diff --git a/awx/main/tasks.py b/awx/main/tasks.py
index 8360ae82db..0a3fe95a44 100644
--- a/awx/main/tasks.py
+++ b/awx/main/tasks.py
@@ -148,38 +148,55 @@ def inform_cluster_of_shutdown(*args, **kwargs):
 @shared_task(bind=True, queue=settings.CELERY_DEFAULT_QUEUE)
 def apply_cluster_membership_policies(self):
     with advisory_lock('cluster_policy_lock', wait=True):
-        considered_instances = Instance.objects.all_non_isolated().order_by('id')
-        total_instances = considered_instances.count()
+        all_instances = list(Instance.objects.order_by('id'))
+        all_groups = list(InstanceGroup.objects.all())
+        iso_hostnames = set([])
+        for ig in all_groups:
+            if ig.controller_id is not None:
+                iso_hostnames.update(ig.policy_instance_list)
+
+        considered_instances = [inst for inst in all_instances if inst.hostname not in iso_hostnames]
+        total_instances = len(considered_instances)
         actual_groups = []
         actual_instances = []
         Group = namedtuple('Group', ['obj', 'instances'])
         Node = namedtuple('Instance', ['obj', 'groups'])
 
         # Process policy instance list first, these will represent manually managed memberships
-        for ig in InstanceGroup.objects.all():
-            logger.info(six.text_type("Applying cluster policy instance list to Group {}").format(ig.name))
-            ig.instances.clear()
+        instance_hostnames_map = {inst.hostname: inst for inst in all_instances}
+        for ig in all_groups:
             group_actual = Group(obj=ig, instances=[])
-            for i in ig.policy_instance_list:
-                inst = Instance.objects.filter(hostname=i)
-                if not inst.exists():
+            for hostname in ig.policy_instance_list:
+                if hostname not in instance_hostnames_map:
                     continue
-                inst = inst[0]
+                inst = instance_hostnames_map[hostname]
                 logger.info(six.text_type("Policy List, adding Instance {} to Group {}").format(inst.hostname, ig.name))
                 group_actual.instances.append(inst.id)
-                ig.instances.add(inst)
-            actual_groups.append(group_actual)
+                # NOTE: arguable behavior: policy-list-group is not added to
+                # instance's group count for consideration in minimum-policy rules
+
+            if ig.controller_id is None:
+                actual_groups.append(group_actual)
+            else:
+                # For isolated groups, _only_ apply the policy_instance_list
+                # do not add to in-memory list, so minimum rules not applied
+                logger.info('Committing instances {} to isolated group {}'.format(group_actual.instances, ig.name))
+                ig.instances.set(group_actual.instances)
 
         # Process Instance minimum policies next, since it represents a concrete lower bound to the
         # number of instances to make available to instance groups
         actual_instances = [Node(obj=i, groups=[]) for i in considered_instances if i.managed_by_policy]
-        logger.info("Total instances not directly associated: {}".format(total_instances))
+        logger.info("Total non-isolated instances:{} available for policy: {}".format(
+            total_instances, len(actual_instances)))
         for g in sorted(actual_groups, cmp=lambda x,y: len(x.instances) - len(y.instances)):
             for i in sorted(actual_instances, cmp=lambda x,y: len(x.groups) - len(y.groups)):
                 if len(g.instances) >= g.obj.policy_instance_minimum:
                     break
+                if i.obj.id in g.instances:
+                    # If the instance is already _in_ the group, it was
+                    # applied earlier via the policy list
+                    continue
                 logger.info(six.text_type("Policy minimum, adding Instance {} to Group {}").format(i.obj.hostname, g.obj.name))
-                g.obj.instances.add(i.obj)
                 g.instances.append(i.obj.id)
                 i.groups.append(g.obj.id)
 
@@ -188,15 +205,20 @@ def apply_cluster_membership_policies(self):
             for i in sorted(actual_instances, cmp=lambda x,y: len(x.groups) - len(y.groups)):
                 if i.obj.id in g.instances:
                     # If the instance is already _in_ the group, it was
-                    # probably applied earlier via a minimum policy
+                    # applied earlier via a minimum policy or policy list
                     continue
                 if 100 * float(len(g.instances)) / len(actual_instances) >= g.obj.policy_instance_percentage:
                     break
                 logger.info(six.text_type("Policy percentage, adding Instance {} to Group {}").format(i.obj.hostname, g.obj.name))
                 g.instances.append(i.obj.id)
-                g.obj.instances.add(i.obj)
                 i.groups.append(g.obj.id)
 
+        # On a differential basis, apply instances to non-isolated groups
+        with transaction.atomic():
+            for g in actual_groups:
+                logger.info('Committing instances {} to group {}'.format(g.instances, g.obj.name))
+                g.obj.instances.set(g.instances)
+
 
 @shared_task(exchange='tower_broadcast_all', bind=True)
 def handle_setting_changes(self, setting_keys):