Use these comparisons to predict behavior before investigating a missing button or an unexpected approval. They describe product rules; internal QA records track which combinations were actually exercised in the captured examples.
Availability and execution
| Combination | Expected experience |
|---|---|
| AI Assist off | The support session does not offer AI Assist through this feature |
| AI Assist on, AI Agent off | Conversational AI Assist is available; Agent mode is not offered |
| AI Agent on, incompatible client | Ask remains useful; Agent is unavailable with a compatibility explanation |
| Compatible device connected, Remote Control absent | Agent is unavailable until the normal Remote Control requirement is met |
| Agent available, user privilege | User-role tools can run; required protected work can trigger an inline admin offer |
| Administrator access present, low/medium/high set to Human-in-the-loop | Protected operations can be eligible, but changing operations still wait for approval |
| Read-only set to Human-in-the-loop | Ordinary diagnostics wait for review; the initial environment probe remains automatic |
| One pending plan contains both automatic and review-required steps | The proposed batch waits as a block |
| A tool is disabled in the active effective policy | It is unavailable; a chat instruction or an old approval cannot grant permission to use it |
| Software tools enabled, High set to Auto-execute | Software and driver suites remain unavailable |
| A skill is enabled, but its required tool is off | The procedure cannot enable that tool; expect an available alternative, limitation, or clarification |
Generated-code behavior
| Combination | Expected experience |
|---|---|
| Organization allows generation; session Code Off | The agent does not generate scripts automatically |
| Organization allows generation; session review mode | Each proposed script waits for review |
| Organization also allows auto-run; session still review | Scripts continue to wait; the allowance alone changes nothing |
| Session switches from review to auto-run with an existing held script | That existing script remains held for review |
| Session turns generation Off with a pending agent-written script | That generated binding cannot dispatch |
| Session Code Off, organization allows explicit code execution | The technician can still choose an eligible code block's Execute path; it enters the applicable review workflow |
| Start over while Code is enabled | Conversation context resets, but the Code mode is retained |
Source and authoring behavior
| Combination | Expected experience |
|---|---|
| Ask with enabled KB/web sources | Retrieval can occur even though device tools are not dispatched |
| A saved knowledge/web source is off | Ordinary conversation does not use it; an administrator can still deliberately test it |
| Entire Web selected | Trust is Unverified; Every new topic is unavailable |
| Organization knowledge fully answers a non-explicit unverified-web request | Web search can be skipped and organization evidence reused |
| Source search failed | Unavailable/failed outcome, not a valid “no matching articles” answer |
| Custom tool saved but not published | A durable draft exists; it is not available to the agent as a published capability |
| AI proposal accepted but not saved | Only the working draft changed |
| Custom draft tested successfully | Only the exact tested snapshot/context has been validated; publication remains separate |
| A new skill saved but not enabled | The procedure is stored but unavailable for selection |
| Preview gives a successful simulated tool result | AI behavior was exercised; actual native-device execution is still unverified |
Suggested rollout sequence
Start with a small, well-understood built-in tool set and the default approval split. Validate representative support questions in a new session. Add focused organization knowledge and verify citations. Introduce tested custom tools or procedures when the built-in behavior does not meet a real support need. Review generated-script and installation allowances as separate decisions with the people responsible for device changes.
Change one material setting at a time during QA. Record the starting values, create a fresh session where required, send the same representative request, and compare the plan, gate, result, and explainability evidence. A correct-looking configuration screenshot alone does not prove enforcement at dispatch.