What Is Shadow Mode?
Shadow Mode lets AI observe human-handled support tickets and post draft replies as internal notes instead of sending real messages to players. Human agents see what the AI would have said — without any impact on the actual conversation. This enables:- Safe onboarding — see AI quality on real conversations before going live
- Copilot workflow — agents get AI-drafted replies they can review and copy-paste
- Instruction tuning — spot where AIP instructions need improvement using real tickets
- Gradual rollout — enable shadow on specific AIPs to validate quality before full automation
Shadow Mode is a workspace-level feature that must be enabled by your CSM via a feature flag before it appears in Settings.
How It Works
- A customer sends a message on the helpdesk
- The ticket is assigned to a human agent (auto-escalated with Shadow Mode type)
- AI pipeline runs normally — classification, AIP routing, response generation
- Instead of sending to the customer, the response is posted as an internal note on the helpdesk
- All AI actions (escalate, resolve, transfer) also become notes instead of real actions
What the Player Experiences
Nothing changes. The player only interacts with the human agent. AI never sends a message to the player.What the Agent Sees
- Normal ticket in their helpdesk queue
- Internal notes from Cevro prefixed with
[Shadow]showing what the AI would have replied - If the AI would escalate, a note explains why (e.g., “Cevro would escalate here. Reason: player requesting account closure”)
- If the AI would send an external message (email, Slack), a note describes what would have been sent
Setting Up Shadow Mode
Navigate to Shadow Mode settings
Go to Settings > Agent Configuration > Shadow Mode. This page is only visible if the feature flag is enabled for your workspace.
Choose AIP scope
Select which AIPs should run in shadow mode:
- All responses — shadow mode applies to every AIP in the workspace. The greeting is suppressed and the ticket is immediately escalated to a human agent.
- Only select AIPs — pick specific AIPs from the list. The AI will still send the greeting message to the player normally, but once the ticket is classified into a selected AIP, all subsequent AI responses become shadow notes instead of real messages.
Supported Channels
| Channel | Supported |
|---|---|
| Intercom | Yes |
| LiveChat | Yes |
| Zoho | No |
| Cevro | Yes |
| Zendesk | Yes |
| Respond.io | No |
What Gets Suppressed
Shadow Mode intercepts all outbound AI behavior. Nothing reaches the player or external systems.| AI Action | Normal Behavior | Shadow Mode Behavior |
|---|---|---|
| Send message | Message sent to player | Posted as internal note with [Shadow] prefix |
| Escalate | Ticket transferred to human team | Note posted: “Cevro would escalate here” (once per ticket) |
| Resolve | Ticket closed | Note posted instead of closing the ticket |
| Transfer | Ticket moved to another group | Note posted instead of transferring |
| Action tools | Tool executes (e.g., issue bonus) | Suppressed — returns synthetic result so AI continues, but no mutation happens |
| Data tools | Tool queries data (e.g., get player info) | Executes normally — AI works with real player data |
| External messages | Email/Slack sent to external target | Note posted: “Cevro would send external message…” |
AI Context in Shadow Mode
Shadow notes are excluded from the AI’s own conversation context. When the AI generates a new response, it only sees:- Real player messages
- Real human agent replies
Human agent replies are captured via helpdesk webhooks and stored so the AI can see what the human actually said. This gives the AI full conversation context for generating relevant shadow drafts.
Analytics & Costs
- Shadow tickets are excluded from all production metrics — they won’t skew your automation rate, CSAT, or resolution stats
- LLM costs are tracked separately — shadow mode calls are tagged so you can filter them from production costs
- Escalation events are logged with a distinct Shadow type for easy identification in the timeline
Key Behaviors
One escalation note per ticket
One escalation note per ticket
If the AI decides to escalate multiple times during a shadow conversation, only the first escalation posts a note. Subsequent attempts are silently skipped to avoid noise.
System errors don't generate notes
System errors don't generate notes
If the AI encounters an internal error that would normally trigger an escalation, no shadow note is posted. Only intentional AI decisions produce notes.
Greeting behavior depends on AIP scope
Greeting behavior depends on AIP scope
When set to All responses, the greeting is suppressed entirely — the ticket is immediately escalated to a human agent. When set to Only select AIPs, the greeting is sent normally because shadow mode activates only after AIP classification.
Works with AIP scoping
Works with AIP scoping
When set to “Only select AIPs”, shadow mode activates on ticket creation but only generates notes for messages routed to the selected AIPs.
Common Questions
Can I run shadow mode on some AIPs while others handle tickets normally?
Can I run shadow mode on some AIPs while others handle tickets normally?
Yes. Set AIP scope to Only select AIPs and choose the ones you want to shadow. All other AIPs will continue to operate normally.
Will shadow mode affect my existing tickets?
Will shadow mode affect my existing tickets?
No. Shadow mode only applies to new tickets created after it’s enabled. Existing conversations are unaffected.
Can agents copy-paste shadow responses?
Can agents copy-paste shadow responses?
Yes. That’s one of the primary use cases. Agents can review the AI’s draft in the internal note and use it as a starting point for their reply.
Why don't I see Shadow Mode in my settings?
Why don't I see Shadow Mode in my settings?
Shadow Mode requires a feature flag to be enabled on your workspace. Contact your CSM to enable it.
Why is Zendesk not supported?
Why is Zendesk not supported?
Zendesk’s Sunshine Conversations API does not support internal notes. Support may be added in the future via the Zendesk Support API.
Related Documentation
- Content Shield — Detect sensitive content and silently escalate to human agents
- Ticket Escalation vs. Unresolved Tickets — Understanding different types of handoffs
- Human-in-the-Loop (HITL) — How HITL requests work and how to use them in AIPs