diff --git a/crates/polars-core/src/frame/column/mod.rs b/crates/polars-core/src/frame/column/mod.rs
index 585350c7627c..6e2a36be8b75 100644
--- a/crates/polars-core/src/frame/column/mod.rs
+++ b/crates/polars-core/src/frame/column/mod.rs
@@ -605,7 +605,7 @@ impl Column {
             Column::Partitioned(s) => series_agg(s.as_materialized_series(), groups).into_column(),
             Column::Scalar(s) => {
                 if s.is_empty() {
-                    return self.clone();
+                    return series_agg(s.as_materialized_series(), groups).into_column();
                 }
 
                 // We utilize the aggregation on Series to see:
diff --git a/py-polars/tests/unit/test_schema.py b/py-polars/tests/unit/test_schema.py
index 44211da41858..7ebae8e22f01 100644
--- a/py-polars/tests/unit/test_schema.py
+++ b/py-polars/tests/unit/test_schema.py
@@ -336,3 +336,12 @@ def test_raise_subnodes_18787() -> None:
                 pl.first().struct.field("a", "b").filter(pl.col("foo") == 1)
             )
         )
+
+
+def test_scalar_agg_schema_20044() -> None:
+    assert (
+        pl.DataFrame(None, schema={"a": pl.Int64, "b": pl.String, "c": pl.String})
+        .with_columns(d=pl.col("a").max())
+        .group_by("c")
+        .agg(pl.col("d").mean())
+    ).schema == pl.Schema([("c", pl.String), ("d", pl.Float64)])