The cost of running AI agents for a marketing team: a model you fill in with your own numbers
The monthly cost of running AI agents for a marketing team is the sum of four lines: model usage (tokens and tool calls), software and connectors, the people time to review and maintain the agents, and amortized setup. Model usage is often the smallest line; review time is usually the largest, so the most reliable way to estimate cost is to count runs and review minutes per agent, not to start from a price per token.
Key takeaways
- Cost per agent per month = runs x cost per run + tools + review hours x loaded rate + setup amortization.
- Token cost per run depends on input size, output size and model choice; measure it from a pilot rather than guessing.
- Batch processing and prompt caching cut model cost for scheduled, repetitive agents.
- Review and maintenance time decides whether the agent pays back.
Why most cost estimates are wrong
Agent cost conversations usually start with a price per million tokens, which is the easiest number to find and the least useful on its own. A marketing agent's cost depends on how often it runs, how much it reads each time, how many tools it calls, which model it uses, and how long a person spends checking its output. Two pacing agents on the same model can differ in cost by an order of magnitude depending on whether they read a summary table or every ad's daily history.
Cost is also the first reason Gartner gives for agentic projects failing: it predicted that over 40% of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. A cost model built before launch, and checked against actual usage after a month, prevents the first of those.
The four cost lines
Every agent has the same four lines. Fill them in per agent, then add them up.
| Line | What it includes | How to estimate it |
|---|---|---|
| 1. Model usage | Input and output tokens per run, tool call overhead, paid tools such as web search | Run the agent 10 to 20 times in a pilot and read the usage report; multiply average cost per run by runs per month |
| 2. Software and connectors | Model plan or API account, workflow or orchestration tool, connectors to CRM and ad platforms, logging | List prices for the seats and tiers you need; allocate shared tools across agents |
| 3. People time | Reviewing outputs, approving changes, fixing errors, updating instructions and the knowledge base | Review minutes per run x runs per month, plus maintenance hours per month, x loaded hourly rate |
| 4. Setup, amortized | Design, build, testing, permissions and documentation | One-time cost divided by the months you expect the agent to run before a rebuild |
The formula per agent is simple: monthly cost = (runs per month x model cost per run) + software share + (review and maintenance hours x loaded rate) + (setup cost / months of use). Our AI agent ROI calculator applies the same structure to the value side.
We set up shared knowledge, agents and QA so your team gets consistent output.
A fill-in worksheet
Copy this table for each agent. The rows describe what to measure, not what your numbers will be.
| Input | Where the number comes from | Your value |
|---|---|---|
| Runs per month | Schedule (daily runs on working days or calendar days) or event volume (leads, replies, pages) | ... |
| Average input tokens per run | Pilot usage report; driven by how much data and instruction the agent reads | ... |
| Average output tokens per run | Pilot usage report; driven by report or draft length | ... |
| Model price per million input and output tokens | Your provider's current pricing page for the model you chose | ... |
| Paid tool calls per run | For example web searches, priced per 1,000 by some providers | ... |
| Software share | Seats, orchestration and connectors allocated to this agent | ... |
| Review minutes per run | Timed during the pilot; falls as trust grows | ... |
| Maintenance hours per month | Instruction updates, knowledge base changes, fixing broken connectors | ... |
| Loaded hourly rate of the reviewer | Salary plus benefits and overhead, from finance | ... |
| Setup cost and expected months of use | Internal hours or partner fee; months before a planned rebuild | ... |
Two facts from one provider show why the pilot matters. Anthropic's pricing documentation lists a 50 percent discount on both input and output tokens for its Batch API, and prices web search on the Claude API at $10 per 1,000 searches plus standard token costs. Check your own provider's current page; prices and discounts change.
Levers that cut the bill
Once the first month of usage is in, these are the changes that move cost most, roughly in the order we try them.
- Send less: give the agent a summary table instead of raw exports, and only the fields it needs.
- Batch scheduled work: overnight reports, weekly audits and bulk enrichment rarely need an instant answer, and batch processing is discounted by some providers.
- Cache the stable part: instructions, brand rules and the knowledge base change rarely; prompt caching charges cache reads at a fraction of the base input price on providers that offer it.
- Match the model to the step: a smaller model can classify replies or check formats; save the largest model for writing and reasoning.
- Cut review time with structure: a fixed output format and an automatic claims check let reviewers scan instead of reading every word.
- Retire agents nobody reads: if the output of an agent has not changed a decision in a month, switch it off.
Where the money really goes
In the agent programs we set up, the review and maintenance line is the one to watch. A daily agent that saves a specialist 30 minutes but needs 20 minutes of checking is not worth much, whatever its token bill. The fix is rarely a cheaper model; it is a narrower job, a fixed output format and a claims check that makes review fast.
Compare the total with the work it replaces, not with zero. Our business case for AI agents in marketing explains how to value hours returned and speed to lead.
Our guide to the AI agent experiment budget for 2027 shows how to gate spend until the numbers are proven.
Our Claude implementation team builds the cost model with you before launch and reviews it after the first month of real usage.
If you want to see what your first three agents would cost to run, talk to our team.
Frequently asked questions
How much does it cost to run an AI agent for marketing?
It depends on runs per month, tokens per run, model choice, tools and review time. Estimate it per agent: runs x cost per run, plus software share, plus review and maintenance hours x loaded rate, plus setup divided by months of use. Measure cost per run in a short pilot.
What is usually the biggest cost of marketing AI agents?
People time to review outputs and maintain instructions and connectors. Model usage is often smaller, especially for scheduled agents that read summarized data.
How can we reduce token costs for AI agents?
Send less data per run, batch scheduled work, cache stable instructions and knowledge where your provider supports it, and use smaller models for simple steps such as classification.
Do AI agent costs grow with usage?
Model usage grows with runs and data volume, so event-driven agents (per lead or per reply) scale with volume. Review time should fall as the agent proves reliable, which is why we track both monthly.
Should we budget AI agents as software or as headcount?
Budget model usage and tools as software, and review and maintenance as part of the team's time. Showing both lines keeps the business case honest.
How long should an AI agent pilot run before we trust the cost estimate?
Long enough to cover the agent's normal cycle: two to four weeks for a daily agent, and at least two runs for a weekly or monthly one. Read the provider's usage report and time the reviews during the pilot, then recheck after the first full month.
Sources
Related service: Claude implementation. See also pricing and client results.
We run enterprise B2B demand generation for teams in Montreal, Toronto, Ottawa and the United States.
