Best Practices in managing AI agents, skills, and assists in ServiceNow

hussein_alani · 3 weeks ago

This guide provides a technical-operational framework for platform administrators, AI architects, and delivery managers who are responsible for running AI agents, AI skills, and agentic workflows on ServiceNow. It covers five areas:

how assists are consumed and billed,
how to track and monitor that consumption,
how to configure rate limits and safeguards,
how to measure per-agent accuracy,
and how to run automated evaluations before and after production deployment.

Table of Contents

Purpose and scope. 4

Understanding assists consumption. 5

Assists vs. actions. 5

Agentic workflow tier sizing. 5

Key tables for consumption data. 5

Tracking and monitoring assists. 6

Now Assist Analytics. 6

Subscription Management: the cross-instance view.. 8

Assist spike alerting. 8

Building a custom forecasting dashboard. 8

Operational pro tips. 9

Logging and auditability. 10

What the platform logs automatically. 10

Tracing an assist charge to a business record. 10

Logging best practices. 10

Aggregation and reporting. 11

Rate limiting AI 12

One Extend rate limit rules. 12

Trigger throttling (Australia Patch 3+) 12

Recursive check properties. 13

Agent design patterns that reduce rate limit pressure. 14

Measuring accuracy per agent 15

Core evaluation metrics. 15

The Agent Productivity Score. 15

The Workflow Automation Score. 16

Accuracy tracking cadence. 16

Per-agent dashboard design. 16

Automated evaluation practices. 17

The evaluation development cycle. 17

Building evaluation datasets. 17

AI Agent Advisor: opportunity discovery and pre-built datasets. 17

Connecting evaluations to analytics. 18

A/B testing for model and prompt changes. 18

Governance with AI Control Tower. 18

Implementation checklist 18

Week 1–2: Foundation. 18

Week 3–4: Monitoring. 19

Week 5–6: Accuracy and evaluation. 19

Week 7–8: Governance. 19

Ongoing. 19

Appendix: quick reference tables. 20

Key system properties (sn_aia_property) 20

Key platform tables. 21

Purpose and scope

This guide provides a technical-operational framework for platform administrators, AI architects, and delivery managers who are responsible for running AI agents, AI skills, and agentic workflows on ServiceNow. It covers five areas: how assists are consumed and billed, how to track and monitor that consumption, how to configure rate limits and safeguards, how to measure per-agent accuracy, and how to run automated evaluations before and after production deployment.

Content in this guide reflects the Australia release (including Australia Patch 3, June 2026) and references system properties, tables, and dashboards available in Now Assist AI Agents plugin version 6.0 and later. Where a feature requires a specific patch level, that requirement is noted inline.

This document is intended to be used alongside the official ServiceNow product documentation and the AI Center of Excellence community articles. It consolidates scattered guidance into a single operational reference.

Understanding assists consumption

Now Assist does not charge for tokens. It charges for value in the form of assists. An assist is ServiceNow’s unit of AI consumption, triggered each time a Now Assist skill or agentic workflow executes. Different capabilities consume different amounts based on complexity.

Assists vs. actions

Each Now Assist skill is composed of one or more actions. A straightforward skill such as case summarization involves few actions and consumes one assist. A complex agentic workflow involves multiple actions and can consume 25 assists or more. Each customer has a contracted pool of available assists, and consumption is tracked against that pool on a 365-day burn-down model that resets annually on the contract anniversary date.

Agentic workflow tier sizing

For AI agents specifically, the number of tool actions executed during a single agentic workflow determines the tier and therefore the assist cost. Orchestrator and communicator agent actions are excluded from the count.

Tools executed per run	Tier	Assists consumed
0–4 tools	Small	25 assists
5–8 tools	Medium	50 assists
9–20 tools	Large	150 assists

Skills, virtual agent topics, and conversational catalog items that run inside an agentic workflow as tools do not consume additional assists on their own—only the agentic workflow run itself is billed. However, sub-production environments and AI Agent Studio testing also consume assists from your contracted pool.

Pro tip: Review your agents for over-tooling. Can you combine or reduce the number of tools in use? Could a virtual agent topic or conversational catalog item handle the use case instead of a full agentic workflow? Dropping from 9 tools to 8 cuts your per-execution cost from 150 assists to 50—a 67% reduction.

