Somewhere in the next few weeks, a renewal quote is going to land on a marketing operations desk and the number will not make sense. The seat count is flat. The contract term is identical. And the total is meaningfully higher, with a new section of the pricing schedule describing units nobody on the marketing side has ever tracked—credits, tasks, actions, generations, agent runs, tokens.
This is the year that stops being an anomaly. Enterprise martech contracts overwhelmingly renew at calendar year-end, which means the negotiation window opens now, in late August and September. And it opens on a stack whose economics changed underneath marketing while the team was busy figuring out what the tools could do.
The shift is straightforward to describe and awkward to budget for. Between roughly 2024 and 2025, marketing platform vendors competed on AI capability by giving it away—generative features bundled into existing tiers, generous credit allotments, pilot programs at no incremental cost, “included during early access” language in the order form. That was a land-grab phase, and it worked: teams built AI features into daily workflows without ever pricing them. Now those features are load-bearing, usage has compounded, and the subsidy is being withdrawn on schedule. What was a differentiator is becoming a line item with a meter attached.
If this sounds familiar, it should. Infrastructure teams went through exactly this transition with cloud computing, and the lesson they learned expensively is worth borrowing for free: the problem with consumption pricing is not that it is more expensive. Often it is genuinely cheaper. The problem is that it converts a predictable fixed cost into a variable one, and organizations that cannot measure the variable lose the ability to forecast, to attribute, and to negotiate.
Why Marketing Is Structurally Unprepared for This
Marketing has bought software on a per-seat basis for twenty years. That model has one enormous virtue: the cost driver is a number the organization already knows. You can forecast next year’s platform spend from next year’s headcount plan, and the finance partner can check your arithmetic.
Consumption pricing severs that link, and it severs it in a particular direction that catches marketing teams off guard.
The cost driver is no longer headcount—it is activity. A team of the same size producing three times the campaign volume consumes three times the units. Which means the more successful your AI adoption, the higher your bill, with no natural ceiling supplied by hiring. Every efficiency gain you booked has a consumption cost on the other side of the ledger, and only one of those numbers has been showing up in your reporting.
The unit is defined by the vendor and is rarely comparable. One platform’s “credit” is a generation request. Another’s is a weighted composite where a long document costs more than a short one. A third meters “actions” that include internal steps a user never sees. These definitions are not standardized, they are not always documented in the order form, and several of them changed at least once in the past eighteen months. You cannot benchmark a price per unit when the unit is proprietary.
Consumption stacks across the stack. A single workflow—research a target account, draft a sequence, personalize by segment, publish, summarize the result—can touch four vendors that each meter a slice of it. Nobody owns the aggregate. The CRM bill, the automation platform bill, the content tool bill, and the analytics assistant bill are reviewed by different people at different times, and the workflow that generated all four is reviewed by nobody.
And the most important one: agents consume without being asked. This is what makes 2026 different from a straightforward SaaS repricing. When consumption required a human to click a button, usage was bounded by the number of humans and the hours in their day. Automated and agentic execution removes that bound. A monitoring agent that polls hourly, a research agent that expands its own task list, a co-pilot that retrieves context on every query—these generate consumption continuously, at machine cadence, and the cost curve no longer resembles anything in your headcount plan.
That last point connects to something we examined earlier this month. If your AI tools are querying a data layer to answer questions, every query has a price now. Teams that solved for governed access to marketing data are about to discover that governed access also needs a budget, and that an agent configured to be thorough is an agent configured to be expensive.
The Forecasting Problem Comes First
The instinct when a metered quote arrives is to start negotiating. That is the wrong first move, because you are negotiating a rate on a volume you cannot estimate, against a counterparty that knows your volume precisely.
Your vendor has your consumption telemetry. They know your monthly credit burn, its growth rate, which teams drive it, and which workflows are heaviest. If you walk into the conversation without that data, the entire discussion happens on their terms—and the standard outcome is that you buy a credit bundle sized by the vendor’s forecast, which will be sized generously.
So the sequence is: measure, model, then negotiate.
Ask for your own usage history in writing, by month, by unit type, before you discuss price. Twelve to eighteen months if they have it. This is a reasonable request and a revealing one—a vendor unwilling or unable to produce granular consumption data for your own account is telling you something about how well you will be able to manage cost after signing.
Find the growth rate and the drivers. You are looking for two things: the trend line, and whether it is driven by more people using the tool or by the same people doing more per person. Those imply completely different forecasts. Seat-driven growth flattens when hiring flattens. Activity-driven growth does not.
