AI Integration & Data Handling

This page is the integration design reference for the AI features in SN Utils. It is written for security reviewers, procurement, and vendor-onboarding teams who need to know exactly how the extension talks to a Large Language Model (LLM) and what data is included in each request. It complements the high-level Security & Privacy overview and the Privacy & Data Processing Agreement.

Architecture at a Glance

There are two user-facing AI features:

Both follow the same integration principle:

ServiceNow tab / Manage page
        │  (user clicks Send / Generate)
        ▼
SN Utils extension  ──HTTPS + Bearer token──▶  https://snutils.com/api/extension/ai/*
                                                        │
                                                        ▼
                                            LLM provider (OpenAI / Anthropic / Google)

The extension's Content Security Policy restricts outbound connections to https://*.service-now.com (the user's own instance) and https://snutils.com. There is no LLM-provider host in the allow-list, which makes a direct call impossible by design.

AI Endpoints

Both require a valid Authorization: Bearer <token> header, are served over TLS, stream their responses as Server-Sent Events, and are rate-limited to 50 requests per hour per user.

Sidekick AI — Data Sent

When Sidekick AI is enabled (it is opt-in, off by default), each request contains the chat messages plus a small metadata-only context object describing the current page. No field values or record data are included.

The context object can contain:

• Page type — form, list, flow, script_editor, workspace, record, or unknown.
• Table name — e.g. incident, when detectable from the page.
• Field names — up to 50 field names from the current form or list. System fields (sys_*, sysparm_*) are filtered out. Names only, never values.
• Form view — the view name (e.g. Default).
• URL path — location.pathname only. Query strings, hash fragments, and any record sys_id in the URL are not included.
• Instance subdomain — the first label of the hostname (e.g. dev12345 from dev12345.service-now.com), so the assistant can distinguish between instances. The full hostname is not sent.

Sample Sidekick AI request payload

{
  "messages": [
    {
      "role": "user",
      "content": "Write a GlideRecord query for open incidents assigned to me"
    }
  ],
  "context": {
    "pageType": "form",
    "tableName": "incident",
    "fieldNames": [
      "number",
      "short_description",
      "state",
      "assigned_to",
      "priority"
    ],
    "formView": "Default",
    "urlPath": "/incident.do",
    "hostname": "dev12345",
    "tier": "metadata"
  },
  "model": "claude-sonnet-4-6"
}

The hostname field carries only the instance subdomain label, not the fully qualified domain. The tier is fixed to metadata in the current product; a future opt-in "full context" tier would be disabled by default and require explicit consent.

AI Command Builder — Data Sent

The AI Command Builder generates slash commands and snippets. A request contains:

1. Your chat prompt — the text typed in the AI panel.
2. The contents of the editor form, if you have started filling it in — for commands: name, hint, description, URL/script, fields, overwrite URL, inline-only flag; for snippets: prefix, name, description, category, context, and the code in the editor. Long fields (script / snippet body) are truncated before sending.
3. Recent chat turns (up to the last 10 messages) so follow-ups keep context.
4. Table field schema (command generation only — see below).

The full prompt is capped at ~3,900 characters client-side.

Table field schema enrichment (command generation)

When you generate a command whose URL targets a detectable table, the builder asks the LLM to use real field names. To do that it reads the table's field dictionary from an open ServiceNow tab — using your existing session and permissions — and includes a compact schema block in the prompt:

• It queries sys_dictionary for the table and its ancestors and sends field name, label, internal type, and the display flag — never any record values.
• The lookup runs under the logged-in user's own rights (no elevated access) and the result is cached locally per instance for 24 hours.
• If no ServiceNow tab is available, or the lookup is slow, it is skipped — generation continues without it.

Sample AI Command Builder request payload

{
  "prompt": "TABLE METADATA — incident (live from instance)\nDisplay field: number\nAvailable fields (name [type] label):\n- number [string] Number *display*\n- short_description [string] Short description\n- state [integer] State\n- assignment_group [reference] Assignment group\nWhen generating \"fields\", use ONLY names from the list above.\n\nCURRENT COMMAND IN FORM:\n{\n  \"command\": \"\",\n  \"hint\": \"\",\n  \"description\": \"Search incidents by assignment group\",\n  \"script\": \"\",\n  \"fields\": \"number,short_description\",\n  \"overwrite_url\": \"\",\n  \"inline_only\": false\n}\n\nUSER REQUEST: search incidents by assignment group and show priority",
  "type": "command",
  "stream": true,
  "messages": [
    {
      "role": "user",
      "content": "Create a command to search incidents by assignment group"
    }
  ],
  "model": ""
}

The schema block lists field names, labels, and types only. An empty model means the server picks the default model.

What Is Never Sent

Across both features, the following never leaves the browser as part of an AI request:

• ServiceNow record data — no GlideRecord values, no sys_ids, no field values from any record.
• Cookies, session tokens, or credentials from your ServiceNow session. (The session token used to read field schema for command generation stays between your browser and your own instance — it is never sent to SN Utils.)
• Attachments or file contents.
• Query strings, URL fragments, or filters that might contain record identifiers or personal data.
• Page DOM beyond the metadata listed above.
• Anything passive — a request is made only when you click Send / Generate.

If you want the AI to reason about a specific value, script, or error message, you paste it into the chat yourself — keeping you in control of exactly what is shared.

Authentication & Transport

• All requests use HTTPS / TLS.
• Requests are authenticated with the user's Pro access token as a Bearer token. Tokens live in the browser's extension storage and are never exposed to the ServiceNow page context.
• The backend validates the token and the request origin before forwarding anything to a provider.

LLM Providers & Sub-Processors

The backend routes each request to an LLM provider based on the selected model. Providers currently include OpenAI, Anthropic, and Google. Users may pick a model (or leave it on Auto, the recommended default):

• Auto (server default)
• Claude Sonnet 4.6 · Claude Fable 5 · Claude Opus 4.8 (Anthropic)
• GPT-5.5 (OpenAI)
• Gemini 3.5 Flash (Google)

The authoritative list of sub-processors, along with their processing locations and retention terms, is maintained in the Privacy & Data Processing Agreement.

Retention & Logging

• Conversations are not retained as user content. Sidekick chat history lives only in the browser tab's session storage and is cleared when the tab closes; the AI Command Builder chat is kept in memory and cleared when you switch away.
• For operational and abuse-prevention purposes the backend keeps a short request log (feature, model, outcome, duration, token counts, and a truncated prompt snippet). It does not store full conversations.
• The table field-schema cache (names/labels/types only) lives in the browser's local extension storage with a 24-hour expiry.

Controls Summary

Related

• Security & Privacy — High-level security overview
• Privacy & Data Processing Agreement — Sub-processors, retention, legal terms
• AI Assistant — Using the AI Command Builder
• Sidekick AI — Using the in-page assistant