Key tables for consumption data

The following platform tables store the raw consumption and configuration records. Administrators should be familiar with all of them for troubleshooting and custom reporting.

Table	Purpose	Notes
sys_gen_ai_usage_log	Assists count per invocation	Populated automatically on every generative AI feature invocation. If assist calculation fails, a scheduled job re-calculates hourly. Accessible to admins.
sys_generative_ai_log	Case/task reference per skill execution	Links each skill execution to the originating case or task record. Primary table for tracing a specific assist charge back to business context.
sn_aia_tool_execution_tier	AI agent execution plan tier consumption	Shows the tier (Small/Medium/Large) assigned to each execution plan. Essential for understanding why a particular run consumed 25, 50, or 150 assists.
now_assists_skills	Aggregated skill usage	Aggregates usage counts within the applicable billing window. Use for trending and forecasting.
sn_nowassist_skill_config	Skill configuration records	Lists all configured Now Assist skills, their settings, and status.
sn_aia_agent	AI agent definitions	Stores agent configurations including role, instructions, tools, and activation status.
sn_aia_execution_plan	Execution plans for AI agents	Records the full execution plan for each agentic workflow run—every step the agent considered and took.
sn_aia_usecase	Use case definitions	Defines agentic use cases (the business process an agentic workflow automates).

Tracking and monitoring assists

Knowing where to look is half the challenge. ServiceNow provides two complementary layers for monitoring AI consumption: instance-level Now Assist Analytics (for real-time, granular data within a single instance) and cross-instance Subscription Management (for enterprise-wide pool tracking).

Now Assist Analytics

Navigate to All > Now Assist Admin > Analytics to access the centralized dashboard collection. This is a store application that requires no additional licensing beyond your Now Assist subscription. Ensure you are running the latest version from the ServiceNow Store (the dashboards have been updated frequently since initial release).

The dashboards most relevant to consumption and operational tracking are:

Dashboard	What it answers	Key metrics
Usage and adoption	Who is using Now Assist, how much, and where?	Total actions, daily actions, average daily unique users, skill group distribution, adoption by department, feedback and error details
Self-service performance	Is Now Assist actually reducing service desk workload?	Total queries, query resolution rate, resolution status, successful conversations by capability
Skill performance	Which skills are gaining traction and which need attention?	Skill engagement trend, total actions, daily active users, accepted actions, acceptance rate
Custom skill details	How are organization-built skills performing?	Engagement trend by workflow/product/feature, daily unique users, execution success rate, user feedback
Now Assist Guardian	How often are content filters and prompt injection defenses triggered?	Guardrail-added latency, percentage flagged as offensive, prompt injection occurrences by skill
User search analyzer	What are users searching for and where are the content gaps?	Queries with/without genius results, top 5 queries without results, search queries per channel
Value insights	How much time is Now Assist saving? (Moved to AI Control Tower in Zurich.)	Total writing/reading time saved, average time saved per record and per user

Pro tip: The user search analyzer is one of the most actionable dashboards for continuous improvement. The “top 5 queries without results” metric directly tells you where to create or improve knowledge articles to increase deflection.

For a detailed overview of the dashboards, check this article.

Subscription Management: the cross-instance view

Navigate to Admin > Subscription Management > Subscriptions > Now Assist usage for the enterprise-wide perspective. This view shows three critical numbers at the top: total purchased assists, total used assists, and available (remaining) assists across all instances.

Below the summary, a breakdown by instance shows which environments consume the most. Drill into the Account level entitlements tab to see consumption by skill/agent name, action count, assist ratio, and total assists. This breakdown reveals which skills consume the most entitlements—you may discover that a heavily-used lightweight skill consumes far fewer assists than a rarely-used complex agent.

Assist spike alerting

ServiceNow provides system properties (available in the sn_aia_property table as of Australia Patch 3) that automatically alert administrators to sudden consumption spikes. Configure these to catch runaway processes before they drain your assist pool.

Property	Description	Default
alert.assist_spike_hours_to_check	Time interval (hours) between scheduled job runs that check for spikes in assist usage	3
alert.assist_spike_usage_percentage_threshold	Percentage increase from the previous job’s results required to trigger a spike notification	0.5 (50%)
alert.assist_spike_usage_threshold	Minimum number of assists consumed before a spike notification can fire	5,000

Pro tip: Do not turn off assists consumption-related notifications, even for sub-production environments. Sub-prod environments consume assists from your contracted pool and can generate unexpected spikes during development and testing.

