A KPMG survey of over 2,100 business leaders surfaced a number worth sitting with: almost half of organizations have rescheduled or reduced AI deployments after costs exceeded expected value, and roughly a third cite weak understanding of AI economics as a real barrier to deploying agents. Separately, nearly 29% of senior executives report struggling to understand and control AI operating costs as providers shift away from flat subscriptions toward usage-based billing.
If your marketing team is piloting AI agents this year, this is the data point that should shape your budgeting process, not an afterthought to it.
Why AI agent costs are genuinely harder to predict than SaaS subscriptions
A traditional martech subscription has a predictable monthly cost. An AI agent’s cost scales with usage: how many tokens it processes, how many tool calls it makes, how many steps a task takes before it completes. A workflow that seems cheap in a demo can become expensive at production volume in ways that are difficult to estimate until you actually run it.
Token counts, total number of agents deployed, and isolated productivity anecdotes are weak substitutes for measuring actual workflow-level business results, and leaning on them is a big part of why so many organizations are getting surprised by their bills.
A more honest budgeting approach
Map the process before you buy the agent. Before adopting any agent platform, document the end-to-end process it will touch, including which steps need which data and which approvals. This surfaces hidden complexity, and hidden cost, before it shows up on an invoice.
Pilot at real volume, not demo volume. A workflow tested on ten examples tells you almost nothing about cost at the hundreds or thousands of executions a live campaign might require. Run your pilot at something closer to real production scale before committing budget.
Measure outcomes, not activity. Track campaign speed, pipeline generated, or retention impact, not the number of tokens processed or agents deployed. Those business outcomes are what will actually justify the spend to leadership, and they are the numbers that expose whether an agent is genuinely valuable or just expensive automation.
Set a spend ceiling with an alert, from day one. Usage-based billing means a runaway workflow can generate an unpleasant surprise fast. A budget alert costs nothing to set up and prevents the exact scenario the KPMG survey describes.
The organizations getting real value from AI agents are not the ones spending the most, they are the ones who built a measurement discipline before they built the automation.
Why “how many steps a task takes” is the variable nobody budgets for
Of the cost drivers listed above, step count is the one most teams underweight when estimating, because it’s the least intuitive. A single agent task isn’t necessarily one model call — an agent working through a multi-step process might make several tool calls, re-read context after each one, and iterate on its own output before returning a final answer, and each of those internal steps consumes tokens even though the person who requested the task only sees the single final result. A task that looks simple from the outside (“summarize this week’s campaign performance”) can involve the agent querying multiple data sources, reconciling conflicting numbers, and drafting several revisions internally before landing on the output you see — and every one of those internal steps is billable. This is exactly why a demo, which is usually run on a clean, well-structured example, chronically underestimates real-world cost: production data is messier, which means more steps to reconcile it, which means more cost per task than the demo ever showed.
What a real pilot-at-volume test actually looks like
“Pilot at real volume” is the right instinct, but it needs a concrete target to be actionable rather than aspirational. Before rolling out an agent workflow broadly, run it against a representative sample of your actual production data — not your cleanest data, your typical data, including the messy CRM records and inconsistent formatting a demo would never include — at a volume close to a real week’s workload, and track the actual cost per execution from your provider’s usage dashboard, not an estimate. That number, multiplied by your expected monthly volume, is the figure that belongs in a budget conversation. Skipping this step and extrapolating from a ten-example demo is the single most common reason the KPMG survey’s numbers exist in the first place.