diff --git a/pandas/core/arrays/arrow/array.py b/pandas/core/arrays/arrow/array.py
index ddb2f01898ec7..7edd2fa585156 100644
--- a/pandas/core/arrays/arrow/array.py
+++ b/pandas/core/arrays/arrow/array.py
@@ -1032,6 +1032,9 @@ def _reduce(self, name: str, *, skipna: bool = True, **kwargs):
             not_eq = pc.not_equal(data_to_cmp, 0)
             data_to_reduce = not_eq
 
+        elif name in ["min", "max", "sum"] and pa.types.is_duration(pa_type):
+            data_to_reduce = self._data.cast(pa.int64())
+
         if name == "sem":
 
             def pyarrow_meth(data, skip_nulls, **kwargs):
@@ -1066,6 +1069,9 @@ def pyarrow_meth(data, skip_nulls, **kwargs):
             raise TypeError(msg) from err
         if pc.is_null(result).as_py():
             return self.dtype.na_value
+
+        if name in ["min", "max", "sum"] and pa.types.is_duration(pa_type):
+            result = result.cast(pa_type)
         return result.as_py()
 
     def __setitem__(self, key, value) -> None:
diff --git a/pandas/tests/extension/base/groupby.py b/pandas/tests/extension/base/groupby.py
index 3a9dbe9dfb384..200a494997116 100644
--- a/pandas/tests/extension/base/groupby.py
+++ b/pandas/tests/extension/base/groupby.py
@@ -137,6 +137,7 @@ def test_in_numeric_groupby(self, data_for_grouping):
             or is_string_dtype(dtype)
             or is_period_dtype(dtype)
             or is_object_dtype(dtype)
+            or dtype.kind == "m"  # in particular duration[*][pyarrow]
         ):
             expected = pd.Index(["B", "C"])
             result = df.groupby("A").sum().columns
diff --git a/pandas/tests/extension/test_arrow.py b/pandas/tests/extension/test_arrow.py
index af3952a532113..71fc7e17bf808 100644
--- a/pandas/tests/extension/test_arrow.py
+++ b/pandas/tests/extension/test_arrow.py
@@ -539,10 +539,6 @@ def test_reduce_series(self, data, all_numeric_reductions, skipna, request):
             "sem",
         ] and pa.types.is_temporal(pa_dtype):
             request.node.add_marker(xfail_mark)
-        elif all_numeric_reductions in ["sum", "min", "max"] and pa.types.is_duration(
-            pa_dtype
-        ):
-            request.node.add_marker(xfail_mark)
         elif pa.types.is_boolean(pa_dtype) and all_numeric_reductions in {
             "sem",
             "std",