diff --git a/pandas/_libs/tslib.pyx b/pandas/_libs/tslib.pyx
index da9d65bcfc095..ad043d1695eac 100644
--- a/pandas/_libs/tslib.pyx
+++ b/pandas/_libs/tslib.pyx
@@ -693,6 +693,8 @@ def array_to_datetime_with_tz(
         object item
         int64_t ival
         _TSObject tsobj
+        bint infer_reso = creso == NPY_DATETIMEUNIT.NPY_FR_GENERIC
+        DatetimeParseState state = DatetimeParseState(creso)
 
     for i in range(n):
         # Analogous to `item = values[i]`
@@ -707,6 +709,9 @@ def array_to_datetime_with_tz(
                 item, tz=tz, unit="ns", dayfirst=dayfirst, yearfirst=yearfirst, nanos=0
             )
             if tsobj.value != NPY_NAT:
+                state.update_creso(tsobj.creso)
+                if infer_reso:
+                    creso = state.creso
                 tsobj.ensure_reso(creso, item, round_ok=True)
             ival = tsobj.value
 
@@ -715,4 +720,20 @@ def array_to_datetime_with_tz(
 
         cnp.PyArray_MultiIter_NEXT(mi)
 
+    if infer_reso:
+        if state.creso_ever_changed:
+            # We encountered mismatched resolutions, need to re-parse with
+            #  the correct one.
+            return array_to_datetime_with_tz(values, tz=tz, creso=creso)
+
+        # Otherwise we can use the single reso that we encountered and avoid
+        #  a second pass.
+        abbrev = npy_unit_to_abbrev(creso)
+        result = result.view(f"M8[{abbrev}]")
+    elif creso == NPY_DATETIMEUNIT.NPY_FR_GENERIC:
+        # We didn't find any non-NaT to infer from, default to "ns"
+        result = result.view("M8[ns]")
+    else:
+        abbrev = npy_unit_to_abbrev(creso)
+        result = result.view(f"M8[{abbrev}]")
     return result
diff --git a/pandas/core/arrays/datetimes.py b/pandas/core/arrays/datetimes.py
index 86402fe05e08a..8305da745ab00 100644
--- a/pandas/core/arrays/datetimes.py
+++ b/pandas/core/arrays/datetimes.py
@@ -2241,14 +2241,14 @@ def _sequence_to_dt64(
             data = data.astype(np.int64)
         elif tz is not None and ambiguous == "raise":
             obj_data = np.asarray(data, dtype=object)
-            i8data = tslib.array_to_datetime_with_tz(
+            result = tslib.array_to_datetime_with_tz(
                 obj_data,
                 tz=tz,
                 dayfirst=dayfirst,
                 yearfirst=yearfirst,
                 creso=abbrev_to_npy_unit(out_unit),
             )
-            return i8data.view(out_dtype), tz, None
+            return result, tz, None
         else:
             # data comes back here as either i8 to denote UTC timestamps
             #  or M8[ns] to denote wall times
diff --git a/pandas/tests/tslibs/test_array_to_datetime.py b/pandas/tests/tslibs/test_array_to_datetime.py
index 829bb140e6e96..86560969d10ce 100644
--- a/pandas/tests/tslibs/test_array_to_datetime.py
+++ b/pandas/tests/tslibs/test_array_to_datetime.py
@@ -10,13 +10,43 @@
 import pytest
 
 from pandas._libs import (
+    NaT,
     iNaT,
     tslib,
 )
+from pandas._libs.tslibs.dtypes import NpyDatetimeUnit
 
 from pandas import Timestamp
 import pandas._testing as tm
 
+creso_infer = NpyDatetimeUnit.NPY_FR_GENERIC.value
+
+
+class TestArrayToDatetimeWithTZResolutionInference:
+    def test_array_to_datetime_with_tz_resolution(self):
+        tz = tzoffset("custom", 3600)
+        vals = np.array(["2016-01-01 02:03:04.567", NaT], dtype=object)
+        res = tslib.array_to_datetime_with_tz(vals, tz, False, False, creso_infer)
+        assert res.dtype == "M8[ms]"
+
+        vals2 = np.array([datetime(2016, 1, 1, 2, 3, 4), NaT], dtype=object)
+        res2 = tslib.array_to_datetime_with_tz(vals2, tz, False, False, creso_infer)
+        assert res2.dtype == "M8[us]"
+
+        vals3 = np.array([NaT, np.datetime64(12345, "s")], dtype=object)
+        res3 = tslib.array_to_datetime_with_tz(vals3, tz, False, False, creso_infer)
+        assert res3.dtype == "M8[s]"
+
+    def test_array_to_datetime_with_tz_resolution_all_nat(self):
+        tz = tzoffset("custom", 3600)
+        vals = np.array(["NaT"], dtype=object)
+        res = tslib.array_to_datetime_with_tz(vals, tz, False, False, creso_infer)
+        assert res.dtype == "M8[ns]"
+
+        vals2 = np.array([NaT, NaT], dtype=object)
+        res2 = tslib.array_to_datetime_with_tz(vals2, tz, False, False, creso_infer)
+        assert res2.dtype == "M8[ns]"
+
 
 @pytest.mark.parametrize(
     "data,expected",