Building a custom forecasting dashboard

Beyond the out-of-the-box dashboards, you can build your own visualization using Performance Analytics with forecasting and threshold indicators. Query the sys_gen_ai_usage_log table to create time-series data, set threshold alerts for approaching your contracted pool limit, and overlay actual consumption against a linear burn-down projection to flag months where you risk overage. The ServiceNow Community article “Build an assists forecasting dashboard” provides step-by-step instructions.

Operational pro tips

Establish baseline metrics in the first 30 days of deployment so you can measure growth, adoption, and anomalies over time.
If Subscription Management shows a sudden spike in one instance, drill into Now Assist Analytics on that specific instance to understand what changed. It could indicate successful adoption—or it could reveal that someone accidentally left a development process running.
Correlate usage with outcomes. Track Now Assist adoption alongside your existing service management KPIs—are departments with higher Now Assist usage also seeing improved resolution times or higher satisfaction scores?
Verify your clone configuration settings. The usage log table should be excluded from cloning by default, but confirm this in your clone settings—if usage data gets cloned from production to development, your analytics will show misleading information.
Create role-specific dashboard views. Executives care about ROI and deflection rates. Service desk managers care about which fulfillers are adopting AI tools. Platform owners care about technical performance and consumption. Build targeted views for each audience.
Schedule monthly reviews of your usage dashboards with stakeholders to identify trends, celebrate wins, and address underutilized capabilities.

Logging and auditability

Every AI interaction on the platform generates execution data that serves as both an operational log and a compliance audit trail. Effective logging practices turn this data into a strategic asset for debugging, chargeback, and continuous improvement.

What the platform logs automatically

ServiceNow captures detailed execution records for every agentic workflow run. These logs include the agent that executed (sn_aia_agent), the execution plan with every step considered and taken (sn_aia_execution_plan), the tools invoked and their inputs/outputs, the use case being automated (sn_aia_usecase), and the final outcome. All cross-platform actions are authenticated using OAuth 2.0 Authorization Code flows, which guarantees complete audit trails.

Tracing an assist charge to a business record

To trace a specific assist charge back to the case, incident, or task that triggered it, join across two tables:

sys_generative_ai_log — contains the reference to the originating case or task record for skill executions.
sys_gen_ai_usage_log — contains the assist count for each invocation.

For AI agent executions specifically, the sn_aia_tool_execution_tier table shows which tier (Small, Medium, or Large) was assigned and therefore how many assists were consumed. Cross-reference this with sn_aia_execution_plan to see the full reasoning chain and tool sequence.

Logging best practices

Enable comprehensive monitoring from day one of agent deployment. Do not wait until an issue arises—you need baseline data for comparison.
Track acceptance rates for AI-generated outputs (summaries, resolution notes, categorizations). Acceptance rates typically range from 60–100% and directly impact the productivity value of task automation.
Log deflection events explicitly. A deflection occurs when a user interacts with AI self-service and does not submit a ticket within 24 hours. ServiceNow counts it as a deflection if the user gives a thumbs-up, engages by scrolling or expanding the response, or spends at least 15 seconds viewing it. Calculate your deflection rate by dividing total deflections by the sum of deflections plus tickets created.
Preserve execution plans (sn_aia_execution_plan) for every production agent. These plans are essential for debugging non-deterministic behavior—they show the exact reasoning chain and tool invocations for a given run.
Use the AI Agent Analytics dashboard (Performance Analytics indicators filtered by Application = Now Assist) as the first diagnostic step when something looks off. Drill into individual execution records only after identifying the pattern.

Aggregation and reporting

Use ServiceNow Performance Analytics to build and analyze AI metrics. Partner with business owners to identify key indicators and build PA dashboards or workspaces at a monthly or quarterly cadence. For AI Control Tower users, the Observe capability (enhanced at Knowledge 2026 with Traceloop-based runtime observability) provides continuous monitoring with live metrics and alerts, replacing periodic audits with real-time visibility into how agents reason and where they make decisions.