Identify what a marginal unit buys. For each significant consumption category, establish what the organization actually gets from it. Not ROI theater—just a plain mapping from units to output. This is what lets you distinguish between consumption that produces work you would pay a person to do and consumption that produces drafts nobody opened. Both show up identically on the invoice.
Assume automated consumption grows faster than you expect. If any part of your 2027 plan involves agents executing on a schedule, model that as its own line with its own growth assumption rather than folding it into a general uplift. It behaves differently, and it is the piece most likely to surprise you.
What to Negotiate For
Once you can forecast, the negotiation has more useful targets than the headline rate.
Rate protection over volume commitment. Vendors prefer to sell large prepaid bundles because unused units are pure margin. What you want is a locked unit price—ideally with tiered rates that improve as volume grows—on a smaller commitment, plus the right to add units at the same rate mid-term. Overcommitting to a bundle in an environment where model costs continue to fall is a bet against your own leverage next year.
A contractual definition of the unit, and notice before it changes. Insist that the metering definition be written into the agreement, and that material changes to how units are calculated require advance notice and give you a right to renegotiate. Silent redefinition of a consumption unit is a price increase that never appears as one.
Overage terms you can survive. Find out exactly what happens at the threshold: does the platform stop, throttle, or continue billing at a punitive rate? All three exist in the market and the difference matters enormously in the middle of a quarter-end campaign. Negotiate the overage rate, not just the included volume.
Rollover and pooling. Unused units carrying forward, and credits poolable across teams or products rather than stranded in silos, are both routinely available and rarely requested.
Visibility as a deliverable. Real-time consumption dashboards, per-user and per-workflow attribution, configurable spend alerts, and API access to your own usage data. Ask for these explicitly and get them in the order form. “You can see it in the admin panel” is not the same as “we will notify you at 80 percent.”
A price for turning it off. Understand what the platform costs without the AI features, and what happens to workflows built on them if you downgrade. If the honest answer is that your operation no longer functions without metered capability, that is worth knowing before you sign—it is the actual measure of your negotiating position, and it argues for locking multi-year rate protection now rather than discovering the dependency at the next renewal.
Running a Metered Stack
Signing a good contract is half of it. The other half is operating under a model where marketing activity generates variable cost in real time, which is genuinely new for most teams.
The good news is that you do not need to build a FinOps function. You need three habits.
One person owns the aggregate number. Not per-vendor ownership—someone accountable for total AI consumption spend across the stack, reviewed monthly, with authority to investigate anomalies. This belongs in marketing operations, because that is where the workflows driving consumption are configured. Splitting it across tool owners guarantees that the stacked-workflow problem stays invisible.
Alerts and circuit breakers on anything automated. Every scheduled or agentic process gets a consumption ceiling and an alert well before it. A misconfigured loop, a retry storm, or an agent that expands its own task list should hit a limit rather than an invoice. Infrastructure teams learned this from runaway compute jobs; the marketing version is a monitoring agent that quietly polls a thousand accounts hourly.
A cost-per-outcome metric that leadership sees. Cost per qualified opportunity, per published asset, per campaign, whatever fits your model—but computed with AI consumption included. This is what keeps the conversation strategic instead of defensive. It also arms you for the argument we walked through in July: if finance is going to ask why marketing needs more money when AI made the work cheaper, the strongest answer is a defensible unit cost that has been trending in the right direction with the consumption cost visible inside it.
One caution on optimization: the cheapest configuration is not the goal. It is entirely possible to tune consumption down by making tools less thorough, less current, and less accurate, and to book that as a saving while quietly degrading the output. Consumption reduction should be evaluated against quality the same way any other efficiency measure is.
The Honest Summary
Consumption pricing for AI capability is not a trick, and resisting it on principle is a losing position. It reflects a real cost structure—inference is a marginal cost in a way that seats never were—and for teams with uneven usage it is often cheaper than the bundled alternative would have been.
What it does is transfer a forecasting burden onto the buyer, at a moment when marketing is simultaneously increasing the share of its work that machines initiate. Those two things compound. A variable cost model plus autonomous execution equals a budget line that moves on its own unless somebody instruments it.
The teams that will handle this well are not the ones who negotiate hardest. They are the ones who walk into the renewal already knowing their consumption curve, its drivers, and what a marginal unit buys them—because that is the only position from which the rate conversation is symmetric.
The window for building that knowledge is the next few weeks, before the quotes turn into signatures and the assumptions inside them become next year’s plan.