diff --git a/doc/source/whatsnew/v1.3.0.rst b/doc/source/whatsnew/v1.3.0.rst
index 258e391b9220c..9b96c6b95301d 100644
--- a/doc/source/whatsnew/v1.3.0.rst
+++ b/doc/source/whatsnew/v1.3.0.rst
@@ -872,6 +872,7 @@ Indexing
 - Bug in :meth:`DataFrame.__setitem__` and :meth:`DataFrame.iloc.__setitem__` raising ``ValueError`` when trying to index with a row-slice and setting a list as values (:issue:`40440`)
 - Bug in :meth:`DataFrame.loc` not raising ``KeyError`` when key was not found in :class:`MultiIndex` when levels contain more values than used (:issue:`41170`)
 - Bug in :meth:`DataFrame.loc.__setitem__` when setting-with-expansion incorrectly raising when the index in the expanding axis contains duplicates (:issue:`40096`)
+- Bug in :meth:`DataFrame.loc.__getitem__` with :class:`MultiIndex` casting to float when at least one column is from has float dtype and we retrieve a scalar (:issue:`41369`)
 - Bug in :meth:`DataFrame.loc` incorrectly matching non-boolean index elements (:issue:`20432`)
 - Bug in :meth:`Series.__delitem__` with ``ExtensionDtype`` incorrectly casting to ``ndarray`` (:issue:`40386`)
 - Bug in :meth:`DataFrame.__setitem__` raising ``TypeError`` when using a str subclass as the column name with a :class:`DatetimeIndex` (:issue:`37366`)
diff --git a/pandas/core/indexing.py b/pandas/core/indexing.py
index 0a06dff790cbf..be5b89f08b5ca 100644
--- a/pandas/core/indexing.py
+++ b/pandas/core/indexing.py
@@ -886,26 +886,22 @@ def _getitem_nested_tuple(self, tup: tuple):
         # handle the multi-axis by taking sections and reducing
         # this is iterative
         obj = self.obj
-        axis = 0
-        for key in tup:
+        # GH#41369 Loop in reverse order ensures indexing along columns before rows
+        # which selects only necessary blocks which avoids dtype conversion if possible
+        axis = len(tup) - 1
+        for key in tup[::-1]:
 
             if com.is_null_slice(key):
-                axis += 1
+                axis -= 1
                 continue
 
-            current_ndim = obj.ndim
             obj = getattr(obj, self.name)._getitem_axis(key, axis=axis)
-            axis += 1
+            axis -= 1
 
             # if we have a scalar, we are done
             if is_scalar(obj) or not hasattr(obj, "ndim"):
                 break
 
-            # has the dim of the obj changed?
-            # GH 7199
-            if obj.ndim < current_ndim:
-                axis -= 1
-
         return obj
 
     def _convert_to_indexer(self, key, axis: int, is_setter: bool = False):
diff --git a/pandas/tests/indexing/multiindex/test_loc.py b/pandas/tests/indexing/multiindex/test_loc.py
index 558270ac86532..f9e2d1280b33a 100644
--- a/pandas/tests/indexing/multiindex/test_loc.py
+++ b/pandas/tests/indexing/multiindex/test_loc.py
@@ -831,3 +831,16 @@ def test_mi_add_cell_missing_row_non_unique():
         columns=MultiIndex.from_product([[1, 2], ["A", "B"]]),
     )
     tm.assert_frame_equal(result, expected)
+
+
+def test_loc_get_scalar_casting_to_float():
+    # GH#41369
+    df = DataFrame(
+        {"a": 1.0, "b": 2}, index=MultiIndex.from_arrays([[3], [4]], names=["c", "d"])
+    )
+    result = df.loc[(3, 4), "b"]
+    assert result == 2
+    assert isinstance(result, np.int64)
+    result = df.loc[[(3, 4)], "b"].iloc[0]
+    assert result == 2
+    assert isinstance(result, np.int64)