First Marketing AI Agent With Approval: n8n, Make or Gumloop

The short answer: if 50 to 300 press, partnership and sponsorship emails a month land in one marketing inbox and wait days for an answer, make that inbox your first agent: it sorts each email, drafts a reply from your written policy, and nothing leaves until a person approves it. On n8n Cloud each email is one execution, so 200 a month use 200 of the 2,500 executions in Starter, $24 a month on monthly billing (n8n pricing, checked on 26 October 2026); Make's Free plan, 1,000 credits a month, carries the same volume by the arithmetic further down if it checks every two hours (Make pricing), and Gumloop's Pro plan starts at $37 (Gumloop pricing). For example, at 4 minutes to answer an email by hand and 1 minute to approve a draft, 200 emails drop from 200 x 4 = 800 minutes to 200 x 1 = 200, so someone gets 600 minutes back a month.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

The inbox earns the first slot for reasons that have little to do with AI. Mail arrives daily, a reply can be judged in a minute, and a bad draft costs nothing while a person stands between it and the sender. To score other jobs first, use the table in building your first marketing AI agent.

Before you pick a builder:

1. Price the checking, not only the replies. n8n counts a polling trigger only when it finds new mail (executions). Make's operations guide counts a trigger as one operation "per check" (operations), and a scenario runs every 15 minutes by default (scheduling). For example, if every check counts, 96 checks a day for 30 days is 2,880 credits, almost three times the Free plan, before a single reply; every 120 minutes is 360 a month.
2. Point it at a label you already control. n8n's Gmail Trigger, Make's Watch emails (Gmail modules) and Gumloop's Gmail trigger (triggers) can all watch one label. A Gmail filter applies it (Gmail filters), and the agent never sees the rest of the inbox.
3. Write the policy before the prompt. One paragraph per category (press, partnership, sponsorship, event invitation, vendor pitch): what you agree to, what you never promise, who gets copied, plus three past replies you would send again. It becomes the instructions in every tool.
4. Take the baseline this week. From the last four weeks under the label: emails received, emails never answered, and hours to first reply for 30 of them.

One job, four moves

The trigger is a new email under the label. The model reads it against your policy, picks a category and drafts a reply for press, partnership and sponsorship; vendor pitches get a label, and anything unclear goes to a person as "other". Then someone sees the draft beside the original and decides. The output is a reply inside the original thread, plus a label recording what happened.

The tools differ at the human check, and that matters more than price. n8n and Gumloop ask a yes-or-no question. Make leaves a draft the approver can rewrite.

Free first: a label, a policy and a habit

By hand, the job needs nothing new: the filter collects the mail, and twice a day one person answers press and partnership mail from the policy, archives vendor pitches and forwards the rest. That routine is enough on its own while one person owns the inbox.

Two free tiers run the agent version without a licence fee. Make's Free plan runs its agent module on Make's AI Provider (agent guide). n8n's self-hosted Community edition is free with most features (n8n docs), but needs a server and your own Google Cloud OAuth client, since one-click Google sign-in exists only on n8n Cloud (Google credentials). Gumloop has no free plan; its 14-day Pro trial needs a card and turns paid unless cancelled (credits page).

In n8n: approve from an email

1. Gmail Trigger on your label, unread only, with Max Emails per Poll set to 1, so each email runs as its own execution and the rest wait for the next poll (Gmail Trigger).
2. Basic LLM Chain with the policy in the prompt and Require Specific Output Format on, feeding a Structured Output Parser with category, draft and reason (Basic LLM Chain). One model call and no tools: the right size for a first build.
3. Switch on category: the three reply categories continue, the others go to a Gmail Add Label step (Switch).
4. Send and Wait for Approval, a Gmail operation that emails the approver the original and the draft, then holds the run. Set Type of Approval to Approve and Disapprove so a decline is recorded (Gmail message operations). There is no edit box: to change a word, decline and answer by hand.
5. If approved, Gmail's Reply operation with Reply to Sender Only on; if declined, an "answer by hand" label (If). Check the attribution setting covered below first.
6. An error workflow starting with the Error Trigger, selected in the workflow's settings (error handling).

