An Approval Workflow for AI Marketing Automation

Marketing automation agents draft campaigns, schedule sends, and touch ad spend — an approval workflow keeps a human signing off before copy goes out or budget moves.

Two different risk profiles, one gate

Marketing automation covers a wide range of actions, and they don't all carry the same risk:

Reversible, low-stakes: scheduling a social post in a drafts queue, generating A/B variants of subject lines nobody sees yet.
Irreversible or costly: sending an email to your full list, publishing a post live, raising a daily ad budget cap.

The second category is where an autonomous agent can do real damage fast — a typo in a subject line reaching 50,000 inboxes, or a budget change that silently 10x's your daily ad spend because the agent misread a KPI. Those are the actions worth gating. The first category can usually run unattended; gating everything just trains reviewers to rubber-stamp without reading, which defeats the point.

Wiring it up over MCP

If your marketing agent runs inside Claude Code, Claude Desktop, or another MCP client, the Impri MCP server handles the push/poll/report cycle as tool calls — no HTTP client needed in your agent code. Configure it once:

json

{
  "mcpServers": {
    "impri": {
      "command": "npx",
      "args": ["@impri/mcp"],
      "env": {
        "IMPRI_API_KEY": "im_your_key_here"
      }
    }
  }
}

Then the agent's flow for a campaign send looks like this (shown as the tool calls it makes):

impri_push_action(
  kind="campaign.send",
  title="Send: August newsletter — 'What we shipped this month'",
  preview={
    format: "markdown",
    body: "Subject: What we shipped this month\n\nHi {{first_name}},\n\n..."
  },
  expires_in=14400,
  editable=["preview.body"]
)
→ { action_id: "act_9f2", status: "pending", inbox_url: "..." }

impri_await_decision(action_id="act_9f2", timeout_s=14400)
→ { status: "approved", preview: { body: "..." }, edited_by_human: true }

# only now does the agent call your ESP
send_campaign(list_id="newsletter", body=decision.preview.body)

impri_report_result(action_id="act_9f2", status="executed")

If a marketer edits a line in the subject or a CTA before approving, edited_by_human: true and the corrected body come back in the same response — the agent never has to ask for the edit separately, it just uses what it's given.

Ad spend changes need a different kind of care

A budget change isn't content a reviewer reads and approves — it's a number they need to sanity-check against context they have and the agent doesn't (this week's performance, upcoming promotions, finance limits). Two fields make that review possible instead of a leap of faith:

bash

curl -X POST https://api.impri.dev/v1/actions \
  -H "Authorization: Bearer $IMPRI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "kind": "ad_campaign.budget_update",
    "title": "Raise daily budget: Q3 retargeting — $80 → $150",
    "preview": {
      "format": "plain",
      "body": "Campaign: Q3 Retargeting\nCurrent daily budget: $80\nProposed: $150\nReason: CTR up 34% over last 7 days, spend capped before 6pm daily"
    },
    "idempotent": false,
    "undo": "PATCH ad campaign budget back to $80 via platform API",
    "expires_in": 21600
  }'

undo tells the reviewer how to reverse it if the new budget turns out to be wrong after the fact — cheap to write, and it changes a "trust the agent" decision into a "worst case is a two-minute fix" decision. idempotent: false puts a warning on the card so a distracted reviewer can't approve the same bump twice from two different notifications.

What this workflow does and doesn't cover

Impri stores the proposed action, notifies whoever reviews marketing sends, and holds the decision — it doesn't know your brand voice, your send-time best practices, or whether "150" is a reasonable ad budget for your account. That judgment stays with the human and with however you've prompted the agent to draft in the first place.

It's also not a campaign scheduler or a workflow engine — if you need multi-step sequencing (draft → legal review → design review → send), build that in your own pipeline or a tool like n8n and call Impri for the final human sign-off step. See MCP integration for wiring this into a Claude Agent SDK or LangChain-based marketing agent, and approve AI-generated content before publishing for the content-review side of this same pattern.

Next step: quickstart to get an API key and wire this into your first campaign.