diff --git a/src/scanpy/plotting/_anndata.py b/src/scanpy/plotting/_anndata.py
index 30156bf482..97d82ab7c1 100755
--- a/src/scanpy/plotting/_anndata.py
+++ b/src/scanpy/plotting/_anndata.py
@@ -36,6 +36,7 @@
 from ._utils import (
     _deprecated_scale,
     _dk,
+    _obs_vector_compat,
     check_colornorm,
     scatter_base,
     scatter_group,
@@ -55,13 +56,7 @@
     from seaborn.matrix import ClusterGrid
 
     from .._utils import Empty
-    from ._utils import (
-        ColorLike,
-        DensityNorm,
-        _FontSize,
-        _FontWeight,
-        _LegendLoc,
-    )
+    from ._utils import ColorLike, DensityNorm, _FontSize, _FontWeight, _LegendLoc
 
 # TODO: is that all?
 type _Basis = Literal["pca", "tsne", "umap", "diffmap", "draw_graph_fr"]
@@ -324,7 +319,7 @@ def _scatter_obs(  # noqa: PLR0912, PLR0913, PLR0915
             # ignore the '0th' diffusion component
             if basis == "diffmap":
                 components += 1
-            xy = adata.obsm["X_" + basis][:, components]
+            xy = adata.obsm[f"X_{basis}"][:, components]
             # correct the component vector for use in labeling etc.
             if basis == "diffmap":
                 components -= 1
@@ -332,19 +327,10 @@ def _scatter_obs(  # noqa: PLR0912, PLR0913, PLR0915
             msg = f"compute coordinates using visualization tool {basis} first"
             raise KeyError(msg) from None
     elif x is not None and y is not None:
-        if use_raw:
-            if x in adata.obs.columns:
-                x_arr = adata.obs_vector(x)
-            else:
-                x_arr = adata.raw.obs_vector(x)
-            if y in adata.obs.columns:
-                y_arr = adata.obs_vector(y)
-            else:
-                y_arr = adata.raw.obs_vector(y)
-        else:
-            x_arr = adata.obs_vector(x, layer=layers[0])
-            y_arr = adata.obs_vector(y, layer=layers[1])
-
+        x_arr, y_arr = (
+            _obs_vector_compat(adata, k, use_raw=use_raw, layer=layer)
+            for layer, k in zip(layers, [x, y], strict=False)
+        )
         xy = np.c_[x_arr, y_arr]
     else:
         msg = "Either provide a `basis` or `x` and `y`."
@@ -399,10 +385,10 @@ def _scatter_obs(  # noqa: PLR0912, PLR0913, PLR0915
             else:
                 c = adata.obs[key].to_numpy()
         # coloring according to gene expression
-        elif use_raw and adata.raw is not None and key in adata.raw.var_names:
-            c = adata.raw.obs_vector(key)
-        elif key in adata.var_names:
-            c = adata.obs_vector(key, layer=layers[2])
+        elif (use_raw and adata.raw is not None and key in adata.raw.var_names) or (
+            key in adata.var_names
+        ):
+            c = _obs_vector_compat(adata, key, use_raw=use_raw, layer=layers[2])
         elif is_color_like(key):  # a flat color
             c = key
             colorbar = False
diff --git a/src/scanpy/plotting/_tools/scatterplots.py b/src/scanpy/plotting/_tools/scatterplots.py
index 5ac7c1cda3..f4044a9fa0 100644
--- a/src/scanpy/plotting/_tools/scatterplots.py
+++ b/src/scanpy/plotting/_tools/scatterplots.py
@@ -31,7 +31,7 @@
     doc_scatter_spatial,
     doc_show_save_ax,
 )
-from .._utils import check_colornorm, check_projection, circles
+from .._utils import _obs_vector_compat, check_colornorm, check_projection, circles
 
 if TYPE_CHECKING:
     from collections.abc import Callable, Collection, Mapping
@@ -266,6 +266,8 @@ def embedding(  # noqa: PLR0912, PLR0913, PLR0915
     # ]
     for count, (value_to_plot, dims) in enumerate(zip(color, dimensions, strict=True)):
         kwargs_scatter = kwargs.copy()  # is potentially mutated for each plot
+        # TODO: It might be worth not returning `NumpyExtensionArray` objects out of the dataframes via accessors because we have a lot of np.ndarray checks.
+        # Setting np.array here prevents the  `NumpyExtensionArray` from propagating.
         color_source_vector = _get_color_source_vector(
             adata,
             value_to_plot,
@@ -275,6 +277,8 @@ def embedding(  # noqa: PLR0912, PLR0913, PLR0915
             gene_symbols=gene_symbols,
             groups=groups,
         )
+        if isinstance(color_source_vector, pd.arrays.NumpyExtensionArray):
+            color_source_vector = color_source_vector.to_numpy()
         color_vector, color_type = _color_vector(
             adata,
             value_to_plot,
@@ -1221,10 +1225,7 @@ def _get_color_source_vector(
         # We should probably just make an index for this, and share it over runs
         # TODO: Throw helpful error if this doesn't work
         value_to_plot = adata.var.index[adata.var[gene_symbols] == value_to_plot][0]
-    if use_raw and value_to_plot not in adata.obs.columns:
-        values = adata.raw.obs_vector(value_to_plot)
-    else:
-        values = adata.obs_vector(value_to_plot, layer=layer)
+    values = _obs_vector_compat(adata, value_to_plot, use_raw=use_raw, layer=layer)
     if mask_obs is not None:
         values = values.copy()
         values[~mask_obs] = np.nan
diff --git a/src/scanpy/plotting/_utils.py b/src/scanpy/plotting/_utils.py
index d648e3569b..91c79ccd3b 100644
--- a/src/scanpy/plotting/_utils.py
+++ b/src/scanpy/plotting/_utils.py
@@ -29,6 +29,7 @@
     from matplotlib.figure import Figure
     from matplotlib.typing import MarkerType
     from numpy.typing import ArrayLike
+    from pandas.api.extensions import ExtensionArray
     from PIL.Image import Image
 
