Which AI in your marketing budget makes the most decisions? In most 2027 plans it is not the AI on the line called AI. The software you licensed, the agents your team built, the copilots with a seat price: those sit on a line, and that line is getting an ROI test. The models that set bids, split budgets, pick audiences and assemble creative inside your ad platforms are priced inside the media rate. They have no line. Nobody asks them for a return, because the only return anyone has for them is the one the platform reports.
I am not arguing that the AI line should be spared. Most of what sits on it should face a hard test, and some of it should go. My argument is narrower: in budget season the test lands on the half of your AI you own and skips the half the seller operates. I made the unit version of this argument in Nobody Is Buying Tokens, and the ownership version in The Default Owner. This is the budget version, and it matters this month because the plan you sign in December is where the default gets written down. The platform's number gets a pass for a reason I spent a whole essay on in The Seller's Counterfactual: it is a quote from the party paid when you act on it.
What the survey asked, and what it could not ask
Validity published a marketer survey on November 11 (more than 250 B2B and B2C marketers; read on PR Newswire November 14). Sixty-one percent said the C-suite is pressing them to show measurable AI gains. Fifty-three percent plan to put at least 10 percent of their 2027 budget into AI. Half said their 2026 AI spending met or beat expectations, 26 percent said it fell short or showed nothing measurable, and 24 percent could not yet tell. It is a vendor survey with no fielding details in the release, and I would not build a case on its percentages.
What caught me was the structure of one question. Asked what they would protect if forced into mid-year cuts, 20 percent named paid media and 16 percent named AI tools and agents, as two separate answers. I think most budget documents carry the same split, and I think it is wrong in an interesting way. The paid media line is full of AI. The respondents protecting paid media are protecting the largest pool of automated decision-making in the company, and the survey, like the plan, files it as something else.
That 53 percent figure has the same blind spot. Ten percent of the budget "into AI" means ten percent into things that arrive with an AI invoice. A company that runs most of its search and shopping spend through fully automated campaign types is already spending far more than that on machine decisions, and none of it shows up in the count.
Why the biggest AI line is invisible
This is plumbing, not a conspiracy. Platform automation has no price of its own. You cannot see a fee for the bidder, or for the model that chose which audience saw which asset, because those costs live inside the cost per click and the cost per conversion. You can cancel a seat. You cannot cancel the bidder without cancelling the media it buys.
The chart of accounts is the list of things a finance team can review. Whatever has no line has no reviewer. So when a CFO asks "what is our AI costing and what is it returning," the honest answer covers the copilots and the agent platform, and leaves out the automation that decides how most of the media money is spent. The question is fair. The ledger it is asked of just cannot see the biggest item.
A practitioner reading this will object that platform automation is not "AI spend," it is just how media is bought now. I agree, and that is my whole point. Because it is how the media buys, it has no owner who chose it as a decision, no denominator anyone wrote down, and no review date. That is what an unpriced decision looks like.
The double standard
Here is where the two lines stop being merely different and start being unfair to each other.
The owned tool has to prove itself against evidence the company controls: a pilot, a holdout, a before and after, a finance model. Often it cannot, because the outcome it improved is credited somewhere else. A better product feed, a cleaner conversion signal, a sharper audience list built by your own agent: all of that flows into the platform's reported return. I wrote about this in Who Counts the Conversion?. Each closed loop counts with its own definition and its own window, and the loop that counts last tends to keep the credit.
The platform automation is graded by its own reporting. Its return is credited conversion value, the same number that I argued in Break-Even ROAS Is a Floor mixes caused sales with sales that would have happened anyway. Causal tests exist, but on the largest platforms they are granted, not held. Google's Conversion Lift page, read November 14, says the product is not available to every account and tells you to contact your Google representative. I made the general point in Who Is Allowed to Measure?, and the specific one about paying the operator for its own audit in You Are Buying the Audit From the Auditee.
Put those together and the outcome follows. In a tight year the owned tool shows a weak number, partly because its contribution was absorbed by the platform's number, and gets cut. The platform automation shows a strong number, partly built on the contribution of the tool being cut, and keeps its budget. Each line looks like it was judged on evidence. Only one of them supplied its own.
The plan does not decide to hand more decisions to the seller. It just applies an ROI test to the only AI it can see.
