diff --git a/doc/source/whatsnew/v0.20.2.txt b/doc/source/whatsnew/v0.20.2.txt
index 6d6a148ed025f..b327e69348755 100644
--- a/doc/source/whatsnew/v0.20.2.txt
+++ b/doc/source/whatsnew/v0.20.2.txt
@@ -84,6 +84,7 @@ Reshaping
 - Bug in ``DataFrame.stack`` with unsorted levels in MultiIndex columns (:issue:`16323`)
 - Bug in ``pd.wide_to_long()`` where no error was raised when ``i`` was not a unique identifier (:issue:`16382`)
 - Bug in ``Series.isin(..)`` with a list of tuples (:issue:`16394`)
+- Bug in ``maybe_infer_to_datetimelike()`` in ``core.dtypes.cast`` where result was not reshaped in all cases (:issue:`16395`)
 
 
 Numeric
diff --git a/pandas/core/dtypes/cast.py b/pandas/core/dtypes/cast.py
index 19d3792f73de7..fd61813a57c98 100644
--- a/pandas/core/dtypes/cast.py
+++ b/pandas/core/dtypes/cast.py
@@ -837,7 +837,7 @@ def try_timedelta(v):
         try:
             return to_timedelta(v)._values.reshape(shape)
         except:
-            return v
+            return v.reshape(shape)
 
     inferred_type = lib.infer_datetimelike_array(_ensure_object(v))
 
diff --git a/pandas/tests/dtypes/test_cast.py b/pandas/tests/dtypes/test_cast.py
index e92724a5d9cd4..767e99d98cf29 100644
--- a/pandas/tests/dtypes/test_cast.py
+++ b/pandas/tests/dtypes/test_cast.py
@@ -9,7 +9,7 @@
 from datetime import datetime, timedelta, date
 import numpy as np
 
-from pandas import Timedelta, Timestamp, DatetimeIndex
+from pandas import Timedelta, Timestamp, DatetimeIndex, DataFrame, NaT
 
 from pandas.core.dtypes.cast import (
     maybe_downcast_to_dtype,
@@ -213,6 +213,17 @@ def test_maybe_convert_scalar(self):
         result = maybe_convert_scalar(Timedelta('1 day 1 min'))
         assert result == Timedelta('1 day 1 min').value
 
+    def test_maybe_infer_to_datetimelike(self):
+        # GH16362
+        # pandas=0.20.1 raises IndexError: tuple index out of range
+        result = DataFrame(np.array([[NaT, 'a', 'b', 0],
+                                     [NaT, 'b', 'c', 1]]))
+        assert result.size == 8
+        # this construction was fine
+        result = DataFrame(np.array([[NaT, 'a', 0],
+                                     [NaT, 'b', 1]]))
+        assert result.size == 6
+
 
 class TestConvert(object):