diff --git a/src/plugins/intel_cpu/src/nodes/conv.cpp b/src/plugins/intel_cpu/src/nodes/conv.cpp
index 5fb5e6ddc6813e..e22a36af852a14 100644
--- a/src/plugins/intel_cpu/src/nodes/conv.cpp
+++ b/src/plugins/intel_cpu/src/nodes/conv.cpp
@@ -531,8 +531,9 @@ void Convolution::getSupportedDescriptors() {
         auto dt = memory::data_type::f32;
 
         // supported lower precisions: bf16, f16
-        if (one_of(originalDT, memory::data_type::bf16, memory::data_type::f16))
+        if (one_of(originalDT, memory::data_type::bf16, memory::data_type::f16) && hasHardwareSupport(originalPrec)) {
             dt = originalDT;
+        }
 
         // fallback to f32 on special case for performance reasons
         if (isDepthWise() && ndims == 5)
diff --git a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
index 3a50f9260d83be..63a946cfa69955 100644
--- a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
+++ b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
@@ -356,7 +356,8 @@ void Transformations::PreLpt(const std::vector<ov::element::Type>& defaultPrecis
         // @todo should we always convert to f32 regardless of hardware support, as it is done for f16?
         if (!hasHardwareSupport(ov::element::bf16))
             map.insert({ov::element::bf16, ov::element::f32});
-        if (!one_of(inferencePrecision, element::f16, element::undefined)) {
+        // TODO: Remove 'hasHardwareSupport' when all nodes are able to handle f16 properly.
+        if (!one_of(inferencePrecision, element::f16, element::undefined) || !hasHardwareSupport(element::f16)) {
             map.insert({ov::element::f16, ov::element::f32});
         }
         return map;
diff --git a/src/plugins/intel_cpu/tests/functional/custom/subgraph_tests/src/classes/undefined_et.cpp b/src/plugins/intel_cpu/tests/functional/custom/subgraph_tests/src/classes/undefined_et.cpp
index 40b3277c8708b3..f232a2d641eb55 100644
--- a/src/plugins/intel_cpu/tests/functional/custom/subgraph_tests/src/classes/undefined_et.cpp
+++ b/src/plugins/intel_cpu/tests/functional/custom/subgraph_tests/src/classes/undefined_et.cpp
@@ -26,6 +26,7 @@
 //       -----------
 
 #include "custom/subgraph_tests/include/undefined_et.hpp"
+#include "utils/precision_support.h"
 
 namespace ov {
 namespace test {
@@ -74,6 +75,12 @@ void UndefinedEtSubgraphTest::SetUp() {
     auto logical_not = std::make_shared<op::v1::LogicalNot>(cvt_f32);
 
     function = std::make_shared<ov::Model>(OutputVector{logical_not->output(0)}, ParameterVector{param_0, param_1, param_2}, "UndefinedET");
+
+    // TODO: Need to remove when the hardware checking for f16 will be eliminated in the Transformations pipeline.
+    if (m_data_et == element::f16 && !ov::intel_cpu::hasHardwareSupport(m_data_et)) {
+        abs_threshold = 1.f;
+        rel_threshold = 0.1f;
+    }
 }
 
 template<typename TD, typename TS>
@@ -146,6 +153,10 @@ TEST_P(UndefinedEtSubgraphTest, CompareWithRefs) {
 
     size_t rnd_unfm_counter = 0lu;
     size_t logical_not_counter = 0lu;
+    auto expected_dt = m_data_et;
+    if (!ov::intel_cpu::hasHardwareSupport(expected_dt)) {
+        expected_dt = element::f32;
+    }
     for (const auto& node : compiledModel.get_runtime_model()->get_ops()) {
         auto rt_info = node->get_rt_info();
         auto it = rt_info.find(exec_model_info::LAYER_TYPE);
@@ -153,7 +164,7 @@ TEST_P(UndefinedEtSubgraphTest, CompareWithRefs) {
         auto op_name = it->second.as<std::string>();
 
         if (op_name == "RandomUniform") {
-            ASSERT_EQ(node->get_output_element_type(0), m_data_et);
+            ASSERT_EQ(node->get_output_element_type(0), expected_dt);
             rnd_unfm_counter++;
         }
         if (op_name == "Eltwise") {