Your AI Marketing Agent Needs a Compute Budget, Not Just an Ad Budget

A marketing team can approve an advertising budget and still leave the cost of its AI agent undefined. The media spend has a ceiling. The software preparing audiences, researching offers and revising campaign copy may not. Those are different budgets, and they deserve different controls.

September's agent infrastructure announcements make this more than a theoretical concern. My view is that marketing leaders should treat an autonomous workflow as a running service with an owner and a spending boundary, not simply as another software seat.

The cloud bill is becoming part of the campaign brief

DigitalOcean's release notes record that its Managed Agents platform entered public preview on 21 September 2026. It brings an agent's execution environment, model access and persistent data together. This is an infrastructure announcement, not proof that a marketing department will save money, but it illustrates the shift from buying an assistant to operating ongoing work.

The billing detail deserves as much attention as the launch headline. DigitalOcean's Harness Runtime pricing page lists compute, storage and other charges separately. It also says there is no per-session or per-product spending limit: usage draws on a shared prepaid balance. At the time of writing, its footnote labels actual active-CPU billing as coming soon and specifies an interim allocation-based charging rule. Buyers should check the implemented terms, not assume that a headline description settles every line of the bill.

Elsewhere, Geordie announced general availability of its Cost Intelligence capability on 9 September. The vendor describes it as connecting expenditure to the agents and activities that generated it. That is a supplier claim about visibility, not evidence of a particular return on investment. Still, the buying question is useful: can we tell which campaign task consumed the money?

A small request need not mean a small workload

Forcepoint's 21 September analysis of unbounded consumption describes how agent workflows can accumulate resource usage without an enforced stopping point. Its examples include accidental misconfiguration as well as malicious activity. The article includes a simulation, not a measurement of typical marketing costs, so its demonstration should not become a forecast for every campaign.

For a marketing example, imagine asking an agent to prepare a seasonal email and SMS campaign. The task might include researching competing offers, proposing audience groups, drafting alternatives and checking a landing page. Now imagine the brief also says to keep improving the work. Without an agreed endpoint, more activity can look like diligence even when the team already has enough material to make a decision.

I would distinguish the useful expansion from the unnecessary one before the work begins. Three genuinely different creative routes may answer the brief. Repeatedly polishing the same route may not. That judgement belongs to the campaign owner, rather than being left entirely to the agent.

Cheap intelligence is not a campaign business case

In August, Gartner forecast that inference costs per agentic workflow would rise more than fivefold through 2028 as workloads became more complex. That is a forecast, not an observed result for marketing teams. Its more relevant warning is that improving unit prices do not, on their own, establish the economics of the whole job.

My proposed measure for a pilot is cost per accepted campaign deliverable. Record the running costs of the task alongside the human time needed to review and repair it. Keep message-delivery fees and media spend visible as separate items. Otherwise, a cheap draft can appear economical while the effort required to make it usable disappears from the assessment.

Accepted should mean something specific: an audience proposal the campaign owner can use, a checked email draft, or an SMS variant that meets the agreed brief. It should not mean that the agent produced a large folder of alternatives. Volume is an especially poor purchasing criterion when the team must inspect everything it receives.

Give the service an owner and an expiry date

The server-management lesson is not that marketers should administer machines themselves. It is that someone must own the operational consequences of leaving an agent running, whether the infrastructure is managed by a supplier or an internal team.

For a first pilot, I would put five decisions in the brief:

• Define the permitted workload. Specify the research scope, deliverables and deadline. Ask for a decision-ready output rather than unlimited improvement.
• Agree an enforceable limit. Have the technical owner confirm what stops a run when its allowance is exhausted. An emailed warning and a hard stop are not interchangeable.
• Separate campaigns in the cost record. Require enough attribution to distinguish a useful research task from an abandoned experiment. A shared account balance is not a campaign-level budget.
• Decide what survives the campaign. Ask which sessions, stored workspaces and recurring tasks remain chargeable afterwards, who closes them, and what must be retained for review.
• Name the person who can authorise more work. When the agent reaches its boundary, the campaign owner should decide whether another research pass is worth buying.

These are operating choices, not an argument against automation. A longer, more expensive task may be justified when the resulting insight is valuable. The point is to make that decision deliberately, rather than discover it on an invoice.

Buy useful work, not continuous activity

The opportunity for AI in marketing is not an endlessly busy assistant. It is a better campaign prepared with an amount of time, attention and money the business understands.

Before approving an always-on agent, ask a less glamorous question than what it can do: what will make it stop? A clear answer belongs beside the audience, offer and media budget in the campaign brief.