diff --git a/scripts/load_npi_companion_data.py b/scripts/load_npi_companion_data.py
index ba2d766..3916863 100644
--- a/scripts/load_npi_companion_data.py
+++ b/scripts/load_npi_companion_data.py
@@ -27,6 +27,11 @@ from psycopg2.extras import execute_values
 
 DATABASE_URL = os.environ.get("DATABASE_URL", "postgresql://pw:pw@localhost:5432/performancewest")
 
+# CMS/OIG exports are not always clean UTF-8 (stray latin-1 bytes like 0xa0).
+# Decode leniently so a few bad bytes don't abort a multi-hundred-MB load.
+def open_csv(path):
+    return open(path, newline="", encoding="utf-8-sig", errors="replace")
+
 
 def parse_date(s):
     if not s:
@@ -49,7 +54,7 @@ def clean_npi(s):
 
 def load_revalidation(conn, path):
     rows = []
-    with open(path, newline="", encoding="utf-8-sig") as f:
+    with open_csv(path) as f:
         for r in csv.DictReader(f):
             npi = (r.get("National Provider Identifier") or "").strip()
             if not (npi.isdigit() and len(npi) == 10):
@@ -84,7 +89,7 @@ def load_revalidation(conn, path):
 
 def load_exclusions(conn, path):
     rows = []
-    with open(path, newline="", encoding="utf-8-sig") as f:
+    with open_csv(path) as f:
         for r in csv.DictReader(f):
             rows.append((
                 clean_npi(r.get("NPI")),
@@ -114,7 +119,7 @@ def load_exclusions(conn, path):
 
 def load_optout(conn, path):
     rows = []
-    with open(path, newline="", encoding="utf-8-sig") as f:
+    with open_csv(path) as f:
         for r in csv.DictReader(f):
             npi = (r.get("npi") or r.get("NPI") or "").strip()
             if not (npi.isdigit() and len(npi) == 10):