n8n Human-in-the-Loop: Approval Workflows for AI Agents (2026)

Automation has one fatal flaw: it never hesitates. Your AI agent doesn't feel a twinge of doubt before emailing a client or deleting a row — it just does it. The fix isn't slowing the agent down. It's giving it a supervisor: human-in-the-loop approval, the checkpoint where a person says yes or no before the machine acts.

---

1. Which Actions Need Approval

Approval checkpoints slow things down — add them where the cost of a mistake outweighs the cost of waiting. The rule of thumb that has survived every agent project I've run:

Apply the same lens to non-agent workflows. A nightly backup doesn't need a human. A workflow that deletes inactive users — that one does.

---

2. The 3 Ways to Do HITL in n8n

n8n gives you three approaches, from native to DIY. Know all three — each earns its place:

The honest guidance: approaches 1 and 2 are easier and native — use them first. Approach 3 (section 9) is the escape hatch for when approvals must survive days, because a workflow paused for days is an execution waiting to time out, get pruned, or tie up a queue-mode worker.

---

3. Native #1: Human Review on AI Agent Tools

Since n8n 2.6 (January 2026), the cleanest approval in the whole platform is built into the AI Agent itself. Here's the flow:

On denial, the tool call is canceled and the AI is informed of the rejection — which is exactly where the next section begins.

---

4. Teaching the Agent to Handle Denials

The most overlooked part of HITL is the prompt. A denied agent with no instructions does the dumbest possible thing: it retries the same tool, gets denied again, and loops until your token budget cries. Your system prompt needs explicit denial behavior:

If an approval for a tool is denied, do NOT retry that tool.
Instead, inform the user of the rejection, suggest an alternative,
or ask the user for clarification — whichever fits the situation.
Never loop on a denied tool call.Copy

Also describe which tools require approval and why — an agent that knows "sending an email needs a human" asks cleaner questions and sets better user expectations.

---

5. Native #2: The Chat Node (Send & Wait)

The n8n Chat has two human-in-the-loop operations, and they fit both positions in a workflow:

• Send a message — notify the user and continue; no pause.
• Send a message and wait for response — pause execution until the user replies, with free text or inline approval buttons you define in the node.

Two placements, matching the two philosophies:

Setup detail that trips people up: for the Chat node to pause and reply mid-workflow, the Chat Trigger must run with Response Mode: Using Response Nodes — set in the trigger's parameters. Without it, the chat answers once and the wait never happens.

---

6. Slack Approvals (Full Setup)

Slack is the favorite approval channel for teams — and the only one with real setup requirements. From the official changelog, the complete checklist:

---

7. Telegram Approvals (Full Setup)

Telegram approvals work the same way but with one advantage: no public HTTPS requirement — your bot's webhook setup already handled that, or the polling community node avoids it entirely (both covered in our Telegram guide).

• Operation: choose Send and Wait for Response on the Telegram node, with approval buttons in the reply markup.
• One-tap approvals: the approver taps Approve/Deny inside Telegram itself — no links, no forms.
• Capture Who Responded: enable it in the same Advanced Interactivity section as Slack — the output records who tapped, their username, chat, and message ID.
• Restrict Who Can Approve: list allowed Telegram user IDs; everyone else's taps are ignored.

If your approval needs to reach a single person, DM them directly. If a team should share the load, route requests to a private group — the approvers get one shared queue, and Capture Who Responded tells you which of them acted.

---

8. Timeouts & Auto-Decisions

Every approval request will, one day, go unanswered. Design for it before it happens:

• Short waits (minutes–hours): native review steps hold the execution. Acceptable while your execution timeouts allow it.
• Long waits (days): a paused workflow is a hostage — it occupies an execution slot and can hit the execution timeout. Use the two-stage pattern (section 9): the requesting workflow ends after sending; a second workflow reacts whenever the answer arrives. Nothing hangs.
• Auto-decisions: decide your policy in advance — auto-approve when the risk is low and the delay is worse than the mistake; auto-reject when the action is irreversible. Write it down, because at 2 AM you'll be tempted to improvise.

