diff --git a/CHANGELOG.md b/CHANGELOG.md
index ae3c75cc..71d5684b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,4 +1,10 @@
 ## [7.0.0] - Unreleased
+- Fix: Raise ArgumentError when using delay with FIFO queues
+  - FIFO queues do not support per-message DelaySeconds
+  - Previously caused confusing AWS errors with ActiveJob's `retry_on` (Rails 6.1+ defaults to `wait: 3.seconds`)
+  - Now raises clear ArgumentError with guidance to use `wait: 0`
+  - Fixes #924
+
 - Enhancement: Use fiber-local storage for logging context
   - Replaces thread-local storage with Fiber[] for proper isolation in async environments
   - Ensures logging context doesn't leak between fibers in the same thread
diff --git a/lib/active_job/queue_adapters/shoryuken_adapter.rb b/lib/active_job/queue_adapters/shoryuken_adapter.rb
index 3f5c673f..2c7cf25c 100644
--- a/lib/active_job/queue_adapters/shoryuken_adapter.rb
+++ b/lib/active_job/queue_adapters/shoryuken_adapter.rb
@@ -98,8 +98,25 @@ def enqueue(job, options = {}) # :nodoc:
       # @param job [ActiveJob::Base] the job to enqueue
       # @param timestamp [Float] Unix timestamp when the job should be processed
       # @return [Aws::SQS::Types::SendMessageResult] the send result
+      # @raise [ArgumentError] if delay is used with a FIFO queue
       def enqueue_at(job, timestamp) # :nodoc:
-        enqueue(job, delay_seconds: calculate_delay(timestamp))
+        delay = calculate_delay(timestamp)
+
+        # FIFO queues do not support per-message delays
+        # Check early to fail synchronously (before any async wrapping in subclasses)
+        # Note: negative delays (past timestamps) don't need handling here -
+        # SQS treats them as immediate delivery (delay_seconds=0)
+        # See https://github.com/ruby-shoryuken/shoryuken/issues/924
+        if delay.positive?
+          queue = Shoryuken::Client.queues(job.queue_name)
+          if queue.fifo?
+            raise ArgumentError,
+                  "FIFO queue '#{queue.name}' does not support per-message delays. " \
+                  'When using ActiveJob retry_on with FIFO queues, set `wait: 0`.'
+          end
+        end
+
+        enqueue(job, delay_seconds: delay)
       end
 
       # Bulk enqueue multiple jobs efficiently using SQS batch API.
diff --git a/spec/shared_examples_for_active_job.rb b/spec/shared_examples_for_active_job.rb
index c7c0ba56..cef92f19 100644
--- a/spec/shared_examples_for_active_job.rb
+++ b/spec/shared_examples_for_active_job.rb
@@ -281,6 +281,26 @@ class TestJob < ActiveJob::Base; end
 
       subject.enqueue_at(job, nil)
     end
+
+    context 'when fifo' do
+      let(:fifo) { true }
+
+      it 'raises ArgumentError when delay is positive' do
+        allow(subject).to receive(:calculate_delay).and_return(3)
+        allow(queue).to receive(:name).and_return('test.fifo')
+        expect(queue).not_to receive(:send_message)
+
+        expect { subject.enqueue_at(job, nil) }.to raise_error(
+          ArgumentError, /FIFO queue.*does not support per-message delays/
+        )
+      end
+
+      it 'does not raise when delay is zero' do
+        allow(subject).to receive(:calculate_delay).and_return(0)
+        expect(queue).to receive(:send_message).with(hash_including(delay_seconds: 0))
+        expect { subject.enqueue_at(job, nil) }.not_to raise_error
+      end
+    end
   end
 end
 # rubocop:enable Metrics/BlockLength