Using Sensitive Topic Filtering with Now Assist for Virtual Agent Without HRSD

When introducing AI into Virtual Agent, the goal is simple: give users faster, smarter answers while still keeping control over what the AI should and should not respond to.

In a standard HRSD implementation, ServiceNow provides a more visible path for configuring sensitive topic handling through Now Assist Guardian and the Now Assist Admin experience. That makes sense for HR use cases, where topics such as employee relations, wellbeing, grievances, and personal matters need careful handling.

But what happens when HRSD is not installed?

This post walks through how Sensitive Topic filtering can still be configured for Now Assist for Virtual Agent by using the underlying GenAI filter framework. The capability is not limited to HRSD conceptually, but some of the admin experience is easier to access when HRSD is present.

The Scenario

I am working with Now Assist for Virtual Agent, but HRSD is not implemented.

The requirement is to govern AI-generated responses and prevent the assistant from responding directly to sensitive or restricted topics. Instead of allowing the AI to generate an answer, the conversation should be redirected into a controlled Virtual Agent topic.

To solve this, I need a way for ServiceNow to:

1. Detect when a user asks something that broadly matches a sensitive topic.
2. Prevent the normal AI response from being returned.
3. Redirect the user into a designed Virtual Agent fallback or error topic.
4. Apply the filter specifically to Now Assist for Virtual Agent.

This is where the sys_gen_ai_filter framework becomes useful.

What is Sensitive Topic Filtering?

Sensitive Topic filtering allows administrators to define topics that should be handled carefully by Now Assist and Generative AI capabilities.

Instead of relying only on prompt design or user guidance, the platform can evaluate the user's input against configured filters. If a sensitive topic is detected, the normal AI response path can be interrupted and the user can be redirected to a specific Virtual Agent topic.

ServiceNow documents this behaviour as part of Now Assist Guardian, where sensitive topic filters can redirect Virtual Agent users when sensitive subjects are detected. ServiceNow also documents that sample phrases can be configured for filters and that applicability can be selected for active Virtual Agents. :contentReference[oaicite:0]{index=0} :contentReference[oaicite:1]{index=1}

Why This Matters Outside HRSD

Sensitive topic handling is often associated with HRSD, but the need is broader than HR.

The same governance pattern can apply to:

• Legal or compliance questions
• Security-related questions
• Financial advice
• Health and safety guidance
• Internal policy exceptions
• Restricted operational procedures
• Topics requiring human review before advice is given

In these cases, the objective is not always to block the user. A better approach is usually to guide them safely.

For example, rather than allowing the AI to answer a sensitive compliance question directly, the Virtual Agent can redirect the user to an approved topic that explains the correct process, captures context, or routes the conversation to the right team.

The Key Platform Tables

The main table used for this configuration is:

sys_gen_ai_filter

This is the Generative AI Filter table.

The related sample utterances are stored in:

sys_gen_ai_filter_sample

ServiceNow documentation for generative AI filters references the Generative AI Filter [sys_gen_ai_filter] table and the use of a Filter Error Topic to redirect users when a filter is detected. :contentReference[oaicite:2]{index=2}

The important point is that the underlying framework exists as part of the Now Assist and GenAI filtering capability, even if the full HRSD-oriented navigation experience is not available.

Prerequisites

Before configuring this, make sure the following are in place:

• Access to Virtual Agent Designer or Assistant Designer.
• Now Assist and Generative AI capabilities are enabled.
• Admin access to the required system tables.
• A clear understanding of the sensitive topic you want to detect.
• A designed fallback or handling topic for users who trigger the filter.

Step 1: Create a Filter Error Topic

Start in Assistant Designer or Virtual Agent Designer.

Create a topic that will handle the sensitive topic scenario. This is the topic the user will be redirected to when the filter is triggered.

The topic should not feel like a hard error unless that is genuinely the right experience.

A better pattern is to use the topic to:

• Acknowledge that the request needs careful handling.
• Explain what the user can do next.
• Offer a safe route to approved information.
• Create a case, request, or interaction if human review is required.
• Route to a live agent where appropriate.

For example, rather than saying:

Blocked. This topic is not allowed.

A more helpful response would be:

This topic needs to be handled by the appropriate team. I can help route your request so it is reviewed correctly.

The quality of this topic matters. It is the experience the user sees when the AI decides not to answer directly.

Step 2: Create the Sensitive Topic Filter

Next, create a new record in:

sys_gen_ai_filter

Complete the key fields as follows:

The threshold controls how closely the user's input needs to align with the sample utterances before the filter triggers.

A lower threshold may catch more variations but can increase false positives. A higher threshold may reduce false positives but miss more loosely worded queries.

A value around 0.81 is a practical starting point, but it should be validated against real examples before being treated as final.