Rate limiting AI

Rate limiting controls the traffic flow to AI capabilities and LLM providers by restricting the number of requests within a given time frame. ServiceNow provides three layers of protection: One Extend rate limit rules for provider-level throttling, trigger-level throttling to prevent runaway agent executions, and recursive check properties to cap execution loops.

One Extend rate limit rules

Navigate to the One Extend Rate Limit Rules [sys_one_extend_rate_limit_rules] table and select New to create a rate limit rule. Role required: admin.

Field	Description	Options / guidance
Name	Identifies the rate limit rule	Use a descriptive name: e.g., “AIA ReAct Engine — instance cap”
Application	Application scope	Must be set to Global
Resource	The target to rate-limit	LLM Provider: select a specific generative AI provider mapping. Capability Definition: select a One Extend Capability Definition (e.g., AIA ReAct Engine for agentic executions). All Capabilities: applies to all providers and all definitions.
Rule Type	Granularity of the rate limit	Instance: accumulates data from all users on the instance. User: keeps the request count per individual user.
Request limit per hour	Maximum requests accepted per hour	Set based on your observed peak usage plus 20–30% headroom. Start conservatively and adjust.
Active	Whether the rule is enforced	Toggle on after testing in sub-prod.

Pro tip: To rate-limit agentic executions specifically, create a rule with Resource = Capability Definition and select the AIA ReAct Engine capability. This caps the total agentic LLM calls without affecting other Now Assist skills.

Trigger throttling (Australia Patch 3+)

ServiceNow provides five configurable system properties in the sn_aia_property table that prevent AI agent triggers from over-executing on the same record within a short period. Any over-use within a time window is classified as a “breach.” The system can warn about runaway triggers or automatically deactivate them.

Property	Description	Default
kill_switch.mode	Enforcement mode. off = disabled. warn_only = sends a daily warning email. enforce = warns on Days 1 and 2, then deactivates the trigger on Day 3.	warn_only
kill_switch.max_fires_per_window	Number of fires on the same record that marks it as breaching	5
kill_switch.min_distinct_records	Number of breaching records needed for the observation window to count as runaway	25
kill_switch.window_size	Length of one observation window (minutes)	1,440 (24 hours)
kill_switch.consecutive_windows_duration	Total look-back span across consecutive windows (minutes)	4,320 (3 days)

Pro tip: Start with warn_only to understand your trigger patterns before switching to enforce. Once in enforce mode, a trigger that breaches thresholds for three consecutive days will be automatically deactivated—which could take an agent offline if you have not built redundancy into your design.

Recursive check properties

These properties cap how many executions a trigger can create or update within a defined time window, preventing infinite-loop scenarios where an agent’s output re-triggers itself.

Property	Description	Default
recursive_check.create_max_executions	Maximum matching executions that can create records	50
recursive_check.create_time_window	Time window (minutes) for checking create executions	15
recursive_check.update_max_executions	Maximum matching executions that can update a record	5
recursive_check.update_time_window	Time window (minutes) for checking update executions	15

Agent design patterns that reduce rate limit pressure

Avoid combining “On Record Create” and “On Record Update” triggers on the same agent without clear condition filters—this is the most common source of recursive loops.
Have the agent modify its loop-breaking field (e.g., assignment group or a custom flag field) early in execution so subsequent condition filters exclude already-processed records.
Limit agents to 10 or fewer per use case and keep tools per agent lean—fewer tools mean less LLM context, faster execution, and lower assist consumption.
Set the sn_aia.continuous_tool_execution_limit property to cap how many times a single tool can run continuously without interruption.
Always review and test trigger configurations in sub-prod before activating in production. Refer to the ServiceNow Community article on AI Agent trigger best practices for detailed guidance.

Measuring accuracy per agent

AI agents are non-deterministic systems. They will not produce identical outputs every time, and expecting 100% accuracy is unrealistic. The goal is to establish a measurement regime that quantifies accuracy across meaningful dimensions, identifies degradation early, and drives iterative improvement.

For more details on the figures mentioned in this section, please review this AI Value Framework whitepaper.

