Approve AI Agent Ad Spend and Marketing Budget Changes

Approve AI agent ad spend and marketing budget changes before money moves: the agent proposes each bid, budget or campaign change, and a human signs off first.

Why spend is different from content

A bad draft costs you embarrassment. A bad budget change costs you cash by the hour. An agent that optimizes campaigns can double a daily cap after misreading a spike, unpause a campaign that was paused for a legal reason, or shift budget on stale numbers. Platforms will happily spend whatever the API allows.

A gate before the change is applied turns "the agent wasted 2,000 overnight" into "the reviewer saw a card that said the daily cap goes from 50 to 500 and tapped reject."

What the reviewer needs on the card

The approval is only as good as what the human can see. Put the numbers in the preview, not just a summary sentence: current value, proposed value, the delta, the scope, and the agent's stated reason. A markdown table renders well in the card.

const IMPRI = "https://api.impri.dev";
const headers = {
  Authorization: `Bearer ${process.env.IMPRI_API_KEY}`,
  "Content-Type": "application/json",
};

type BudgetChange = {
  campaign: string;
  currentDaily: number;
  proposedDaily: number;
  reason: string;
};

export async function proposeBudgetChange(c: BudgetChange): Promise<boolean> {
  const delta = ((c.proposedDaily - c.currentDaily) / c.currentDaily) * 100;

  const res = await fetch(`${IMPRI}/v1/actions`, {
    method: "POST",
    headers,
    body: JSON.stringify({
      kind: "ads.budget_change",
      title: `${c.campaign}: daily budget ${c.currentDaily} -> ${c.proposedDaily}`,
      preview: {
        format: "markdown",
        body:
          `| Field | Value |\n|---|---|\n` +
          `| Campaign | ${c.campaign} |\n` +
          `| Current daily | ${c.currentDaily} |\n` +
          `| Proposed daily | ${c.proposedDaily} (${delta.toFixed(0)}%) |\n\n` +
          `**Agent's reason:** ${c.reason}`,
      },
      expires_in: 3600, // a bid decision based on old metrics should lapse
      editable: ["preview.body"],
      idempotent: true,
      undo: `Set ${c.campaign} daily budget back to ${c.currentDaily}`,
    }),
  });
  const { id } = (await res.json()) as { id: string };

  let action: any;
  do {
    await new Promise((r) => setTimeout(r, 15_000));
    action = await (await fetch(`${IMPRI}/v1/actions/${id}`, { headers })).json();
  } while (action.status === "pending");

  if (action.status !== "approved") return false;

  // Your own function that calls the ad platform; Impri never touches it.
  const applied = await applyBudgetToPlatform(c.campaign, action.decision.final_preview.body);

  await fetch(`${IMPRI}/v1/actions/${id}/result`, {
    method: "POST",
    headers,
    body: JSON.stringify({ status: applied ? "executed" : "execute_failed" }),
  });
  return applied;
}

Note the edit path. If you set editable: ["preview.body"], the reviewer can change the proposed number before approving, so parse the amount from final_preview rather than trusting the value the agent sent. applyBudgetToPlatform must validate whatever it parses.

Layering your own limits

Impri does not know what a sensible budget is. Keep hard limits in code as well as in the human gate:

• Cap outside the agent. Enforce a maximum per-campaign and per-day figure in the executor, so an approval cannot push past a ceiling you set on purpose.
• Split by risk. Send small nudges (a few percent) straight through and route only changes above a threshold to the inbox. Deciding the threshold is your logic, not Impri's.
• Distinct kind values such as ads.budget_change, ads.pause, ads.creative_launch keep the inbox sortable and let you filter the audit log when finance asks who approved a change.

Boundaries

Impri is only the approval gate. It stores the proposal, notifies the reviewer, and records the decision. It does not analyze performance, decide whether a budget is reasonable, or talk to Google Ads, Meta, or any other platform.

It is a real control only when the ad-platform credential lives in the executor that runs after approval. If the agent also holds a token that can change budgets directly, it can bypass the gate. Confine it by wrapping the platform call, as covered in the integrations guide.

Use a short expires_in. Approving a change an hour late means approving it against data that no longer holds.

Next step

Start with the quickstart to get a key, then gate a single action type, budget increases, before adding pauses and creative launches. If your reviewers live in chat, Slack approval puts the card where they already are.