diff --git a/apps/web/src/app/api/internal/usage/record/route.ts b/apps/web/src/app/api/internal/usage/record/route.ts
index 35a2c8f825..d123f4f297 100644
--- a/apps/web/src/app/api/internal/usage/record/route.ts
+++ b/apps/web/src/app/api/internal/usage/record/route.ts
@@ -11,6 +11,7 @@ import {
 import {
   createPhaseTimer,
   emitUsageRecordTiming,
+  isPrimaryPoolSaturated,
   readPoolGauges,
   shouldEmitUsageRecordTiming,
 } from '@/lib/ai-gateway/usage-record-diagnostics';
@@ -46,6 +47,14 @@ export async function POST(request: NextRequest): Promise<NextResponse> {
   // the pre-check gets confirmed.
   const timer = createPhaseTimer();
   const poolBefore = readPoolGauges();
+  if (isPrimaryPoolSaturated(poolBefore)) {
+    // Do not join an unbounded pg-pool queue. A 503 proves no write started, so
+    // the caller can safely retry this delivery without creating a duplicate.
+    return NextResponse.json(
+      { error: 'Usage record sink busy' },
+      { status: 503, headers: { 'retry-after': '1' } }
+    );
+  }
   let poolWaitingPeak = poolBefore.waiting;
   const samplePool = () => {
     const gauges = readPoolGauges();
diff --git a/apps/web/src/lib/ai-gateway/usage-record-diagnostics.test.ts b/apps/web/src/lib/ai-gateway/usage-record-diagnostics.test.ts
index 95a51e5975..018d47356d 100644
--- a/apps/web/src/lib/ai-gateway/usage-record-diagnostics.test.ts
+++ b/apps/web/src/lib/ai-gateway/usage-record-diagnostics.test.ts
@@ -7,6 +7,7 @@ import { describe, expect, test } from '@jest/globals';
 import {
   createPhaseTimer,
   describeDatabaseError,
+  isPrimaryPoolSaturated,
   isUsageRowConflict,
   readPoolGauges,
   shouldEmitUsageRecordTiming,
@@ -252,6 +253,20 @@ describe('readPoolGauges', () => {
   });
 });
 
+describe('isPrimaryPoolSaturated', () => {
+  test('rejects work when every configured connection is checked out', () => {
+    expect(isPrimaryPoolSaturated({ total: 10, idle: 0, waiting: 245 })).toBe(true);
+  });
+
+  test('accepts work when an idle connection is available', () => {
+    expect(isPrimaryPoolSaturated({ total: 10, idle: 1, waiting: 0 })).toBe(false);
+  });
+
+  test('accepts work when the pool can still open a connection', () => {
+    expect(isPrimaryPoolSaturated({ total: 9, idle: 0, waiting: 0 })).toBe(false);
+  });
+});
+
 describe('shouldEmitUsageRecordTiming', () => {
   test('always emits at or above the slow threshold', () => {
     expect(shouldEmitUsageRecordTiming(1_000, () => 1)).toBe(true);
diff --git a/apps/web/src/lib/ai-gateway/usage-record-diagnostics.ts b/apps/web/src/lib/ai-gateway/usage-record-diagnostics.ts
index d1de25d674..053b99d53f 100644
--- a/apps/web/src/lib/ai-gateway/usage-record-diagnostics.ts
+++ b/apps/web/src/lib/ai-gateway/usage-record-diagnostics.ts
@@ -78,6 +78,16 @@ export function readPoolGauges(): PoolGauges {
   };
 }
 
+/**
+ * Whether a new request would have to queue for the primary pool.
+ *
+ * Fail open if the pool implementation does not expose its configured maximum.
+ */
+export function isPrimaryPoolSaturated(gauges: PoolGauges): boolean {
+  const max = poolMax();
+  return max !== null && gauges.idle === 0 && gauges.total >= max;
+}
+
 function eventLoopLagMs() {
   return {
     mean_ms: Math.round(eventLoopDelay.mean / 1e6),