Step 3: Add Sample Utterances

The next step is to define the phrases that represent the topic you want to catch.

These are stored in:

sys_gen_ai_filter_sample

A useful tip is to add this table as a related list on the sys_gen_ai_filter form. That makes managing the sample utterances much easier.

Create child sample records linked to your filter. Add representative phrases that users may type when asking about the sensitive topic.

Do not only add one perfect phrase. Users rarely ask questions in one consistent way.

For example, include variations such as:

• Direct questions
• Informal wording
• Short phrases
• Longer natural language questions
• Common misspellings where relevant
• Different ways of referring to the same topic

ServiceNow's Now Assist Guardian FAQ also notes that filters can be configured by adding sample phrases through the setup dialogue or directly in the sys_gen_ai_filter_sample table. :contentReference[oaicite:3]{index=3}

Step 4: Enable Applicability in Now Assist Admin

Once the filter exists, open the Now Assist Admin filter experience directly:

/now/now-assist-admin/filters

Locate your filter and open it.

In the Applicability tab, enable:

NAVA

This makes the filter applicable to Now Assist for Virtual Agent.

Without this step, the filter may exist in the table but not apply to the NAVA runtime experience.

How the Runtime Behaviour Works

Once configured, the user experience is straightforward.

1. The user enters a message in Virtual Agent.
2. Now Assist evaluates the message against the active filters.
3. If the message is similar enough to the configured sample utterances, the filter is triggered.
4. The normal AI-generated response is interrupted.
5. The configured Filter Error Topic is executed instead.

This gives administrators a controlled way to handle sensitive or restricted areas without relying solely on the model to make the right judgement in every scenario.

Known Limitation: Filter Visibility in Navigation

One behaviour that can cause confusion is that configured filters may not appear in the standard Application Navigator.

This is expected in many cases.

The sys_gen_ai_filter table is part of the Now Assist Admin configuration experience. It is not necessarily exposed as a standard UI16 module in the Application Navigator.

This means administrators may be able to create and manage the records directly, but not see a friendly navigation module unless the relevant admin experience or application configuration exposes it.

When Does It Appear Natively?

Filters are more naturally exposed when using the Now Assist Admin experience, particularly where Now Assist Guardian and HRSD-related capabilities are installed and configured.

The direct workspace path is:

/now/now-assist-admin/filters

ServiceNow documentation describes sensitive topic filters as appearing in the Active tab of the Filters section of Now Assist Guardian settings once activated. :contentReference[oaicite:4]{index=4}

However, this does not mean a UI16 Application Navigator module will automatically be created for every implementation.

Recommended Workaround

If administrators need easier access, create a custom UI16 module that points directly to:

/now/now-assist-admin/filters

This gives admins a clear navigation entry without changing the underlying filter framework.

For table-level access, admins can also navigate directly to:

sys_gen_ai_filter.list

and:

sys_gen_ai_filter_sample.list

Access should still be controlled carefully, as these records govern how AI interactions are handled.

Testing the Filter

Testing is the most important part of this configuration.

Do not test only the exact phrases that were added as samples. That only proves the obvious path.

Test with:

• Exact sample utterances
• Similar wording
• Short phrases
• Longer questions
• Borderline phrases that should not trigger
• Unrelated questions that must continue through the normal AI path

The aim is to find the right balance between protection and usability.

If the filter triggers too often, review the sample utterances and threshold. If it does not trigger enough, add more representative samples or adjust the threshold based on evidence.

Best Practice

Treat Sensitive Topic filters as part of your AI governance model, not as a one-off configuration task.

• Use clear, representative sample utterances.
• Test with realistic user language, not just admin-created examples.
• Design the Filter Error Topic to guide users safely.
• Avoid blunt blocking unless there is a clear reason.
• Document the purpose of each filter.
• Record who owns the filter from a process or policy perspective.
• Review filters periodically as AI use cases evolve.
• Use direct workspace access or a custom module to make administration easier.

The Takeaway

Sensitive Topic filtering is not just an HRSD feature. The underlying GenAI filter framework can be used to govern Now Assist for Virtual Agent in broader enterprise scenarios.

The key is to configure the filter record, define useful sample utterances, link the filter to a controlled Virtual Agent topic, and enable applicability for NAVA in Now Assist Admin.

There is a navigation limitation to be aware of. Without the right admin experience or HRSD-related exposure, filters may not appear naturally in the standard Application Navigator.

That does not prevent the approach from working.

By using the sys_gen_ai_filter framework and the Now Assist Admin filter page directly, organisations can introduce a scalable governance layer for AI-powered Virtual Agent conversations.

The result is a better balance between innovation and control: users still get the benefit of AI-powered self-service, while sensitive topics are handled through a designed and auditable path.