AI budget allocation for ads: when to move spend and how much

AI budget allocation works when the AI does the reading and you do the deciding. Point it at live account data and it is fast at two things that eat your Monday: finding the campaigns spending with nothing to show for it, and finding the ones that are capped by budget while hitting target. Then it can make the change, one step at a time, after you confirm. What it cannot do is weigh what is not in the data: margin, inventory, the week your client sells the most. Here is what to ask, when and how much to move, and the guardrail that keeps a typo from becoming a US$ 2,000 day.

How to split spend between Meta, Google and TikTok in the first place is covered in how to split budget across channels, and the signals that justify growth are in when to scale a campaign. This post is about what changes when an AI assistant joins the process. It is one step in the full method for optimizing ad campaigns with AI.

What the AI has to see before it suggests anything

Without data, an AI's budget recommendation is a persuasive paragraph about nothing. Before you ask "where should I put more money?", make sure it can read:

• Spend and results by campaign over the same window, 14 or 30 days. Shorter than that is mostly noise.
• Cost per result by day, so you know whether the number is stable or two lucky days are carrying the average.
• How much of each budget actually gets spent. A campaign that spends its full budget every day may be held back. One that spends half does not need more.
• Qualified leads and sales by campaign, from your CRM. Each platform counts conversions its own way; the CRM is the common yardstick.
• What already changed, in the change history. Stacking a new edit on top of yesterday's makes both impossible to read.

Prompts that end in a decision

Five that work, in this order:

1. "Which campaigns spent more than US$ 500 in the last 14 days with zero conversions?" Candidates to lose budget.
2. "Which campaigns spent their full daily budget every day with cost per result under US$ 40?" Candidates to gain.
3. "Using the CRM: cost per qualified lead and cost per sale for each campaign, last 30 days." This is the ranking that matters, and it sometimes flips the cost-per-lead ranking.
4. "If I take US$ 75 a day from campaign A and give it to campaign B, what is the percentage change for each?" The AI does the math, so you see how big the step really is.
5. After the change: "Compare the 3 days after the edit with the 7 days before, for both campaigns." This tells you whether the move worked.

The thresholds are examples. Use your own, but always give one. "Expensive campaign" with no number turns into the AI's opinion.

When to move spend, and how much at a time

Our rule of thumb is to move 10 to 20% of total spend at a time, as laid out in the post on splitting budget across channels. That is not an official number from any platform. It is caution.

Why slow? Two reasons. A large budget change can send a campaign back into the learning phase, and costs swing until it settles. And if you change three campaigns at once, you will never know which change did the work. One move, a few days of data, prompt 5, then the next move.

The cap is the guardrail

AI gets budget wrong in a predictable way: it misreads the request. "Bump it 50" can become US$ 50 a day or plus 50%. One extra zero typed in a hurry turns US$ 200 into US$ 2,000. So the safety net should not be the model's good intentions. It should be the system.

In Vazante, the per-action cap is the highest daily budget a single request can set. Say you set it at US$ 400: a request to put a campaign at US$ 600 is refused before it reaches Meta, Google or TikTok. Below the cap, the AI shows the campaign, the current value and the new one, and waits for you to confirm.

A setup that works: put the cap a little above the largest daily budget you run today. The AI can reallocate inside the account's normal range, and a fat-fingered request never turns into a US$ 2,000 day.

What the AI does not know about your budget

Anything that matters for the decision but lives outside Meta, Google, TikTok and the CRM is a blind spot. The usual ones:

• Margin. Product A's campaign can have the best cost per sale and the worst margin.
• Inventory and capacity. More spend where the client is out of stock or short-staffed just generates complaints.
• The client's calendar. The AI sees account history, not the fact that next week is the client's biggest of the year.
• Commitments. If the contract fixes how much goes to each channel, the AI has no idea.
• Deals with no source. Sales that land in the CRM without UTM data vanish from any per-campaign ranking. Ask for the unattributed share every time.

The fix is to say it. "Product A's margin is half of product B's; redo the ranking with that in mind" is a perfectly good prompt. The AI works with the context you give it in the conversation.

Your next move

This week, run prompts 1 and 2 on your biggest-spending account. Pick one move, one step in size, and note the date. Three days later, run prompt 5. If CPL spiked before you even got to budget, start with the diagnosis in lower cost per lead with AI.

To do this on a live account with your cap in place, plans start at US$ 39 a month, and the first 7 days cost US$ 2.99.