---

9. Free Template: Two-Stage Telegram Approval

The two-stage pattern from section 8, implemented: Workflow A sends the approval request with Approve/Reject buttons and finishes; Workflow B listens for the tap and executes or cancels. No hung executions, works on every n8n version, and every node type below matches verified exports.

Workflow A — Request Approval

{
  "name": "Approval Requester",
  "nodes": [
    {
      "parameters": {},
      "id": "manual",
      "name": "When Tested / Triggered",
      "type": "n8n-nodes-base.manualTrigger",
      "typeVersion": 1,
      "position": [250, 300]
    },
    {
      "parameters": {
        "jsCode": "// Replace with the action your workflow is proposing\nconst action = 'Publish post #42 to LinkedIn';\nconst chatId = 'YOUR_TELEGRAM_CHAT_ID'; // personal ID or private group\n\nconst payload = {\n  chat_id: chatId,\n  text: `🔍 Approval needed:\\n\\n${action}\\n\\nApprove or reject:`,\n  reply_markup: {\n    inline_keyboard: [[\n      { text: '✅ Approve', callback_data: 'approve_42' },\n      { text: '❌ Reject', callback_data: 'reject_42' }\n    ]]\n  }\n};\n\nreturn [{ json: { body: JSON.stringify(payload) } }];"
      },
      "id": "build",
      "name": "Build Request",
      "type": "n8n-nodes-base.code",
      "typeVersion": 2,
      "position": [450, 300]
    },
    {
      "parameters": {
        "method": "POST",
        "url": "=https://api.telegram.org/bot{{ $env.TELEGRAM_TOKEN }}/sendMessage",
        "sendBody": true,
        "specifyBody": "json",
        "jsonBody": "={{ $json.body }}",
        "options": {}
      },
      "id": "send",
      "name": "Send Request (Bot API)",
      "type": "n8n-nodes-base.httpRequest",
      "typeVersion": 4.2,
      "position": [650, 300]
    }
  ],
  "connections": {
    "When Tested / Triggered": {
      "main": [
        [
          {
            "node": "Build Request",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "Build Request": {
      "main": [
        [
          {
            "node": "Send Request (Bot API)",
            "type": "main",
            "index": 0
          }
        ]
      ]
    }
  },
  "active": false,
  "settings": {
    "executionOrder": "v1",
    "saveManualExecutions": true
  }
}Copy

