Cost per task: budgeting model spend for marketing agents
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Marketing agents are moving from pilots into regular use, and their running costs are moving with them from rounding errors to budget lines. During a pilot, model spend is usually small enough to ignore.
Then volume grows, a second and third agent arrive, and someone in finance asks why a new line item keeps rising. Teams that tracked costs from the start can answer that question. Teams that did not are left guessing, and the guessing tends to erode trust in the whole program.
We recommend a simple unit: cost per completed task.
Why cost per task
Model pricing is typically quoted per unit of text processed, often per thousand or per million tokens. That is the right unit for engineers tuning a system. It is the wrong unit for a marketing budget, because a marketer cannot easily judge whether a given volume of tokens was money well spent.
A marketer can judge whether a research brief, a draft nurture email, or a weekly pipeline report is worth a certain amount. Cost per task connects model spend to the work it produces. It also makes comparisons with human effort straightforward, since teams already think about how long a brief or a report takes a person.
The word "completed" matters. Count only tasks that produced a usable output. For example, if an agent attempts 100 briefs and 20 are discarded, the cost of those 20 belongs to the other 80.
What to include
The full cost of a task includes more than the final model call. We suggest counting four components.
- Model calls: every call in the run, including planning, retries, self-checks, and summarization steps.
- Tools and data: paid search, enrichment, scraping, or other per-call services the agent uses.
- Infrastructure: hosting, storage, and orchestration, if they are significant at your volume.
- Human review: reviewer time, priced at a loaded hourly rate, if you want a complete picture.
The gap between a single call and a full run is often large. A task that looks cheap in one call can cost several times more once retries, long contexts, and multi-step checks are added. The only reliable way to see this is to tag every call with a task ID and total them. Our post on observability for marketing agents describes the run histories that make this possible.
A worked example
Consider a hypothetical B2B software team with an account research agent. Every figure below is invented for illustration and does not reflect any provider's pricing.
Suppose each brief takes six model calls: one to plan, three to research, one to draft, and one to check. Assume those calls cost a combined $0.40 per brief. An enrichment lookup adds $0.10, so the automated cost is $0.50 per brief.
Now add the failures. In this example, one in five briefs is discarded and rerun. The team pays for 1.25 runs per usable brief, so the automated cost rises to about $0.63.
Finally, add review. A marketer spends six minutes checking each brief. At a hypothetical loaded rate of $60 an hour, that is $6.00 per brief. The full cost is about $6.63, and review accounts for more than 90% of it.
The lesson is general even though the numbers are not. Model spend is often the smaller part of the true cost. Reducing review time through better quality can save more than switching to a cheaper model.
Ways to reduce cost without losing quality
Most savings come from a short list of changes. We suggest trying them roughly in this order, since the early ones carry the least risk to quality.
- Trim context: smaller inputs cost less, and they often improve focus. See our post on why context windows are a budget.
- Match the model to the task: classification, extraction, and formatting often run well on smaller models. We discuss this in small models, big jobs.
- Cache repeated work: if many tasks share reference material or research, reuse results instead of regenerating them.
- Cap retries: set a maximum number of attempts, and investigate any task that frequently hits the cap.
- Stop loops early: detect agents that repeat steps without progress, and end the run.
Each change should be tested against quality before it goes live. A cheaper configuration that raises the edit rate may cost more in total once review time is counted.
Setting budgets
Once you know cost per task and expected volume, budgeting becomes arithmetic. Multiply the two, add a margin for retries and growth, and set the result as a monthly ceiling per agent.
We recommend alerts at two levels: a daily threshold that catches sudden spikes and a monthly threshold that catches slow drift. A sudden jump often points to a bug, such as a loop or a malformed input that triggers endless retries, rather than real demand. Treat each alert as an investigation, not a reason to raise the ceiling.
For annual planning, cost per task gives finance a model they can check. Our guide to budgeting for 2027 covers how to fold agent costs into a wider program budget.
Put it next to value
Cost alone says nothing about whether an agent is worthwhile. We suggest reporting three numbers together for each agent: cost per completed task, time saved per task, and a quality measure such as human edit rate.
An agent that costs more per task but saves an hour of skilled work may be a far better investment than a cheap one that needs heavy editing. The reverse also happens. A low-cost agent with a high edit rate can quietly shift work onto reviewers while looking efficient on paper.
What changes at scale
At small volume, one person can review costs by hand each month. At larger volume, with several agents sharing tools and context, costs become harder to assign. A shared retrieval step or a supervisor agent serves many tasks at once.
We suggest allocating shared costs by task count, and documenting the rule so finance can audit it. Any allocation rule is imperfect. A stable, written rule is more useful than a precise one that changes every quarter.
We are now studying how cost per task behaves as agents improve. We expect review time to fall faster than model spend, which would shift where savings come from over a program's first year. We are testing how to measure that shift without adding reporting overhead.
How to apply this: for each agent, calculate last month's cost per completed task, including retries, checks, and discarded runs. Report it next to time saved and edit rate, and set daily and monthly alerts before volume grows.
Related research
Shadow mode: run agents alongside humans before going live
The safest way to learn how an agent will perform on real work is to let it do the work in parallel, without its output reaching anyone. That is shadow mode.
Mid-year review: a scorecard for your agentic marketing program
Six months in, leadership will ask whether agents are paying off. This scorecard answers with evidence on value, quality, risk, operations, and readiness.
Synthetic audiences are a thinking tool, not an evidence tool
Asking AI personas to react to messaging is fast and cheap. It can sharpen early thinking, but it cannot replace evidence from real customers.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