Core evaluation metrics

ServiceNow’s evaluation framework measures agent accuracy across three primary metrics:

Metric	What it measures	Scale
Overall task completeness	Whether the agent successfully completed its assigned task, considering all required steps	1 (unsuccessful), 2 (partially successful), 3 (successful)
Tool choice accuracy	Whether the agent selected the most appropriate tool for each step	Rated from Excellent through Moderate to Poor. A drop from Excellent to Moderate typically indicates gaps in agent instructions or tool descriptions.
Tool calling correctness	Whether the agent constructed tool calls correctly—checking accuracy, completeness, and formatting of input parameters	Scored against expected parameter values and formats

Default rating thresholds are customizable to fit your risk tolerance. For Overall Task Completeness, each rating maps to a recommended deployment action—agents scoring poorly should be held back for instruction refinement before production exposure.

The Agent Productivity Score

The Agent Productivity Score quantifies the combined work output of AI and human agents. A 100% score means the agent is completing the maximum possible work actions in a given time period. ServiceNow’s internal data shows human agents complete a qualified work action (comments, work notes) in an average of 8.3 minutes, yielding a theoretical maximum of 7.3 actions per hour. If the AI agent handles 3 actions and the human handles 4, the score is (4+3)/7.3 = 100%, with AI contributing 42%.

The Workflow Automation Score

The Workflow Automation Score measures the percentage of qualified AI actions completed per day compared to customer or human agent actions. A qualified action is a comment, work note, or change to a case record attribute (category, status, priority, assignment). ServiceNow’s internal benchmarks show 43% automation for low-complexity cases and 37% for complex cases.

Accuracy tracking cadence

Timeframe	Activity	Who
Daily	Review the AI Agent Analytics dashboard for tool failure rates, agent slowdowns, and input-shape failures	Platform administrators
Weekly	Examine execution plans (sn_aia_execution_plan) for the worst-performing agents. Correlate with assist consumption spikes in sys_gen_ai_usage_log.	AI architects
Monthly	Run formal agentic evaluation jobs against representative datasets of 20–100 records per agent	AI Center of Excellence
Quarterly	Conduct human review of automated scores to validate alignment with business expectations	Business stakeholders
At every upgrade or patch	Re-run evaluations to catch regressions from platform changes, LLM updates, or prompt modifications	Platform administrators

Per-agent dashboard design

Build a Performance Analytics workspace for each production agent that surfaces: task completion rate over time, tool choice accuracy trend, average assist consumption per execution (from sn_aia_tool_execution_tier), acceptance rate of AI-generated outputs, mean time to resolution for AI-handled vs. human-handled workflows, and error/fallback rate. Filter indicators by Application = Now Assist to access the out-of-the-box set, then extend with custom metrics.

Automated evaluation practices

Manual testing catches obvious failures, but it does not scale. ServiceNow’s Agentic Evaluations capability provides an automated quality validation framework that functions as the quality gate between development and production. Find it in AI Agent Studio → Testing → Start automated evaluation. Required role: sn_aia.admin.

The evaluation development cycle

Plan and build the agent in AI Agent Studio. Configure tools, instructions, and workflows.
Run manual tests in AI Agent Studio to validate basic behavior. Refine agent responses through iterative testing.
Run automated evaluation jobs. Select all three metrics (task completeness, tool calling correctness, tool choice accuracy) for a comprehensive view. Define a representative dataset and execute.
Analyze results. Do not just look at scores—drill into individual executions to understand what the evaluation is telling you. Walk the execution traces for the worst offenders.
Fix and re-evaluate. Adjust instructions, tool descriptions, or permissions. Clone the evaluation for a clean before-and-after comparison, then re-run.
Iterate until metrics meet deployment criteria. Most teams iterate 2–4 times before achieving deployment-ready scores.

Pro tip: Running evaluations consumes assists. Factor evaluation cadence into your assist budget, especially if you plan to run large datasets (50–100 records) on a regular schedule.

Building evaluation datasets

Three dataset construction options are available:

Existing execution logs: Best for agents already running in production—carries real-world input distribution and edge cases.
Fresh logs from AI Agent Studio: Best when iterating on a new agent and validating specific scenarios before production exposure.
Scaled workflow runs: The platform drives the agent across a filtered record set (e.g., the most recent 50 P3 incidents) and evaluates the generated logs. Best for breadth before deployment.

Start small with 10–20 logs to understand how evaluation works, then scale to 20–100 records for production-grade validation.