The fifteen-minute plan audit
Open your draft 2027 plan. List every place where a model, rather than a person, decides one of four things: what to bid, how to split budget, who sees the ad, or which creative they see. Do not limit yourself to Google. Meta's automated campaign types, Amazon's bidding, retail media networks, the ad products arriving in chat assistants: all of them qualify. Then fill in one row per item.
| Where | What the model decides | Own budget line? | Named owner | Denominator | Who reports the return |
|---|---|---|---|---|---|
| Shopping, fully automated | Bid, audience, placement, creative mix | No (inside paid media) | None written | Credited revenue | The platform |
| In-house audience agent | Which customers to suppress and which to seed | Yes (AI tools) | Marketing ops lead | Asked to prove ROI | Nobody, its outcomes are credited elsewhere |
Now count. How many rows have their own budget line? How many have an owner whose name is written down? My bet is that in most companies the answer for the platform rows is zero on both counts. That count is the finding. It does not say the platform is doing a bad job. It says the company has no way to know, and has agreed in advance not to ask.
One test, both lines
The fix is not to protect the tools. It is to grade both lines on the same three questions.
First, what is the denominator? Name the business unit the decision is supposed to produce, not the activity it generates. On this site that unit has a name for each vertical: cost per logo in B2B software, cost per bound policy in insurance, cost per graduate in higher education, cost per financed vehicle for dealer groups, cost per booked job in home services. A platform campaign that reports cost per lead and an in-house agent that reports hours saved fail this question equally. Both get rewritten into the same unit before either is judged.
Second, who reports the return? If the answer is the party that sells the media, the number goes in the plan as a claim, with a date by which you will check it independently: a geographic holdout, a matched-market test, a lift study you commission rather than accept. If no independent check is possible this year, write that down too. A known gap is a managed risk. An unknown one gets renewed every year by default.
Third, what would make you stop? Every line gets a kill condition, stated as a price rather than a feeling. For the owned tool it might be a cost per decision above a threshold for two quarters. For the platform automation it might be a holdout showing incremental return below the floor you set for the account. Either way, it is written before the money is spent.
Some tools will fail this test and should go. I would expect a fair number of seat-priced AI tools to fail it, and I would cut them without regret. Some platform automation will pass easily, and some will turn out to be paid largely for sales you already had. The point is that you find out which is which, instead of letting the line structure decide.
What this changes in the planning meeting
The strongest objection I expect comes from finance, and it deserves a straight answer. "The platform number ties to revenue. When we spend more, sales go up. Your tool shows a slide." That is often true, and it is a reasonable basis for a first cut. Credited return and real return do tend to move together. They move together less tightly than the dashboard implies, and the gap is largest on the spend that would have converted anyway, but a CFO who trusts the line that correlates with revenue is not being foolish.
What I would ask that CFO for is not trust in the tools. It is a calendar. The plan signed in December carries the platform automation at its reported return, flagged as unverified. Some time in the first quarter, one independent test runs on the largest automated line, a regional holdout if nothing else is available. By the end of the second quarter the plan is re-forecast on what the test found, and the tools that could not name a denominator are gone by then too. Nothing about this slows the budget down. It just means that by mid-year both halves of the AI have been graded by someone other than themselves.
I am not sure how many companies will do this in 2027. My guess is not many, because the platform line is comfortable and the test costs money and attention in a year when both are short. That is fine. Write the line down anyway. A decision with a name and an owner, even an unchecked one, is a long way ahead of a decision nobody can see.
The claim
Between now and June 30, 2027, none of the three marketer budget studies most likely to be cited in planning decks (Gartner's annual CMO Spend Survey, IAB's annual Outlook, and Validity's next marketer survey) will count spend that runs through fully automated campaign types as AI spend, or break it out as a separate AI category.
If any one of them does, I was wrong, and the budget conversation will have started to see the line I am describing. I would be glad to be wrong about this one.
Put a line in the 2027 plan called "decisions bought from the seller." List the automated campaign types under it, give it an owner, a denominator and a kill condition, and grade it on the same test as the tools you are about to cut. If the platform automation is as good as its reporting says, it will pass, and you will know why you are paying for it. If you leave it off, the plan has still made a decision. It has just left the seller to make it, and to report how well it went. The working tools for this, a calculator included, are on the Cost Per Decision hub.