diff --git a/CHANGELOG.md b/CHANGELOG.md
index b28905a..1ce0b3f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,8 +8,9 @@ and [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 ## [0.1.2] — 2026-04-27
 
 Security patch covering all High-severity findings from the 0.1.1
-internal security review. Recommended upgrade for anyone running
-0.1.x.
+internal security review, plus an Excel-redaction quality fix that
+restores parity between the CLI/SDK/daemon path and the noirdoc-cloud
+proxy. Recommended upgrade for anyone running 0.1.x.
 
 ### Security
 - **PDF metadata leak.** PII embedded in a PDF's `/Info` dictionary
@@ -59,6 +60,21 @@ internal security review. Recommended upgrade for anyone running
   command now exits with an error and points at `noirdoc ns
   summary` unless `--unsafe` is passed.
 
+### Fixed
+- **XLSX redaction quality regression.** `Redactor.redact_file` (used
+  by `noirdoc redact`, the Python SDK, and the daemon) flattened every
+  cell across every sheet into a single ` | `-joined string before
+  detection, then did substring `cell.value.replace()` on
+  reconstruction. Cell context was destroyed and many entities — short
+  surnames, locations, numerically-typed cells — were silently missed.
+  XLSX inputs now route through
+  `noirdoc.file_analysis.xlsx_inference.pseudonymize_xlsx_smart`, which
+  classifies columns by header keyword, samples the first rows for
+  unclassified columns, and writes per-cell `<<TYPE_N>>` pseudonyms via
+  `mapper.get_or_create()`. The reveal path was already cell-aware and
+  round-trips correctly. This was the same path the noirdoc-cloud proxy
+  has always used; the SDK simply wasn't wired up to it.
+
 ## [0.1.1] — 2026-04-27
 
 ### Added
diff --git a/src/noirdoc/sdk.py b/src/noirdoc/sdk.py
index 9adf6cd..a976cab 100644
--- a/src/noirdoc/sdk.py
+++ b/src/noirdoc/sdk.py
@@ -28,6 +28,8 @@
 Policy = Literal["pseudonymize", "extract_only"]
 DetectorChoice = Literal["presidio", "gliner", "ensemble"]
 
+_XLSX_MIME = "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
+
 
 class RedactionResult:
     """The result of redacting a single file."""
@@ -234,6 +236,30 @@ async def _redact_file_async(self, path: Path, language: str) -> RedactionResult
         if format_for_mime(mime) is None:
             raise ValueError(f"Unsupported MIME type for {path.name}: {mime}")
 
+        # XLSX uses a column-aware pipeline: header keyword classification + per-column NLP
+        # sampling + cell-level pseudonymization. The generic flat-text path destroys cell
+        # context and misses many entities — see xlsx_inference.pseudonymize_xlsx_smart.
+        if mime == _XLSX_MIME:
+            from noirdoc.file_analysis.xlsx_inference import pseudonymize_xlsx_smart
+
+            detector = await self._ensure_detector()
+            xr = await pseudonymize_xlsx_smart(
+                content,
+                detector,
+                self._mapper,
+                language=language,
+                pseudonymize=True,
+            )
+            self._persist()
+            return RedactionResult(
+                input_path=path,
+                output_bytes=xr.new_bytes if xr.new_bytes is not None else content,
+                entity_count=xr.entity_count,
+                entity_types=dict(xr.entity_types),
+                mime_type=mime,
+                reconstructed=True,
+            )
+
         block = FileBlock(
             source_path="sdk",
             source_type="file",
diff --git a/tests/test_sdk_xlsx.py b/tests/test_sdk_xlsx.py
new file mode 100644
index 0000000..630ae01
--- /dev/null
+++ b/tests/test_sdk_xlsx.py
@@ -0,0 +1,74 @@
+"""Regression test for XLSX redaction via the SDK.
+
+Confirms that ``Redactor.redact_file`` routes ``.xlsx`` through the
+column-aware ``pseudonymize_xlsx_smart`` pipeline (header keyword
+classification + per-cell pseudonyms) instead of the flat-text fallback,
+which silently dropped most entities on real spreadsheets.
+"""
+
+from __future__ import annotations
+
+import io
+
+import pytest
+
+pytestmark = pytest.mark.slow
+
+
+def _build_workbook(path):
+    from openpyxl import Workbook
+
+    wb = Workbook()
+    ws = wb.active
+    ws.title = "Sheet1"
+    ws.append(["Name", "Email", "Notes"])
+    ws.append(["Anna Müller", "anna@example.com", "leave alone"])
+    ws.append(["Ben Schulz", "ben@example.com", "also untouched"])
+    wb.save(path)
+
+
+def test_redact_file_xlsx_uses_smart_pipeline(tmp_path):
+    from openpyxl import load_workbook
+
+    from noirdoc.sdk import Redactor
+
+    src = tmp_path / "in.xlsx"
+    dst = tmp_path / "out.xlsx"
+    _build_workbook(src)
+
+    r = Redactor(detector="presidio", language="de")
+    result = r.redact_file(src, output=dst)
+
+    assert result.reconstructed is True
+    assert result.mime_type.endswith("spreadsheetml.sheet")
+    assert result.entity_count >= 4
+
+    out = load_workbook(dst)["Sheet1"]
+    assert out.cell(2, 1).value.startswith("<<PERSON")
+    assert out.cell(3, 1).value.startswith("<<PERSON")
+    assert out.cell(2, 2).value.startswith("<<EMAIL")
+    assert out.cell(3, 2).value.startswith("<<EMAIL")
+    # Non-classified column passes through untouched.
+    assert out.cell(2, 3).value == "leave alone"
+    assert out.cell(3, 3).value == "also untouched"
+
+
+def test_reveal_file_xlsx_roundtrips(tmp_path):
+    from openpyxl import load_workbook
+
+    from noirdoc.sdk import Redactor
+
+    src = tmp_path / "in.xlsx"
+    dst = tmp_path / "out.xlsx"
+    _build_workbook(src)
+
+    r = Redactor(detector="presidio", language="de")
+    r.redact_file(src, output=dst)
+    revealed = r.reveal_file(dst)
+
+    assert revealed is not None
+    out = load_workbook(io.BytesIO(revealed))["Sheet1"]
+    assert out.cell(2, 1).value == "Anna Müller"
+    assert out.cell(2, 2).value == "anna@example.com"
+    assert out.cell(3, 1).value == "Ben Schulz"
+    assert out.cell(3, 2).value == "ben@example.com"