I was on a call with a new AE hire last week. She was asking, reasonably, what our campaign-cost numbers looked like for the quarter - the kind of question every AE should ask when they're trying to understand the shop they just joined. And I started to answer the way I would have answered it eighteen months ago: cost per lead, cost per opportunity, content production cost, paid media spend, the standard ledger.
Then I stopped, because the answer I was giving her wasn't actually true anymore. The biggest variable in our campaign cost structure right now isn't paid media. It isn't content production. It isn't the agency line. It's the AI tokens we're burning to run the campaign - and I genuinely cannot tell you, in advance, what a campaign will cost in tokens. I can tell you what it cost after we ran it. I cannot tell you what it will cost while I'm planning it.
So I told her the truth, which felt awful: I have a black box at the heart of my budget right now, and the black box is making me a worse CMO than I was last year.
The biggest variable in my campaign cost structure is the one I literally cannot forecast. That's not a planning problem. That's a planning paralysis problem.
What used to work
CMO budget planning, for as long as the role has existed, has been a discipline of bounded uncertainty. You don't know exactly what a campaign will return, but you know what the inputs will cost. Paid media is a quoted CPM. Agency work is a quoted retainer. Content production is a quoted day rate. Software is a quoted seat fee. The uncertainty is on the upside — will the campaign perform? - not on the cost side.
That structure is what makes campaign planning possible. You allocate against known costs, you set targets against expected returns, you commit to a number in front of your CFO with reasonable confidence that the actual spend will land within ten percent of the plan. The variance you absorb is on returns, not on costs. The CFO accepts that uncertainty because at least one half of the equation is locked.
The agentic era broke this. The cost half of the equation is no longer locked. And nothing in the existing CMO playbook prepared me for what that does to planning.
What a single campaign actually costs in tokens, and why I can't tell you in advance
Consider the simplest version of a campaign I'm running this quarter. The plan is to produce a series of articles, a set of LinkedIn posts, a webinar follow-up sequence, a few customer case studies, and the supporting visuals. Conventional cost structure: I'd quote you a number within a few thousand dollars.
Here's why I can't anymore.
Each article draft goes through somewhere between four and twelve agent passes — research, structure, draft, edit, fact-check, voice-match, format. I don't decide that number in advance; the draft does, based on how clean the first pass comes out. A piece that lands well in the first draft costs me a third of what a piece that needs heavy iteration costs. I can't predict, before I see the first draft, which one this article will be. The cost isn't decided at planning time. It's decided at execution time, in real time, by the work itself.
Then there's the model-tier question. Some passes need a heavier model; some can run on a cheaper one. The orchestration layer makes a judgment in the moment about which to use. Sometimes that judgment is wrong and I have to re-run the pass on a heavier model, which means I paid for both passes. The decision isn't mine to make in advance; it's the system's to make in flight. The cost is, again, a function of the work.
Then there's the consumption-creep problem Sree wrote about in his Honest Math piece. Every time the team gets more comfortable with the tools, they ask the tools to do more. Last quarter we used the visual-generation agent for the hero images. This quarter we're using it for every supporting graphic in every article, plus social variants, plus the carousel formats, plus the reel storyboards. The capability didn't cost more per call — in fact, per-call costs have come down. But the total spend has gone up significantly because we're using it for ten times as many things.
I am, in practice, planning a campaign whose total cost is the product of three unknowns: how cleanly the AI drafts the work (variable per asset), which model tier the orchestration layer picks (variable per pass), and how much new ground the team will cover with the tools by the time the campaign runs (variable per quarter, always upward). Three unknowns multiplied together is not a forecast. It is a wish.
The half-true answers I have right now
I want to be honest about what I'm doing in the meantime, because the version of this article that pretends I've solved it would be a worse article than the one that admits I haven't.
I am tracking historical token spend by campaign type and using that as a rough multiplier for similar campaigns. This works as long as the next campaign looks like the last one. The problem is that the next campaign never looks like the last one — the team's capability has grown, the agents have changed, the model providers have updated their pricing, and the scope has shifted. The historical multiplier is directionally correct and specifically wrong, which is the worst kind of correct for a planning input.
