The news
Axios and The Wall Street Journal report that some enterprises have burned through their entire annual AI budget in just a few months, with others watching spend double or triple with little warning. The culprit isn't reckless experimentation — it's agentic AI, where a single task triggers dozens or hundreds of model calls instead of one. Goldman Sachs projects token consumption will multiply 24x between 2026 and 2030.
Our take
This is a structural problem, not a discipline problem, and the distinction matters.
Most marketing AI budgets were set for chatbot-era usage: a prompt in, an answer out. Agentic workflows don't work that way. When you ask an agent to draft a content brief, research a prospect list, or pull campaign performance, it doesn't make one API call — it chains calls, checks its own output, backtracks, retries. The token meter is running the whole time. Nobody budgeted for that because most teams don't have visibility into it until the invoice arrives.
The visibility gap is the real problem here. Most marketers have no idea what their AI workflows actually cost per run. They picked a tool, got a flat monthly seat license, and assumed the economics were fixed. But agentic layers — whether that's a purpose-built agent platform, an AI feature in HubSpot, or a Zapier workflow calling GPT-4o on every new contact — don't price like SaaS. They price like cloud compute. Usage spikes, costs spike.
There's also a compounding dynamic that's easy to miss. Workflow automation compounds — that's the point. Small automations chain into bigger ones. That's good for productivity and genuinely dangerous for unmonitored compute spend. A single content workflow running five times a day looks fine. The same logic replicated across campaigns, markets, and personas suddenly doesn't.
The teams that avoid budget shock aren't the ones using less AI. They're the ones who instrumented their usage before they scaled it — who know which workflows are running, how often, and roughly what each run costs.
So now what?
- Audit your agentic surface area. List every place your team is using AI that does more than a single generation — any workflow with multiple steps, loops, or tool calls. That's your cost exposure.
- Pressure-test your pricing model. If a tool is charging per-seat and hiding usage-based costs in "credits" or API passthroughs, find out what happens when volume doubles.
- Set a usage baseline before you scale. Before expanding any agentic workflow to more records, more campaigns, or more users, run it at small scale and clock the cost per execution.
The teams getting surprised by AI bills weren't using AI recklessly — they were just scaling faster than their visibility.