diff --git a/src/azure-cli-core/azure/cli/core/util.py b/src/azure-cli-core/azure/cli/core/util.py
index e4c319ad55f..1a691d417a5 100644
--- a/src/azure-cli-core/azure/cli/core/util.py
+++ b/src/azure-cli-core/azure/cli/core/util.py
@@ -633,6 +633,7 @@ def todict(obj, post_processor=None):
     """
     from datetime import date, time, datetime, timedelta
     from enum import Enum
+    from azure.core.serialization import attribute_list
     if isinstance(obj, dict):
         result = {k: todict(v, post_processor) for (k, v) in obj.items()}
         return post_processor(obj, result) if post_processor else result
@@ -646,9 +647,12 @@ def todict(obj, post_processor=None):
         return str(obj)
     # This is the only difference with knack.util.todict because for typespec generated SDKs
     # The base model stores data in obj.__dict__['_data'] instead of in obj.__dict__
-    # We need to call obj.as_dict() to extract data for this kind of model
-    if hasattr(obj, 'as_dict') and not hasattr(obj, '_attribute_map'):
-        result = {to_camel_case(k): todict(v, post_processor) for k, v in obj.as_dict().items()}
+    # The way to detect if it's a typespec generated model is to check the private `_is_model` attribute
+    # azure-core provided new function `attribute_list` to list all attribute names
+    # so that we don't need to use raw __dict__ directly
+    if getattr(obj, "_is_model", False):
+        result = {to_camel_case(attr): todict(getattr(obj, attr), post_processor)
+                  for attr in attribute_list(obj) if hasattr(obj, attr)}
         return post_processor(obj, result) if post_processor else result
     if hasattr(obj, '_asdict'):
         return todict(obj._asdict(), post_processor)