AI Agent Advisor: opportunity discovery and pre-built datasets

AI Agent Advisor analyzes your business data, surfaces the highest-impact agentic use cases, and automatically generates an evaluation dataset from the clustered records. When you navigate to the auto-evaluation tool, this dataset is pre-selected—run auto-eval without manually choosing or preparing data. This capability does not consume assists.

Connecting evaluations to analytics

When the AI Agent Analytics dashboard shows a tool failing more frequently, an agent slowing down, or a specific input shape failing repeatedly, add that scenario to your evaluation test set so the next run catches the regression. The feedback loop works both ways—evaluation findings inform what to watch in analytics, and analytics anomalies inform what to add to evaluation datasets.

A/B testing for model and prompt changes

When changing the underlying LLM or modifying agent prompts, run parallel evaluation jobs—one against the current configuration and one against the proposed change—using the same dataset. Compare scores side by side. Do not deploy prompt or model changes without a before-and-after evaluation comparison. Third-party LLMs can be selected as part of an agentic workflow as of Yokohama Patch 6.

Governance with AI Control Tower

AI Control Tower is the central governance layer for managing every AI system, agent, and workflow across the enterprise—regardless of where it runs. It operates across five dimensions: Discover (inventory AI assets across 30+ integrations including AWS, Google Cloud, Azure, SAP, Oracle, and Workday), Observe, Govern, Secure, and Measure (value and adoption tracking).

For consumption management specifically, AI Control Tower provides consumption dashboards for proactive monitoring and integrates with the assist spike alerting properties described earlier in this guide. In the Zurich release, the value insights tab from Now Assist Analytics moves into AI Control Tower, consolidating all measurement into a single governance layer.

Implementation checklist

Use this checklist to adopt the practices in this guide incrementally.

Week 1–2: Foundation

Audit all active AI agents, agentic workflows, and configured skills across instances.
Verify Now Assist AI Agents plugin is at version 6.0+ and apply Australia Patch 3 for trigger throttling properties.
Install or update Now Assist Analytics from the ServiceNow Store.
Configure assist spike alert properties (alert.assist_spike_hours_to_check, alert.assist_spike_usage_percentage_threshold, alert.assist_spike_usage_threshold).
Set kill_switch.mode to warn_only on all instances.
Set sn_aia.continuous_tool_execution_limit on all agents as a safety mechanism.
Confirm clone configuration excludes the usage log table.

Week 3–4: Monitoring

Establish per-agent and per-workflow consumption baselines using Now Assist Analytics.
Review Subscription Management for cross-instance pool status.
Set up a monthly review cadence with stakeholders for usage dashboards.
Build custom forecasting dashboards if your contracted pool is tight.

Week 5–6: Accuracy and evaluation

Run initial agentic evaluations against your top 3–5 production agents.
Establish baselines for task completeness, tool choice accuracy, and tool calling correctness.
Build per-agent Performance Analytics workspaces with key accuracy indicators.
Integrate evaluation runs into your development release cycle as a mandatory quality gate.

Week 7–8: Governance

Deploy or configure AI Control Tower for AI asset discovery across instances.
Review trigger throttling data (if in warn_only mode) and determine whether to switch to enforce mode.
Create One Extend rate limit rules for your highest-volume agentic capability definitions.
Document rate limit policies and publish as organizational standards.

Ongoing

Review evaluation results monthly and iterate on agent instructions and tool configurations.
Expand AI Agent Advisor to discover new automation opportunities continuously.
Refine consumption forecasting models based on historical data.
Re-run evaluations at every platform upgrade, patch, or LLM change.

Appendix: quick reference tables

Key system properties (sn_aia_property)

Property	Purpose	Default
kill_switch.mode	Trigger throttling enforcement: off, warn_only, or enforce	warn_only
kill_switch.max_fires_per_window	Fires per record that mark it as breaching	5
kill_switch.min_distinct_records	Breaching records needed for window to count as runaway	25
kill_switch.window_size	Observation window length (minutes)	1,440
kill_switch.consecutive_windows_duration	Total look-back span (minutes)	4,320
recursive_check.create_max_executions	Max executions creating records	50
recursive_check.create_time_window	Create check window (minutes)	15
recursive_check.update_max_executions	Max executions updating a record	5
recursive_check.update_time_window	Update check window (minutes)	15
alert.assist_spike_hours_to_check	Spike check interval (hours)	3
alert.assist_spike_usage_percentage_threshold	Spike percentage threshold	0.5 (50%)
alert.assist_spike_usage_threshold	Minimum assists before spike fires	5,000
sn_aia.continuous_tool_execution_limit	Max continuous executions for one tool	Configurable

