Actions Wavo can take (approvals, drafts, queries)

Wavo's tools fall into 3 modes: read (instant, no approval), write (instant + audited), approval (pending → user confirms → execute). Risk classification via TOOL_RISK_MAP in approval-payload.ts: HIGH / MEDIUM / LOW. Approval requests expire after 30 minutes (ACTION_TTL_MS=1800000). Every approval logged to ai_action_log with SHA-256 payload hash + idempotency key.

Who is this for

Anyone wondering "what can Wavo actually change in my account?" or "why did Wavo ask me to confirm?"

The 3 tool modes

Mode 1: read

Runs immediately. No approval. No risk to your account.

Examples:

• campaign_search, campaign_performance_summary

• get_performance_data, get_audience_insights

• get_tracker_data, compare_tracker_meta

• cdm_orders_search, cdm_products_search

• list_automation_rules, draft_get, draft_validate

• analyze_existing_ad, analyze_landing_page

• web_search

Wavo invokes, gets data, answers. Done.

Mode 2: write

Runs immediately + writes audit log entry. LOW risk by design.

Examples:

• create_campaign_draft — creates a draft (not published; safe to edit)

• generate_image, generate_video, generate_copy — creates Creative Hub assets

• create_support_ticket — opens Pylon ticket

No approval card. Wavo just does it + tells you what happened.

Mode 3: approval

Pending → user confirms → execute. Required for HIGH and MEDIUM risk tools.

Flow:

1. Wavo plans an action

2. Generates an approval card with:

   • Tool name

   • Risk level (HIGH / MEDIUM)

   • Payload preview (what will change)

   • Execute / Reject buttons

3. You review + click Execute or Reject

4. If Execute: action runs, audit log records SUCCESS

5. If Reject: audit log records DENIED with reason

6. If 30 min passes with no decision: card expires, audit log records EXPIRED

TOOL_RISK_MAP

HIGH risk = significant, hard-to-reverse changes (deleting rules, duplicating campaigns, changing budgets). MEDIUM = reversible but worth confirming. LOW = safe but tracked.

Why approvals matter

LLMs can misinterpret intent. Wavo may understand "increase budget by 20%" as "increase budget by 200%" if your prompt is ambiguous. Approval card lets you:

• Review the exact payload before it executes

• Reject if Wavo got it wrong

• Audit what was approved + by whom

This is a deliberate safety design, not a bug. Cannot be disabled.

Approval card contents

A typical approval card shows:

Action: update_entity_budget
Risk: HIGH
Target: Campaign "Spring Sale Prospecting"
Change: daily_budget €50 → €60 (+20%)

[Execute]  [Reject]

Expires in: 29 min 47 sec

What you see depends on the tool:

• delete_automation_rule: rule details + scope + last 5 executions

• duplicate_entity: source + destination + new name

• update_entity_budget: current + new value + delta %

• create_automation_rule: full rule definition (conditions, actions, schedule)

• toggle_entity_status: from / to state

Audit log

Every approval-mode action writes to ai_action_log:

Idempotency: if you accidentally approve the same action twice (e.g. double-click), the second execution is a no-op — same hash, same key.

Query the audit log via /api/v1/audit-logs?resource_type=ai_action (admin / owner only).

Multiple actions in one turn

Wavo can plan multiple actions. Each gets its own approval card. You approve / reject independently.

Example: "Pause all losing adsets, then create a rule to prevent future losses."

• Approval card 1: bulk toggle_entity_status (HIGH risk)

• Approval card 2: create_automation_rule (MEDIUM risk)

You can approve one + reject the other.

ACTION_TTL — 30 min

ACTION_TTL_MS=1800000. After 30 min:

• Approval card expires

• Audit log status flips PENDING → EXPIRED

• Action cannot be executed via that card

If you want to proceed after expiry: ask Wavo again. New approval card generated with fresh TTL.

Why 30 min: balance between giving you time to think + preventing stale approvals from being acted on if your context changed.

Read vs Write mode timing

For async writes (generate_*): Wavo returns immediately + you watch via Jobs panel (see ch-118).

Cross-cluster action flow

When Wavo creates a campaign:

1. create_campaign_draft (LOW write) — draft saved in DB

2. Optional: draft_validate (read) — checks for errors

3. To publish: requires manual approval via Campaign Creator UI (Wavo doesn't auto-publish; PRD-15 cc-* explains)

When Wavo creates a rule:

1. create_automation_rule (MEDIUM approval) — approval card

2. Execute → rule created with status: paused

3. User activates manually (Wavo doesn't auto-activate; PRD-17 rul-* explains)

Defense-in-depth: Wavo doesn't run wild, even with approval.

Common surprises

• "Wavo did something I didn't approve" — only LOW write tools run without approval (drafts, generate). Anything HIGH/MEDIUM requires explicit approval.

• "The approval card disappeared" — 30-min TTL expired. Ask Wavo again.

• "Two approval cards for the same action" — Wavo planned multiple steps; idempotency key prevents duplicate execution if approved twice.

• "I rejected but Wavo still mentions the rule" — Wavo remembers the planning; nothing was actually executed. Audit log confirms.

Related

• Create your first rule — rule actions via Wavo

• Launch Express mode — campaign actions via Wavo

• Data grounded — what Wavo sees before acting