Workflow B — Handle the Answer

{
  "name": "Approval Handler",
  "nodes": [
    {
      "parameters": {
        "updates": ["callback_query"],
        "additionalFields": {
          "download": true
        }
      },
      "id": "trigger",
      "name": "Telegram Trigger",
      "type": "n8n-nodes-base.telegramTrigger",
      "typeVersion": 1.1,
      "position": [250, 300],
      "credentials": {
        "telegramApi": {
          "id": "YOUR_TELEGRAM_CREDENTIAL_ID"
        }
      }
    },
    {
      "parameters": {
        "jsCode": "const item = $input.item.json;\nconst cq = item.callback_query;\nconst data = cq.data;                 // 'approve_42' | 'reject_42'\nconst approved = data.startsWith('approve');\nconst actionId = data.split('_')[1] || data;\n\nreturn [{\n  json: {\n    approved,\n    actionId,\n    callbackQueryId: cq.id,\n    chatId: cq.message.chat.id,\n    approverName: cq.from.first_name || 'Unknown'\n  }\n}];"
      },
      "id": "parse",
      "name": "Parse Tap",
      "type": "n8n-nodes-base.code",
      "typeVersion": 2,
      "position": [450, 300]
    },
    {
      "parameters": {
        "method": "POST",
        "url": "=https://api.telegram.org/bot{{ $env.TELEGRAM_TOKEN }}/answerCallbackQuery",
        "sendBody": true,
        "specifyBody": "json",
        "jsonBody": "={\"callback_query_id\": \"{{ $json.callbackQueryId }}\", \"text\": \"Thanks!\"}",
        "options": {}
      },
      "id": "answer",
      "name": "Close Spinner",
      "type": "n8n-nodes-base.httpRequest",
      "typeVersion": 4.2,
      "position": [650, 300]
    },
    {
      "parameters": {
        "conditions": {
          "options": {
            "caseSensitive": true,
            "leftValue": "",
            "typeValidation": "strict"
          },
          "conditions": [
            {
              "id": "c1",
              "leftValue": "={{ $json.approved }}",
              "rightValue": true,
              "operator": {
                "type": "boolean",
                "operation": "equals"
              }
            }
          ],
          "combinator": "and"
        }
      },
      "id": "if-approved",
      "name": "Approved?",
      "type": "n8n-nodes-base.if",
      "typeVersion": 2,
      "position": [850, 300]
    },
    {
      "parameters": {
        "method": "POST",
        "url": "=https://api.telegram.org/bot{{ $env.TELEGRAM_TOKEN }}/sendMessage",
        "sendBody": true,
        "specifyBody": "json",
        "jsonBody": "={\"chat_id\": \"{{ $('Parse Tap').item.json.chatId }}\", \"text\": \"✅ Approved by {{ $('Parse Tap').item.json.approverName }} — action #{{ $('Parse Tap').item.json.actionId }} is running.\"}",
        "options": {}
      },
      "id": "notify-approve",
      "name": "Notify Approved",
      "type": "n8n-nodes-base.httpRequest",
      "typeVersion": 4.2,
      "position": [1050, 200]
    },
    {
      "parameters": {
        "method": "POST",
        "url": "=https://api.telegram.org/bot{{ $env.TELEGRAM_TOKEN }}/sendMessage",
        "sendBody": true,
        "specifyBody": "json",
        "jsonBody": "={\"chat_id\": \"{{ $('Parse Tap').item.json.chatId }}\", \"text\": \"❌ Rejected by {{ $('Parse Tap').item.json.approverName }} — action #{{ $('Parse Tap').item.json.actionId }} cancelled.\"}",
        "options": {}
      },
      "id": "notify-reject",
      "name": "Notify Rejected",
      "type": "n8n-nodes-base.httpRequest",
      "typeVersion": 4.2,
      "position": [1050, 400]
    }
  ],
  "connections": {
    "Telegram Trigger": {
      "main": [
        [
          {
            "node": "Parse Tap",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "Parse Tap": {
      "main": [
        [
          {
            "node": "Close Spinner",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "Close Spinner": {
      "main": [
        [
          {
            "node": "Approved?",
            "type": "main",
            "index": 0
          }
        ]
      ]
    },
    "Approved?": {
      "main": [
        [
          {
            "node": "Notify Approved",
            "type": "main",
            "index": 0
          }
        ],
        [
          {
            "node": "Notify Rejected",
            "type": "main",
            "index": 0
          }
        ]
      ]
    }
  },
  "active": false,
  "settings": {
    "executionOrder": "v1",
    "saveManualExecutions": true
  }
}Copy

• Where your real action goes: replace the "Notify Approved" message with the actual execution — publish the post, run the refund, delete the rows. The two notify nodes are the placeholders for your approve/reject branches.
• One token, one env var: both workflows read TELEGRAM_TOKEN from the environment — nothing sensitive sits in the JSON (same pattern as our Telegram guide).
• Scale it up: put the action ID and the actual payload into the callback data, or store pending actions in a sheet/database and look them up on approval — the two-stage structure stays the same.

---

10. The 8 Most Common Errors & Fixes

---

11. Frequently Asked Questions

---

Sources & References

• n8n Documentation — Human-in-the-loop for tools
• n8n Documentation — Changelog (approval features & Chat node operations)
• n8n Documentation — AI Agent nodes
• n8n Community — How the community handles HITL steps
• n8n Documentation — Wait node

Explore More Guides