Key platform tables

Table	Contents
sys_gen_ai_usage_log	Assists count per invocation (auto-populated; hourly re-calc on failure)
sys_generative_ai_log	Case/task reference per skill execution
sn_aia_tool_execution_tier	Execution plan tier and assists consumed per run
now_assists_skills	Aggregated skill usage within billing window
sn_nowassist_skill_config	Skill configuration records
sn_aia_agent	AI agent definitions and settings
sn_aia_execution_plan	Full execution plans for agentic workflow runs
sn_aia_usecase	Agentic use case definitions
sys_one_extend_rate_limit_rules	One Extend rate limit rule configurations
sn_aia_property	AI Agent system properties (trigger throttling, recursive checks, alerts)

Content in this guide reflects the Australia release (including Australia Patch 3, June 2026) and references system properties, tables, and dashboards available in Now Assist AI Agents plugin version 6.0 and later. Where a feature requires a specific patch level, that requirement is noted inline.

This document is intended to be used alongside the official ServiceNow product documentation and the AI Center of Excellence community articles. It consolidates scattered guidance into a single operational reference.

Table of Contents

Purpose and scope 4
Understanding assists consumption 5
    Assists vs. actions 5
    Agentic workflow tier sizing 5
    Key tables for consumption data 5
Tracking and monitoring assists 6
    Now Assist Analytics 6
    Subscription Management: the cross-instance view 8
    Assist spike alerting 8
    Building a custom forecasting dashboard 8
    Operational pro tips 9
Logging and auditability 10
    What the platform logs automatically 10
    Tracing an assist charge to a business record 10
    Logging best practices 10
    Aggregation and reporting 11
Rate limiting AI 12
    One Extend rate limit rules 12
    Trigger throttling (Australia Patch 3+) 12
    Recursive check properties 13
    Agent design patterns that reduce rate limit pressure 14
Measuring accuracy per agent 15
    Core evaluation metrics 15
    The Agent Productivity Score 15
    The Workflow Automation Score 16
    Accuracy tracking cadence 16
    Per-agent dashboard design 16
Automated evaluation practices 17
    The evaluation development cycle 17
    Building evaluation datasets 17
    AI Agent Advisor: opportunity discovery and pre-built datasets 17
    Connecting evaluations to analytics 18
    A/B testing for model and prompt changes 18
Governance with AI Control Tower 18
Implementation checklist 18
    Week 1–2: Foundation 18
    Week 3–4: Monitoring 19
    Week 5–6: Accuracy and evaluation 19
    Week 7–8: Governance 19
    Ongoing 19
Appendix: quick reference tables 20
    Key system properties (sn_aia_property) 20
    Key platform tables 21

saileshlank

Hi @hussein_alani

Thank you for the article

I was going through your article written last evening to create a custom PA dashboard - But can't find it now? Has that been archived? Can you share the steps again as I was building it and now lost the track

hussein_alani

Hi @saileshlank , I haven't shared such an article on custom PA dashboards. But I recommend checking the NowLearning course on Platform Analytics as that would be the most comprehensive guide to help you create such dashboards.

Vamsi_Krishna07

Hi @hussein_alani ,

Thank you for the insightful article. I'd like to point out a correction: the "Value Insights" dashboard is no longer available in Now Assist Admin Console -> Performance tab. ServiceNow removed it in ZP7, and I'm unsure if it was reintroduced in AP3.

Thanks!

cartergray7

Great resource—having billing, monitoring, rate limits, and evaluation guidance in one place will make managing AI agents much easier. The version-specific notes are especially helpful for smooth implementation.

hussein_alani

Hi @Vamsi_Krishna07 , thanks for the comment! That's right, the value insights dashboard was moved to AICT since Zurich, this is noted in the column "What it answers", thought it was important to note that it was moved instead of removing the entry.

hussein_alani

Hi @cartergray7 , I appreciate the comment! Thanks.