I am setting soft caps on per-campaign token spend and asking the agent orchestration layer to flag when we're approaching them. This catches runaway consumption after it's happened, not before. It's a smoke alarm, not a thermostat. Useful, but not the planning instrument I need.
I am building a small buffer into every campaign budget — ten to twenty percent reserved against token overruns. The CFO has been reasonable about this. But a twenty percent buffer is not a forecast; it's a confession that I can't forecast. And buffers tend to get used, because the constraint that justified them was real.
A twenty percent buffer is not a forecast. It's a confession that I can't forecast. And buffers tend to get used.
Why this is making me a worse CMO
The reason I called myself a worse CMO at the start of this piece isn't melodrama. It's specific. There are three concrete ways the black box is degrading the quality of work I do.
First, I am making smaller bets than I should be. When you can't forecast cost, the safe move is to scope down — fewer agent passes, less ambitious content, fewer simultaneous experiments. Smaller campaigns have smaller cost variance, but they also have smaller returns. I am hedging on the input side because I can't model the cost side, and the hedging is showing up as conservatism in places where my job is to be ambitious. That's worse for the company.
Second, I am explaining my budget poorly to my CFO. The conversation used to be “here's the spend, here's the expected return, here's the variance.” The conversation now is “here's the spend, plus or minus the variable AI consumption, which I can't fully model but I think will land in this range, with a buffer in case it doesn't.” That's not a budget conversation. That's a hedge conversation, and my CFO can tell the difference. The trust math degrades a little every quarter I cannot give him a number he can plan against.
Third, I am slower to commit to things. The decision to run a particular campaign used to take a day; it now takes a week because I'm modeling token-cost scenarios in spreadsheets that I know are wrong. The slowness is a tax on the work. It is also, ironically, the kind of friction the agentic era was supposed to eliminate.
Why this isn't a vendor problem to solve away
Some of you reading this are about to type me a comment with a tool recommendation. There's a tool that tracks per-task AI cost. There's a vendor that promises consumption forecasting. There's a category of “AI FinOps” solutions emerging to address exactly this problem.
I have looked at most of them. They are useful at the observability layer — they show me what I spent, broken down by call, model, task, team. They are honest about being weak at the prediction layer — because the unknowns I listed above are not unknowns the tools can resolve. How cleanly the AI will draft a particular piece of work is a function of the work and the model, not a function that an observability layer can forecast in advance. The tools can tell me what just happened. They cannot tell me what will happen, because nobody can.
The reason this is hard isn't a missing product. It's that the underlying cost structure of agentic work is genuinely different from anything CMOs have planned against before. Sree's Honest Math piece named this from the CFO's chair — AI tokens are an uncapped variable cost in a budget category that was historically capped. I am writing this piece from the chair next to the campaign dashboard, where the lived experience of that uncapped variable cost is a planning paralysis I haven't seen described anywhere yet.
This is a productivity problem, not a capacity problem
I have been writing this article for about two weeks now and the closing kept refusing to land. The diagnostic was honest; the half-true answers were honest; the worse-CMO admission was honest. But the piece wouldn't close, and I couldn't tell why. Then last Tuesday it landed, and I think the rest of the piece reads differently once you see what I missed.
Everything I have just described is only a problem if you are still planning your marketing function as a productivity function. The whole structure of the anxiety I named — the black box, the failing forecast, the CFO trust degrading, the smaller bets — is the structure that emerges when your governing metric is outputs per dollar of input. Productivity is a denominator metric. The dollar is on the bottom of the fraction. If you can't measure the dollar reliably, the metric breaks. And once the metric breaks, the planning breaks with it.
Capacity is a different metric. Capacity asks what becomes possible that wasn't possible before - how many simultaneous deals a rep can carry, how many campaigns a marketing team can run, how many customer relationships a CSM can hold without dropping any of them. The dollar is still in the equation; it just stops being the central planning question. The central planning question in a capacity worldview is whether the input — whatever it costs, within reason — unlocks possibilities that compound. You don't manage capacity by forecasting the unit cost. You manage it by deciding which capacity unlocks are worth the variance, and then committing.
