How much an AI agent costs in production: what goes into the math beyond the model

Why the question usually comes out wrong

Comparing agents by model price is like comparing cars by the price of a gallon of gas. Fuel matters, but the total includes maintenance, insurance, who’s driving and how long the car sits in the shop.

Agents are the same. There’s the cost of building, the cost of running, the cost of keeping it working and the cost of the people who stay in the loop. An agent that’s cheap to run and makes lots of mistakes can end up costing more than a pricier one that gets things right.

The number that matters is cost per resolved case, not cost per message or per license.

The components of the real cost

So you don’t miss anything, split the math into six blocks:

• Design and build. Mapping the process, writing rules, building integrations, setting up tests. It’s the cost of getting started, and it’s usually underestimated because the real process has more exceptions than the described one.
• Language model. API usage, billed by the amount of text processed. It varies with volume, context size and the model you choose.
• Infrastructure and tools. Servers, databases, document search, queues, log storage, monitoring. Sometimes bundled into a platform, sometimes separate.
• Integrations. Licenses or usage fees for the APIs of the systems the agent accesses: CRM, ERP, WhatsApp, e-signature. Some charge per call or per message.
• Operations and maintenance. Someone has to track metrics, review mistakes, adjust rules when the process changes and update the agent when a system releases a new version.
• Human review. The time of the people who approve, correct and take over the cases the agent hands back. It’s a real cost, and it’s what lets you trust the agent with important work.

There’s a seventh item that never appears on the invoice: the cost of mistakes. A wrong refund, a wrong answer to a customer, exposed data. It won’t be in a vendor’s spreadsheet, but it should be in yours.

How model pricing works

Models are billed per token, which is a piece of a word. In English, a word averages roughly 1.3 tokens, and other languages often use more. Prices are quoted per million tokens, and output text (the answer) usually costs several times more than input text.

Ranges vary a lot and change fast, so treat these as orders of magnitude. Smaller, faster models cost from cents to a few dollars per million tokens. The most capable models run from a few dollars to a few tens of dollars per million, especially for output. The gap between the cheapest and the most expensive can exceed a hundredfold.

What surprises newcomers is how much goes into each call. An agent doesn’t just send the customer’s question. It sends:

• The agent’s instructions, which can run to several pages.
• The description of every tool it can use.
• The conversation history so far.
• The document snippets or records it retrieved.
• The result of every tool it has already called.

And a case is rarely resolved in a single call. The agent queries a system, reads the result, decides, queries another. Each step is a new call, and the context keeps growing.

How to calculate cost per case

You don’t need to-the-cent precision at the start. You need an honest estimate that you’ll later check against real data. A simple approach:

• Define what a case is. An order handled, an invoice checked, a lead qualified. It should be something the operation already counts.
• Count calls per case. How many steps the agent takes, on average, to resolve a typical case.
• Estimate input and output tokens per call. Add up instructions, tools, history and retrieved data. Output is usually much smaller than input.
• Multiply by the model’s price. Input times input price, output times output price, times the number of calls.
• Add a margin for retries. Network failures, badly formatted responses and retries happen. A 20% to 30% buffer is reasonable at the start.
• Spread fixed costs over volume. Infrastructure, maintenance and monthly licenses divided by the number of cases per month.
• Add human time. The share of cases that go to review, times the average review time, times the hourly cost of the reviewer.

An illustrative example, with round numbers made up just to show the math. Imagine an agent that checks orders in six calls, each with 8,000 input tokens and 500 output tokens. That’s 48,000 input tokens and 3,000 output tokens per case. With a model that charges $3 per million for input and $15 for output, that comes to about $0.14 plus $0.05, or roughly $0.20 per case for the model alone. If 10% of cases go to a person who spends five minutes on each, human time per average case is half a minute. Often it’s that half minute, not the model, that weighs most.

Where the bill usually blows up

The most common surprises come from predictable things:

• Context that grows unchecked. Long conversations and whole documents sent on every call multiply the cost without improving the answer.
• A big model for a simple task. Classifying a message or extracting a field rarely needs the most expensive model.
• Loops. An agent that can’t finish and keeps retrying can burn dozens of calls on a single case. Without a step limit, you only see it on the bill.
• Exceptions that land back on the team. If half the cases end up with a person, the gain disappears and the build cost is still there.
• Forgotten maintenance. Systems change, rules change, models get replaced. Without someone looking after it, quality slips slowly and nobody notices until the complaints start.

How to cut costs without losing quality

Before switching models to save money, look at the design:

• Use different models for different steps: a small one to triage and extract, a larger one only where there’s a hard decision.
• Send the model only the part of the document that matters, not the whole file.
• Use the prompt caching many providers offer, which charges less for the part of the instructions that repeats on every call.
• Summarize the history of long conversations instead of resending everything.
• Set a step limit and a spending limit per case, with an alert when it’s exceeded.
• Use regular code for whatever doesn’t need AI: sums, dates, format validation, fixed rules.

And measure from day one. Cost per case, the share of cases resolved without help and human review time. With those three numbers, you know whether the agent is paying its way.

What to ask before signing any proposal

Whether with a vendor or your internal team, get answers to these questions:

• Is model usage included or billed separately? If it’s included, is there a cap?
• Who pays for integration licenses and fees?
• What’s included in maintenance once the agent is in production?
• How will I track cost per case, and how often?
• What happens to the cost if volume doubles?