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

app/services/csat_template_utility_analysis_service.rb

@@ -0,0 +1,86 @@
+class CsatTemplateUtilityAnalysisService
+  include CsatTemplateUtilityRubric
+
+  pattr_initialize [:account!, :inbox!, :message!, { button_text: nil, language: 'en' }]
+
+  def perform
+    baseline = rule_based_result
+    return baseline if baseline[:classification] == 'LIKELY_MARKETING'
+
+    llm_result = llm_result_or_nil(baseline)
+    llm_result || baseline
+  end
+
+  private
+
+  def llm_result_or_nil(baseline)
+    llm_output = Captain::CsatUtilityAnalysisService.new(
+      account: account,
+      message: message,
+      button_text: button_text,
+      language: language,
+      baseline: baseline
+    ).perform
+
+    return nil if llm_output[:error]
+
+    normalize_llm_result(llm_output, baseline: baseline)
+  rescue StandardError => e
+    Rails.logger.error("CSAT utility LLM analysis failed for inbox #{inbox.id}: #{e.message}")
+    nil
+  end
+
+  def normalize_llm_result(result, baseline:)
+    classification = normalized_classification(result[:classification], baseline: baseline)
+    optimized_message = result[:optimized_message].presence || baseline[:optimized_message]
+    optimized_message = baseline[:optimized_message] if baseline[:classification] == 'LIKELY_MARKETING'
+
+    {
+      classification: classification,
+      optimized_message: optimized_message
+    }
+  end
+
+  def normalized_classification(value, baseline:)
+    raw = value.to_s
+    return 'LIKELY_MARKETING' if baseline[:classification] == 'LIKELY_MARKETING'
+
+    raw
+  end
+
+  def rule_based_result
+    text = sanitized_message
+    marketing_hits_count = MARKETING_PATTERNS.count { |pattern| pattern.match?(text) }
+    utility_hits_count = UTILITY_PATTERNS.count { |pattern| pattern.match?(text) }
+    criteria = evaluate_criteria(text: text, marketing_hits_count: marketing_hits_count)
+    classification = classify(criteria: criteria, utility_hits_count: utility_hits_count)
+    build_rule_payload(
+      classification: classification
+    )
+  end
+
+  def build_rule_payload(payload)
+    {
+      classification: payload[:classification],
+      optimized_message: optimized_message_for(payload[:classification])
+    }
+  end
+
+  def sanitized_message
+    message.to_s.squish
+  end
+
+  def classify(criteria:, utility_hits_count:)
+    return 'LIKELY_MARKETING' unless criteria[:marketing_prohibition]
+    return 'LIKELY_MARKETING' unless criteria[:prohibited_content]
+    return 'LIKELY_UTILITY' if criteria.values.all? && utility_hits_count >= 2
+
+    'UNCLEAR'
+  end
+
+  def optimized_message_for(classification)
+    return sanitized_message if classification == 'LIKELY_UTILITY'
+
+    build_input_aware_utility_message
+  end
+end