLM call or per call; evaluations scale with the number of records
Email / activity response; incident assist; code generation; Otto panel conversation	5	A panel conversation ends after 1 hour idle or a change of intent
KB article generation; VA topic; multi-turn catalog order; Moveworks simple automation	10	Per article or per end-to-end execution
L1 IT Service Desk AI Specialist	15	Per assigned incident
Agentic workflow: Small	25	Rate card: fewer than 4 actions. Community guidance: 0–4 tools.
Agentic workflow: Medium	50	5–8 actions
Agentic workflow: Large	150	9–20 actions; beyond 20 starts a new workflow
Voice agents	S / M / L tier	Calls under 30 seconds are free (no longer a flat 50)
LEAP action reducer	125	Per group; was 25 in the September 2025 rate card
Flow generation/widget generation	500	Per generation
Experience (UI) generation	1,000	Per run
App/playbook generation	2,500	Per generation
Q&A Genius (AI Search); lookups	0	Was 1 in 2025
AI Control Tower managed AI asset	200	Per asset per month

Source: Now Assist Overview (legal). This document changes often, so always check the current version.

Seven consumption rules every architect should know

Agents are billed by tool calls, in tiers. One action is one tool invocation. Orchestrator and communicator actions are excluded from the count.
Skills called directly as agent tools are free. Skills, VA topics and conversational catalog items used as tools inside an agent are not billed separately. However, a skill wrapped inside a script or flow tool is billed.
Tier boundaries are cliffs. Cutting an agent from 9 tool calls to 8 drops each run from 150 to 50 assists, a 67% saving.
Long runs restart the meter. A run of 26 tool calls is billed as one Large run (150) plus a new Medium run (50).
Sub-production is not free. AI Agent Studio testing and agentic evaluations draw on the same account-level pool. ServiceNow's governance guide estimates evaluations at roughly 1 assist per record.
The pool is annual and use-it-or-lose-it. Under the legal terms, entitlements run for 12 months and reset on the purchase anniversary, with no rollover. Some third-party sources claim monthly resets, which contradicts the contract terms.
External agents. Skills exposed through the MCP Server cost their usual rate plus 1. An external agent calling a ServiceNow agent is billed on that agent's tool count. A ServiceNow agent calling an external agent uses no assists.

Worked cost comparisons

A. 10,000 incidents a month

Pattern	Calculation	Assists / month
Skills only: summary (70%) + resolution notes (60%) + 300 KB articles	7,000 + 6,000 + 3,000	16,000
Add a Small agent for triage on 30% of incidents	3,000 × 25	+75,000
Same agent, drifting into Medium	3,000 × 50	+150,000
Same agent, over-tooled into Large	3,000 × 150	+450,000

One Small agent run costs as much as 25 summaries. That is not an argument against agents. It is an argument for being deliberate about which work they take on.

B. Cost per resolved outcome, not per run

Failed or escalated agent runs still consume assists. A custom L1 agent that averages 25 assists per run and resolves 60% of tickets costs about 42 assists per resolved ticket. At 50 assists per run, it costs about 83. The L1 AI Specialist, at 15 per assigned incident with the same resolution rate, costs about 25. Filter what enters an agent with deterministic eligibility rules.

C. The skill-vs-agent design trade-off

ServiceNow's CoE gives an example: a Virtual Agent user asks about a policy and wants it emailed. Built as an agent, it costs more than twice as much as Q&A Genius plus a Skill Kit email skill plus an LLM topic (source). On the current rate card the comparison is roughly 25 vs 11 assists.

D. Sizing the pool

ServiceNow PM Neha Agrawal's month-one ITSM baseline is 5,500 assists for skills plus 25,000 for agents, or 30,500 a month. ServiceNow does not publish pool sizes per tier. Advisory estimates put an ITSM fulfiller at about 1,500 / 3,000 / 6,000 assists a year on Foundation / Advanced / Prime (Crossfuze, Redress Compliance). Treat these as estimates, not contractual figures.

What causes runaway consumption

Trigger loops. Combining “On Record Created” and “On Record Updated” triggers without condition filters is the most common source of recursive loops.
Cloned data plus scheduled jobs. A clone can expose an entire incident backlog to a scheduled GenAI job.
Skills switched on by default. Since December 2025, selected skills activate automatically when plugins are installed or upgraded. Review the Admin Console after every upgrade (source).
Retry loops. Analysts have warned about agents that keep consuming assists each time they retry a failed task (CIO.com).
Over-tooling. Keep to 10 or fewer agents per use case and keep each agent's toolset lean.

Controls available

Control	Where / property	Notes
Trigger kill switch	`sn_aia_property`: `kill_switch.mode` (off / warn_only / enforce)	Enforce mode deactivates a runaway trigger on day 3. Zurich Patch 10 / Australia Patch 3+.
Recursive checks	`recursive_check.create_max_executions` (50 / 15 min), `update_max_executions` (5 / 15 min)	Caps loops (guide)
Assist spike alerts	`alert.assist_spike_*` properties	Keep these on in sub-production too
Rate limiting	`sys_one_extend_rate_limit_rules`, scoped to the “AIA ReAct Engine” capability	Caps agentic runs without affecting skills
Tool-loop cap	`sn_aia.continuous_tool_execution_limit`	Stops one tool being called repeatedly
Telemetry	`sys_gen_ai_usage_log`, `sys_generative_ai_log`, `sn_aia_tool_execution_tier`, `sn_aia_execution_plan`	Join these to trace each assist to a record
Dashboards	AI Agent Studio > Analytics; Now Assist Admin > Analytics; Subscription Management	Account-level view by instance and skill
AI Control Tower	Inventory, per-team quotas, AI Gateway pause / kill switch	What's new, Aug–Sep 2026

Recommendations

Write down the design order: flow or Predictive Intelligence first, then skill, then topic, then agent. Require a short justification for every agent, including its expected tier and cost per outcome.
Design agents for the Small tier. Aim for 3 or fewer tool calls, to stay clear of the ambiguity at exactly 4. Move deterministic steps into flows, and call skills directly as Skill tools.
Filter what enters an agent using conditions or Predictive Intelligence confidence, because failed runs still bill.
Turn on the guardrails from day one. Run the kill switch in warn_only mode for 2–4 weeks, then switch to enforce. Add rate limits and spike alerts, and narrow scheduled-job filters.
Budget sub-production explicitly, covering evaluations, testing and prompt iteration.
Negotiate the meter, not just the tier. Get the pool size, overage or top-up rate, alert thresholds and a first-year ramp written into the order form (Redress Compliance).
Report cost per outcome, such as assists per resolved ticket, next to MTTR and deflection rate.

Caveats

The rate card is a living legal document that changed several times between 2025 and 2026. Check the current version before you rely on any figure.
Pool sizes per tier, overage pricing and the dollar value of an assist are not published. Any such figures are third-party estimates.
Two points are ambiguous across ServiceNow's own sources: how a run of exactly 4 tool calls is billed, and how MCP Client tools are counted. Confirm both with your account team.

Conclusion

Skills and agents are layers, not competing generations of the same technology. Skills are high-volume, low-cost and human-reviewed: they make people faster. Agents run less often but deliver more per run: they own an outcome. The best designs combine modular skills with lean agents, behind deterministic gates, and measure consumption from day one.