     from .._utils import Empty
@@ -44,6 +45,7 @@
     "_create_white_to_color_gradient",
     "_deprecated_scale",
     "_dk",
+    "_obs_vector_compat",
     "add_colors_for_categorical_sample_annotation",
     "check_colornorm",
     "check_projection",
@@ -1167,3 +1169,23 @@ def _create_white_to_color_gradient(
     return ListedColormap(
         clipped_rgb, name=color if isinstance(color, str) else hex_color
     )
+
+
+def _obs_vector_compat(
+    adata: AnnData, k: str, *, use_raw: bool, layer: str | None
+) -> np.ndarray | ExtensionArray:
+    try:
+        from anndata.acc import A
+    except ImportError:
+        return (
+            adata.raw.obs_vector(k)
+            if use_raw and k not in adata.obs.columns
+            else adata.obs_vector(k, layer=layer)
+        )
+
+    if k in adata.obs.columns:
+        return adata[A.obs[k]]
+    elif not use_raw:
+        return adata[A.layers[layer][:, k]]
+    else:
+        return adata.raw[A.X[:, k]]
diff --git a/tests/test_get.py b/tests/test_get.py
index 20a36dee24..0f8d28c22f 100644
--- a/tests/test_get.py
+++ b/tests/test_get.py
@@ -204,12 +204,15 @@ def test_column_content():
     adata = pbmc68k_reduced()
 
     # test that columns content is correct for obs_df
-    query = ["CST3", "NKG7", "GNLY", "louvain", "n_counts", "n_genes"]
+    cols = ["louvain", "n_counts", "n_genes"]
+    query = [*cols, "CST3", "NKG7", "GNLY"]
     df = sc.get.obs_df(adata, query)
     for col in query:
         assert col in df
         np.testing.assert_array_equal(query, df.columns)
-        np.testing.assert_array_equal(df[col].values, adata.obs_vector(col))
+        np.testing.assert_array_equal(
+            df[col].values, adata.obs[col] if col in cols else adata[:, col].X.ravel()
+        )
 
     # test that columns content is correct for var_df
     cell_ids = list(adata.obs.sample(5).index)
@@ -217,7 +220,10 @@ def test_column_content():
     df = sc.get.var_df(adata, query)
     np.testing.assert_array_equal(query, df.columns)
     for col in query:
-        np.testing.assert_array_equal(df[col].values, adata.var_vector(col))
+        np.testing.assert_array_equal(
+            df[col].values,
+            adata[col, :].X.ravel() if col in cell_ids else adata.var[col],
+        )
 
 
 def test_var_df(adata: AnnData):
diff --git a/tests/test_metrics.py b/tests/test_metrics.py
index 40c6cbfa58..fb44e7cb9f 100644
--- a/tests/test_metrics.py
+++ b/tests/test_metrics.py
@@ -68,7 +68,9 @@ def test_consistency(metric) -> None:
     )
 
     all_genes = metric(pbmc, layer="raw")
-    first_gene = metric(pbmc, vals=pbmc.obs_vector(pbmc.var_names[0], layer="raw"))
+    first_gene = metric(
+        pbmc, vals=pbmc[:, pbmc.var_names[0]].layers["raw"].toarray().ravel()
+    )
 
     np.testing.assert_allclose(all_genes[0], first_gene, rtol=1e-9)
 
diff --git a/tests/test_plotting_embedded/conftest.py b/tests/test_plotting_embedded/conftest.py
index 42be824899..7b30bad49a 100644
--- a/tests/test_plotting_embedded/conftest.py
+++ b/tests/test_plotting_embedded/conftest.py
@@ -58,7 +58,10 @@ def adata():
     adata.obs["label_missing"] = adata.obs["label"].copy()
     adata.obs.loc[::2, "label_missing"] = np.nan
 
-    adata.obs["1_missing"] = adata.obs_vector("1")
+    # TODO: If we don't `copy`, something about this being an ArrayView means that all values get set to nan?
+    # https://github.com/scverse/anndata/issues/2348
+    adata.obs["1_missing"] = adata[:, "1"].X.flatten().copy()
+
     adata.obs.loc[
         adata.obsm["spatial"][:, 0] < adata.obsm["spatial"][:, 0].mean(), "1_missing"
     ] = np.nan