feat(csat): Add WhatsApp utility template analyzer with rewrite guidance (#13575)

CSAT templates for WhatsApp are submitted as Utility, but Meta may
reclassify them as Marketing based on content, which can significantly
increase messaging costs.
This PR introduces a Captain-powered CSAT template analyzer for
WhatsApp/Twilio WhatsApp that predicts utility fit, explains likely
risks, and suggests safer rewrites before submission. The flow is manual
(button-triggered), Captain-gated, and applies rewrites only on explicit
user action. It also updates UX copy to clearly set expectations: the
system submits as Utility, Meta makes the final categorization decision.

Fixes
https://linear.app/chatwoot/issue/CW-6424/ai-powered-whatsapp-template-classifier-for-csat-submissions


https://github.com/user-attachments/assets/8fd1d6db-2f91-447c-9771-3de271b16fd9

lib/captain/csat_utility_analysis_service.rb

@@ -0,0 +1,66 @@
+class Captain::CsatUtilityAnalysisService < Captain::BaseTaskService
+  pattr_initialize [:account!, :message!, { button_text: nil, language: 'en', baseline: {} }]
+
+  def perform
+    api_response = make_api_call(
+      model: GPT_MODEL,
+      messages: [
+        { role: 'system', content: system_prompt },
+        { role: 'user', content: message }
+      ]
+    )
+
+    return api_response if api_response[:error]
+
+    build_result(api_response[:message])
+  end
+
+  private
+
+  def build_result(response_message)
+    parsed = parse_json_response(response_message)
+    return { error: 'Invalid LLM response format' } if parsed.blank?
+
+    core_result(parsed).merge(message: response_message)
+  end
+
+  def core_result(parsed)
+    {
+      classification: normalize_classification(parsed['classification']),
+      optimized_message: parsed['optimized_message'].presence || baseline[:optimized_message]
+    }
+  end
+
+  def system_prompt
+    template = prompt_from_file('csat_utility_analysis')
+    Liquid::Template.parse(template).render(prompt_variables)
+  end
+
+  def prompt_variables
+    {
+      'message' => message.to_s,
+      'button_text' => button_text.to_s,
+      'language' => language.to_s,
+      'baseline_classification' => baseline[:classification].to_s
+    }
+  end
+
+  def parse_json_response(content)
+    raw = content.to_s.strip
+    json = raw.match(/```json\s*(.*?)\s*```/m)&.captures&.first || raw
+    JSON.parse(json)
+  rescue JSON::ParserError
+    nil
+  end
+
+  def normalize_classification(value)
+    normalized = value.to_s.upcase
+    return normalized if %w[LIKELY_UTILITY LIKELY_MARKETING UNCLEAR].include?(normalized)
+
+    baseline[:classification].presence || 'UNCLEAR'
+  end
+
+  def event_name
+    'csat_utility_analysis'
+  end
+end

lib/integrations/openai/openai_prompts/csat_utility_analysis.liquid

@@ -0,0 +1,27 @@
+You are a WhatsApp template compliance assistant.
+Your task is to evaluate whether a CSAT template message is likely to be approved as UTILITY vs MARKETING under Meta policy.
+
+Rules:
+1. Prefer UTILITY only when the message is tied to an existing support or transactional event.
+2. Avoid promotional language, upsell, cross-sell, offers, discounts, or purchase intent.
+3. Keep the rewritten message concise, explicit, and purely transactional.
+4. Do not invent product offers or marketing phrases.
+
+Input:
+- Message: {{ message }}
+- Button text: {{ button_text }}
+- Language code: {{ language }}
+
+Baseline heuristic:
+- Classification: {{ baseline_classification }}
+
+Return ONLY valid JSON with this shape (example):
+{
+  "classification": "LIKELY_UTILITY",
+  "optimized_message": "rewritten utility-safe message"
+}
+
+Allowed values for "classification": "LIKELY_UTILITY", "LIKELY_MARKETING", or "UNCLEAR".
+
+Important:
+- Write `optimized_message` in the same language as `Language code`.