diff --git a/changelog/10458.bugfix.md b/changelog/10458.bugfix.md
new file mode 100644
index 000000000000..90a85cbf1c5f
--- /dev/null
+++ b/changelog/10458.bugfix.md
@@ -0,0 +1 @@
+Make `action_metadata` json serializable and make it available on the tracker. This is a backport of a fix in 3.0.0.
diff --git a/rasa/core/policies/ensemble.py b/rasa/core/policies/ensemble.py
index 75b9d7d1fcd7..ad5cff602c9b 100644
--- a/rasa/core/policies/ensemble.py
+++ b/rasa/core/policies/ensemble.py
@@ -655,6 +655,7 @@ def _pick_best_policy(
             is_no_user_prediction=best_prediction.is_no_user_prediction,
             diagnostic_data=best_prediction.diagnostic_data,
             hide_rule_turn=best_prediction.hide_rule_turn,
+            action_metadata=best_prediction.action_metadata,
         )
 
     def _best_policy_prediction(
diff --git a/rasa/core/policies/unexpected_intent_policy.py b/rasa/core/policies/unexpected_intent_policy.py
index c3d0b8e50ba8..019e7d775dfc 100644
--- a/rasa/core/policies/unexpected_intent_policy.py
+++ b/rasa/core/policies/unexpected_intent_policy.py
@@ -488,11 +488,11 @@ def _collect_action_metadata(
         def _compile_metadata_for_label(
             label_name: Text, similarity_score: float, threshold: Optional[float],
         ) -> "RankingCandidateMetadata":
-            severity = threshold - similarity_score if threshold else None
+            severity = float(threshold - similarity_score) if threshold else None
             return {
                 NAME: label_name,
-                SCORE_KEY: similarity_score,
-                THRESHOLD_KEY: threshold,
+                SCORE_KEY: float(similarity_score),
+                THRESHOLD_KEY: float(threshold) if threshold else None,
                 SEVERITY_KEY: severity,
             }
 
diff --git a/tests/core/policies/test_unexpected_intent_policy.py b/tests/core/policies/test_unexpected_intent_policy.py
index 0818301caeca..a495700e45ad 100644
--- a/tests/core/policies/test_unexpected_intent_policy.py
+++ b/tests/core/policies/test_unexpected_intent_policy.py
@@ -1,3 +1,4 @@
+import json
 from pathlib import Path
 from typing import Optional, List, Dict, Text, Type
 import tensorflow as tf
@@ -489,9 +490,9 @@ def test_action_unlikely_intent_prediction(
         self,
         trained_policy: UnexpecTEDIntentPolicy,
         default_domain: Domain,
-        predicted_similarity,
-        threshold_value,
-        is_unlikely,
+        predicted_similarity: float,
+        threshold_value: float,
+        is_unlikely: bool,
         monkeypatch: MonkeyPatch,
         tmp_path: Path,
     ):
@@ -538,6 +539,8 @@ def test_action_unlikely_intent_prediction(
             # of the metadata is tested separately and
             # not as part of this test.
             assert prediction.action_metadata is not None
+            # Assert metadata is serializable
+            assert json.dumps(prediction.action_metadata)
 
     @pytest.mark.parametrize(
         "tracker_events, should_skip",