The two costs of an AI agent
Building an AI agent is a one-time cost. Running it is a monthly bill, and unlike ordinary software, the bill grows with every request the agent handles. Understanding that bill before you build is the difference between an agent that pays for itself and one that quietly does not.
This article works through the running cost of a typical customer support agent. The numbers are illustrative, based on the kind of usage we see in real projects; model prices change often, so treat the method as the point, not the exact figures.
A worked example: a support agent
Assume a helpdesk that receives 5,000 tickets a month, and an agent that reads each ticket, searches the help articles, and drafts a reply.
Each ticket involves roughly:
- Reading and context. The ticket, the customer's history and the relevant help articles come to around 4,000 tokens of input. A token is roughly three quarters of a word; models charge per million tokens.
- The reply. Around 400 tokens of output. Output tokens usually cost several times more than input tokens.
- Retries and checks. Some requests run twice: a quality check, a retry after a timeout, a second search when the first found nothing. A realistic overhead is 30 percent.
At an illustrative mid-tier model price of $3 per million input tokens and $15 per million output tokens, the arithmetic looks like this:
| Item | Monthly amount |
|---|---|
| Input: 5,000 tickets x 4,000 tokens x 1.3 overhead | 26 million tokens, about $78 |
| Output: 5,000 tickets x 400 tokens x 1.3 overhead | 2.6 million tokens, about $39 |
| Search index and hosting for the agent service | $50 to $150 |
| Monitoring and logging | $20 to $50 |
| Total | roughly $190 to $320 a month |
Set that against what the agent replaces. If it resolves even half of those 5,000 tickets without a person, and a support agent handles perhaps 400 tickets a month, that is several people's worth of routine work for a few hundred dollars. That is the shape of a good agent project. If your volume is 200 tickets a month, the same maths says do not build it, and we would tell you so.
What people forget to count
- The human hand-off is a feature, not a failure. The 20 to 40 percent of cases the agent passes to a person still cost people time. Count the agent's value on the cases it finishes, not the cases it touches.
- Peak load. Model providers limit requests per minute. A marketing campaign that triples your tickets for a week needs queueing, or the agent falls over exactly when you need it.
- Quality checks cost tokens too. A second model call that reviews the first answer can double your bill and is often worth it. Decide with numbers, not by default.
Five ways to cut the bill
- Use a smaller model for the easy steps. Classifying a ticket does not need the model that writes the reply. Routing simple steps to a cheap model routinely cuts costs by half.
- Cache what repeats. The same product question arrives hundreds of times. Serving a checked answer from cache costs nearly nothing.
- Send less context. Better search over your help articles means fewer tokens per request. Retrieval quality is a cost lever, not just a quality lever.
- Batch the background work. Anything that does not need an instant answer can run in cheaper batch modes overnight.
- Measure per resolved case. The metric that matters is cost per case the agent finished correctly. Optimise that, not cost per request.
How we estimate this before you commit
Every agent project at AppWizards starts with a prototype on your real data, scored against a test set of real examples. The prototype also produces a real usage profile, which is what we use to project the monthly bill. The decision to go to production is made with both numbers on the table: how often the agent is right, and what it costs when it is.
Read more about how we build these on our AI agent development page.