diff --git a/hlink/linking/core/classifier.py b/hlink/linking/core/classifier.py
index bf9f898..65295fe 100644
--- a/hlink/linking/core/classifier.py
+++ b/hlink/linking/core/classifier.py
@@ -117,6 +117,10 @@ def choose_classifier(model_type, params, dep_var):
             **params_without_threshold,
             features_col=features_vector,
             label_col=dep_var,
+            probability_col="probability_array",
+        )
+        post_transformer = SQLTransformer(
+            statement="SELECT *, parseProbVector(probability_array, 1) as probability FROM __THIS__"
         )
     else:
         raise ValueError(
diff --git a/hlink/tests/integration_score_with_trained_models_test.py b/hlink/tests/integration_score_with_trained_models_test.py
index 993a497..c04819c 100644
--- a/hlink/tests/integration_score_with_trained_models_test.py
+++ b/hlink/tests/integration_score_with_trained_models_test.py
@@ -3,6 +3,8 @@
 # in this project's top-level directory, and also on-line at:
 #   https://github.com/ipums/hlink
 
+from hlink.tests.markers import requires_xgboost
+
 
 def test_apply_chosen_model_RF(
     spark,
@@ -859,6 +861,100 @@ def test_step_3_apply_chosen_model_boosted_trees(
     )
 
 
+@requires_xgboost
+def test_apply_chosen_model_xgboost(
+    spark,
+    training,
+    matching,
+    training_conf,
+    datasource_training_input,
+    potential_matches_path,
+    state_dist_path,
+    spark_test_tmp_dir_path,
+):
+    training_data_path, prepped_df_a_path, prepped_df_b_path = datasource_training_input
+    training_conf["comparison_features"] = [
+        {
+            "alias": "regionf",
+            "column_name": "region",
+            "comparison_type": "fetch_a",
+            "categorical": True,
+        },
+        {
+            "alias": "namelast_jw",
+            "column_name": "namelast",
+            "comparison_type": "jaro_winkler",
+        },
+        {
+            "alias": "state_distance",
+            "key_count": 1,
+            "column_name": "bpl",
+            "comparison_type": "geo_distance",
+            "loc_a": "statecode1",
+            "loc_b": "statecode2",
+            "distance_col": "dist",
+            "table_name": "state_distances_lookup",
+            "distances_file": state_dist_path,
+        },
+    ]
+
+    training_conf["training"]["dataset"] = training_data_path
+    training_conf["training"]["dependent_var"] = "match"
+    training_conf["training"]["independent_vars"] = [
+        "namelast_jw",
+        "regionf",
+        "state_distance",
+    ]
+    training_conf["training"]["chosen_model"] = {
+        "type": "xgboost",
+        "max_depth": 5,
+        "eta": 0.5,
+        "threshold": 0.5,
+        "threshold_ratio": 1.3,
+    }
+    training_conf["training"]["score_with_model"] = True
+    training_conf["spark_tmp_dir"] = spark_test_tmp_dir_path
+    training_conf["drop_data_from_scored_matches"] = True
+
+    prepped_df_a = spark.read.csv(prepped_df_a_path, header=True, inferSchema=True)
+    prepped_df_b = spark.read.csv(prepped_df_b_path, header=True, inferSchema=True)
+    potential_matches = spark.read.csv(
+        potential_matches_path, header=True, inferSchema=True
+    )
+    prepped_df_a.write.mode("overwrite").saveAsTable("prepped_df_a")
+    prepped_df_b.write.mode("overwrite").saveAsTable("prepped_df_b")
+    potential_matches.write.mode("overwrite").saveAsTable("potential_matches")
+
+    training.run_all_steps()
+    matching.run_step(2)
+
+    potential_matches_df = spark.table("scored_potential_matches").toPandas()
+
+    # Check one case that we expect to be a match and one case that we expect not
+    # to be a match.
+    should_be_match = potential_matches_df.query(
+        "id_a == '0202928A-AC3E-48BB-8568-3372067F35C7'"
+    )
+    assert (
+        should_be_match.shape[0] == 1
+    ), "expected exactly one potential match for 0202928A"
+    assert should_be_match["probability"].iloc[0] >= 0.5
+    assert should_be_match["prediction"].iloc[0] == 1
+
+    # In the real world, this would probably be a match, depending on how the
+    # additional features looked. But we've included so few training features
+    # for our test model that small differences in names can really hurt a
+    # potential match's chances of being classified as a match.
+    should_not_be_match = potential_matches_df.query(
+        "id_b == '033FD0FA-C523-42B5-976A-751E830F7021'"
+    )
+    assert (
+        should_not_be_match.shape[0] == 1
+    ), "expected exactly one potential match for 033FD0FA"
+    assert should_not_be_match["probability"].iloc[0] <= 0.5
+    assert should_not_be_match["prediction"].iloc[0] == 0
+
+
 def test_step_3_apply_chosen_model_RF_threshold(
     spark,
     training_conf,