For an agent that picks its own tools, how to use n8n puts Gmail behind review in Slack; plans are in the n8n entry.

In Make: the draft in the thread is the approval

1. Watch emails with a simple filter on your label, unread only (Gmail modules), and the scenario set to run every 120 minutes. A personal @gmail.com connection must be reauthorised every six months (Gmail app).
2. Run an agent (Make AI Agent (New)) with the policy as Instructions, the email as Input and Response format set to Data structure (agent guide). Test it in Chat with Agent on ten past emails first.
3. A filter on the line after the agent passes only the reply categories (filtering).
4. Create a draft email with the Thread ID mapped from the trigger, so the draft sits in the original conversation (Gmail modules). Leave Send a draft email out of the scenario.

The person rewrites what needs it and presses Send. Only this build lets the reviewer edit first, and it needs no approval feature, which matters: the help page of Make's Human in the loop app returned a 404 on 26 October 2026.

For example, at 360 checks and 3 credits per email (the agent's operation, its tokens, the draft), 200 emails use 360 + 600 = 960 credits, inside Free. The token share is the assumption to test against History after a week: on the Small tier a credit buys 18,080 input or 2,260 output tokens (credits, agent credit usage). More in how to use Make and the Make entry.

In Gumloop: one card per send

1. Change the Gmail approval setting first. The default is Always allow, under which the agent uses every Gmail tool without asking (human in the loop). Choose Ask for writes/deletes: reading runs freely, sending waits.
2. Give it the policy as instructions, with the past replies as examples.
3. Add the trigger: Gmail, New email in a label, polling roughly every 60 seconds, with @ fields for Sender Address, Subject and Email Body in the prompt (triggers).
4. Approve or reject the card showing To, Subject and Body, in the chat, the notification bell, the Tasks page or Slack; a rejection with a reason makes the agent adjust (human in the loop).

Triggers switch off after 3 failed runs in a row (triggers). Cost per run varies (credits page): for example, if your first ten runs average 80 credits, 200 emails use 16,000 of Pro's 20,000. See how to use Gumloop and the Gumloop entry.

The monthly bill, three ways

Checked on 26 October 2026 on n8n pricing, Make pricing, Make credits, Gumloop pricing and Gumloop credits. Read Make's paid prices in a browser before you outgrow Free.

How to know it worked

• Speed: median hours from arrival to first reply in the reply categories, against the baseline.
• Silence: emails still unanswered after three working days. This should fall fastest.
• The approval split: drafts approved as written, edited or declined, counted weekly. Each decline is a sentence missing from the policy.
• A holdout by category: answer event invitations by hand for four weeks, outside the agent. If reply times fall for the agent's categories and stay flat for invitations, the agent did it; if both fall, the attention did.

The default to switch off first

n8n's Gmail Reply and Send operations end every message with "This email was sent automatically with n8n" unless you turn off Append n8n attribution, so a partner who gets your approved answer also gets an n8n advert that nobody approved (Gmail message operations). Switch it off in step 5 and send one test reply to yourself.

Once one agent runs cleanly, the next question is rarely the tool. It is which jobs a team hands to agents and who signs off on each, and that is the agentic marketing work I sell.

FAQ

Can any of these builds send without a person?

Not as described. n8n replies only after the Approve click, the Make scenario has no send module, and Gumloop holds a gated call until someone approves it.

Which builder suits a team with no developer?

Make, if the reviewer wants to rewrite drafts and the volume fits Free. Gumloop, if the team would rather describe the job than wire modules. n8n Cloud costs least per email at volume, for a team that can read a node's JSON when a run fails.

When can the approval step go?

For replies to people outside the company, never. For the label-only branch, where vendor pitches are filed and nobody is written to, a weekly look is enough once the categories hold for a month.