Ramp published the raw AI spend data from its enterprise customers this week. The number that traveled on LinkedIn is not the most interesting one. The interesting one sits two paragraphs below the headline.

“Since January 2025, average monthly AI token spend across Ramp customers has increased 13x.”

“One customer discovered $120K in annual AI spend that never appeared on a single provider dashboard.”

The second sentence is the one that should change your Monday if you lead marketing. The 13x figure gets discussed in engineering and finance circles. The $120,000 figure has not been connected yet to the department where AI spend is accumulating fastest: marketing.

That is the thesis. The IQ Source piece for marketing teams living through this growth without the ability to measure it is Marketing Automation. The rest of this post explains why marketing is in worse shape than engineering on this story, what the two raw numbers from this week actually mean, and which three metrics you will need before your next AI stack renewal.

Why marketing is in worse shape than engineering on this

Brian Halligan posted a few days ago that after engineering, marketing is the second-heaviest team using AI inside the companies he sees as an investor. That is probably correct. The difference between the two teams is telemetry, not spend.

Engineering pays for AI in visible places. The Vercel bill goes up. The AWS bill goes up. The Anthropic dashboard shows tokens per project. A mediocre CFO can stare at three screens and understand in fifteen minutes what happened to the quarterly budget.

Marketing pays for AI somewhere else entirely. It pays through a Jasper subscription a PM bought last year. Through Copy.ai that someone approved for a specific campaign. Through Claude Team for the content squad. Through ChatGPT Team for the demand-gen squad. Through Gemini Workspace because the department was already on Google. Through AI plugins inside HubSpot, inside Salesforce, inside Adobe Experience. Through Synthesia for video, Descript for audio, Midjourney for imagery.

Each of those subscriptions bills to a different credit card, on a different cycle, against a different accounting code. None of them sit on the same dashboard. The marketing director has no equivalent to the Anthropic dashboard. The closest thing is the monthly report from the agency partner, who is precisely the actor with the opposing incentive to show total cost.

That is what the Ramp case study really implies. $120,000 a year in AI spend that never appeared on a provider dashboard does not mean the customer is disorganized. It means the architecture of AI spend in marketing is structurally designed to never surface in one place. And that pattern does not correct itself. It worsens while the 13x keeps compounding.

The two raw numbers from this week

The two figures that traveled this week measure different things, and they hit marketing for different reasons.

The 13x from the Ramp report measures spend trajectory. Market average, not pioneers. It is the number a CFO will use to project your next year’s budget. If your marketing AI line is at average, you are being projected as a multiplier. If you are below average, you are probably not tracking all the subscriptions. If you are above average, expect a Monday call.

The 82% from the Entelligence analysis measures efficiency, not trajectory. Brandon Carl shared on X this week data from an analysis covering 2,444 companies, attributed to Aiswarya Sankar, founder of Entelligence.AI. The hard number is that for every $100,000 spent on tokens, only $18,000 ends up reaching a stable production feature. The other 82% goes to AI-generated bugs, rework, and review friction. 44% specifically on bug fixes.

The study measures code. But the pattern transfers to marketing content point by point. The discarded draft before publication is the bug fix. The regenerated prompt because the tone was wrong is the rework. The blog post that shipped but moved no metric is the review friction. Marketing has its own 82% and is not measuring it, which is why it does not appear in any conversation.

Now combine the two. Spend up 13x. Internal efficiency at 18%. If your marketing department is tracking the market average on spend and has no efficiency metric, the configuration you have is the worst possible one: increasingly expensive contracts with zero numbers to justify the next renewal in front of your CFO.

Mark Ajzenstadt, founder of Limestone, posted his raw engineering team costs on X this week: $200 per developer per month versus $500 to $2,000 per person at Uber. The difference, in his words, is the cost-control, model-routing, and margin-analysis layer he designed before deployment. Marketing does not have that layer yet. I covered the engineering side in the token addiction architecture post. The marketing version is more urgent because it starts from a zero-visibility baseline.

What marketing FinOps would look like

FinOps for cloud infrastructure is a mature category. It has vendors, certifications, conferences. FinOps for AI is still nascent. FinOps for marketing specifically does not have a name yet. But the mechanics can be assembled this week with three concrete numbers your team can start tracking without buying a new tool.

Number one: cost per MQL generated by workflow, not by channel. The traditional report says how much Google Ads cost, how much LinkedIn cost, how much the blog cost. That ceiling has been reached. The new number must say how much each specific AI workflow that produces leads is costing. The automated blog post workflow with Claude. The email cadence workflow with Jasper. The personalized outbound workflow with Copy.ai. Each workflow has its own token cost, its own human review time cost, its own MQL yield. Without that granularity, what exists is the aggregate number, which is precisely the one the Ramp customer had when they discovered the $120,000 hiding.

Number two: regeneration rate before publishing. For every piece of content that actually ships (to the blog, the email, the ad, the sales deck), how many drafts were generated and discarded? If your rate is 1 to 1, you are not using AI seriously. If it is 5 to 1, that is normal. If it is 20 to 1, you have an operational problem buried in the subscription line. A healthy editorial content range falls between 3 to 1 and 7 to 1 depending on the piece. The metric matters more than the absolute number because it tells you whether your prompts are tuned or improvised.

Number three: cost per piece published versus piece discarded. A discarded piece also costs tokens. When you add up the total token spend for the month and divide it by the pieces that actually reached the world (not the ones that filled the content team’s Notion), you get the cost per effective piece. Compare that figure to what the same piece would cost from an external agency. The conversation with your CFO changes once the number is calculated, in whichever direction it points.

These three numbers do not require new tooling in most cases. They require logging discipline and an operational owner who watches the three of them every week. The problem is not technology; it is ownership. Most marketing departments in Latin America and the US mid-market do not have a single person specifically accountable for the economics of AI workflows. They have a director watching campaign ROAS, a content manager watching engagement, and no one watching token economics. That is why the Ramp number exists.

What IQ Source does about it

Marketing Automation from IQ Source is not an implementation service. It is an operational service that builds marketing automation workflows with AI and installs from day one the measurement layer that engineering uses seriously but that has not reached marketing yet.

Concretely: every workflow we design ships with cost-per-outcome telemetry, not cost-per-subscription telemetry. It ships with measurable regeneration rate, cost per effective piece versus discarded piece, and an internal dashboard that lets the marketing director produce the three numbers when the CFO asks. Not next quarter. The same week.

Discovery happens through AI Maestro — two months of consulting where we map current marketing processes, identify which workflows are wasting tokens on invisible rework, and deliver an AI Opportunity Score with a Go/No-Go gate. The gate matters because most of the tools a marketing director has on the stack are automatable, but not all of them are profitable to automate at current token pricing. The gate tells you which ones are. The math is comparable to what we covered from a different angle in the where the work moved post, only applied to marketing spend.

The piece almost never discussed openly is the crossover cost. The two options a marketing director currently has are hiring an internal specialist dedicated to AI economics for the department (a profile the market does not produce yet, so building it from scratch is expensive), or subcontracting a partner who already has the measurement system assembled. The question is not whether you can skip it. It is how many more quarters you will accept paying 13x without a single number answering what it is producing for you.

Before you close the week, ask your marketing team one concrete question. When they add up every AI subscription currently active across the department (including the ones nested inside HubSpot, Salesforce, and Adobe), what is the monthly total? If no one on the team can produce that number this week, you are the marketing equivalent of the Ramp customer with $120,000 hiding. And if they can produce it, the next question is how many MQLs you attribute specifically to that spend. If the answer to either question is “let me get back to you,” you know exactly where the problem is.

Measure your marketing team’s AI spend