Productivity is a denominator metric. Capacity is a possibility metric. You cannot manage one with the discipline of the other.
The reason I got stuck is that I have been trying to plan a capacity-tool function with the planning discipline of a productivity-tool function. The tools my team uses every day — the agent fleet, the orchestration layer, the model-tiered drafting workflow — are capacity tools. They are not making us cheaper per article. They are making it possible for us to ship articles, campaigns, doctrine pieces, customer programs, and partner content at a volume and depth that was structurally impossible eighteen months ago. The variance in unit cost that has been driving me crazy is the price of a kind of work that didn't exist before. There is no productivity-era benchmark for it because productivity-era frameworks don't measure what these tools actually do.
Sree's Honest Math piece made this argument at the macro level — the cost category itself behaves differently in the agentic era, and CFOs need to update their model of what “marketing spend” even means. I am making the operational version of the same argument from the other chair. The black box at the heart of my budget is not a forecasting failure on my part. It is a worldview mismatch between the planning discipline I was trained on and the tools the function now runs on.
That doesn't make the unit-cost variance disappear. It still exists; it still matters at audit time; the CFO still wants numbers. What changes is that it stops being the governing question. The governing question becomes capacity: what did the function ship this quarter that the function could not have shipped last quarter, and was the input that made that possible worth what it cost? That is a question you can answer honestly. The forecasting question, as currently structured, is one I have stopped pretending to be able to answer.
None of this is solved. The planning discipline for capacity-era marketing functions does not yet exist as a mature practice; we are all writing it. But naming the worldview shift is the move that lets me stop blaming myself for failing to do productivity-era planning on a capacity-era function. The job isn't to forecast the variance. The job is to choose the capacity unlocks worth the variance, and to defend that choice clearly to the people who pay for it.
The question I'm asking other CMOs
I am writing this article in part because I want to know if other CMOs are feeling this and if anyone has gotten further with it than I have. The version of this piece that lectures other CMOs on how to solve their planning problem would be dishonest, because I haven't solved my own. The version that names the problem honestly and invites the conversation is, I think, more useful.
So here is what I would actually like to know from the CMOs reading this:
- Are you planning your function as a productivity function or as a capacity function? And if it's the former - are you feeling the same mismatch I described, or have I overstated it?
- Are your CFOs accepting buffer-based budgeting indefinitely, or are they pushing back on you to find a real forecast? Mine is patient. I don't know how long that lasts.
- Are you scoping down on ambition the way I was? Or are you finding a way to keep the ambition high and absorb the variance somewhere else in your stack?
- Is anyone running internal token-budget chargebacks to teams the way RevOps runs SaaS-license chargebacks? Has that worked, or has it created the wrong incentives?
- If you have made the worldview shift - from productivity to capacity - inside your function: what changed in how you brief your CFO and your board? That is the conversation I most want to have.
Where this goes next
I don't expect the planning discipline for capacity-era marketing to be solved in the next quarter. I expect, honestly, that it is two or three years away from being legible the way conventional CMO planning is legible today. We are in the messy middle of building it — every CMO running an AI-native marketing function right now is, whether they realise it or not, contributing data to a discipline that doesn't fully exist yet.
What I am committed to doing in public is naming the problem honestly. I am a worse CMO than I was a year ago in the specific sense that I cannot forecast my biggest variable cost using the productivity framework I was trained on. I am also a better CMO than I was a year ago in every other sense that matters in a capacity framework — the work the team is producing, the capacity per person, the speed of execution, the depth of what we can ship. The trade is, on balance, worth it. But the asymmetry between what I can do and what I can plan with the old framework is real, and pretending it isn't would be dishonest to anyone trying to figure out how to do this job in 2026.
If you are running an agentic marketing function and the planning side of it feels harder than it used to, you are not behind. You are at the front of a discipline that hasn't been written yet. We're going to have to write it together.
