How Much Does an AI Agent Cost to Build and Run?
What an AI agent costs to build and run, why cost per completed task matters more than token price, and how to find your break-even volume.
By Sameer Rahman- Updated
- 10 min read
On this page
Most AI agent budgets go wrong in the same place. The team compares token prices, picks the cheapest model, and gets a quote for the build. Then the real costs show up, in failed tasks people have to redo, in tool results that bloat every step, and in the hours someone spends checking the agent's work.
So this guide starts with build prices, then gets to what really decides the bill. You'll see how to work out cost per completed task, why a model that's 20 times more expensive can cost half as much, and how to find the volume where an agent pays for itself.
Key takeaways
- The build is mostly engineering time, at about $93 an hour. A one-month proof of concept is about $15,000, and an agent wired into several systems can run $60,000 to $90,000.
- Agent costs don't grow in a straight line. Each step rereads a growing context, so doubling the steps can nearly triple the cost.
- Judge models by cost per completed task. Failures that land on a person often cost more than the model itself.
- Review time decides the payback. One minute of checking on a four-minute task removes a quarter of the saving.
- Long tasks beat frequent ones. A 30-minute task can pay back at about a tenth of the volume of a four-minute one.
How much does an AI agent cost?
An AI agent costs whatever the engineering time to build it costs, plus monthly running costs that range from under $100 to a few thousand dollars. A one-month proof of concept is about $15,000. A production agent connected to several business systems can run $60,000 to $90,000.
An AI agent is a system where the model decides which steps to take and which tools to use until a task is done. It might research a company, sort an inbox, or pull details from documents. If the steps never change, you don't need one, and a workflow with an AI step will be cheaper to build, run, and test.
The easiest way to price a build is in engineering time. The median US software developer earned $135,980 (opens in a new tab) in May 2025, and benefits add about 30% (opens in a new tab) on top of wages in private industry. That's about $93 an hour. Here's what common scopes look like at that rate.
| Scope | Engineering time | At about $93 an hour |
|---|---|---|
| Proof of concept on your own data, read and draft only | About 4 weeks | About $15,000 |
| Production agent for one task, writing to one system | About 8 weeks | About $30,000 |
| Agent across your CRM, documents, and email, with writes | 4 to 6 months | $60,000 to $90,000 |
That's the cost of the time itself, so it also works as a check on any quote you get. Divide the price by $93 and you'll see roughly how many hours you're paying for, and whether that matches the scope.
What really drives the build price
The model is rarely what makes a build expensive. The price mostly follows from three things, and you can reduce all of them before you ask for a quote.
What the agent can write to. Every integration adds work, but they aren't all equal. Reading from a CRM is simple. Writing to it means handling duplicates, retries, permissions, and what happens when the agent gets it wrong. If version one can only read and draft, you can often cut the hardest part of the build.
The state of your data. Cleaning up records and documents can take a big share of the build. It's also the part you can start on yourself, by cleaning up the records and documents the agent will depend on.
How you'll prove it works. A demo needs a few good examples. A production agent needs a test set of real cases, with known right answers, that you rerun every time a prompt or model changes. Quotes that skip this look cheaper, and then you end up paying for it in rework. Ask to see how the test set will be built.
Why agent costs grow faster than you'd expect
A chatbot usually calls the model once per answer. An agent calls it on every step and rereads everything so far, including what its tools returned. So the context gets longer as the task goes on, and you pay for all of it again on each step.
Take an agent that starts with 8,000 tokens of instructions and context, adds 1,500 tokens from a tool on each step, and writes 300 tokens per step. A token is a small chunk of text, roughly part of a word. Here's what that costs at Claude's list prices (opens in a new tab) at the time of writing, which are $0.10 in and $0.50 out per million tokens for Haiku 5.5, and $2 and $10 for Sonnet 5.5.
| Task | Input tokens read in total | Haiku 5.5 | Sonnet 5.5 |
|---|---|---|---|
| 6 steps | 70,500 | About $0.008 | About $0.16 |
| 12 steps | 195,000 | About $0.02 | About $0.43 |
Twice the steps costs about 2.7 times as much. That gives you two levers most budgets ignore.
- Shrink what tools return. If a CRM lookup returns 200 fields and the agent needs five, have the integration send five. That's a design decision, and it lowers the cost of every later step.
- Use prompt caching. On Sonnet 5.5, a cache hit costs $0.10 per million tokens, against $2 for fresh input. Writing to the cache costs a bit more, at $2.50. In the six-step example, that takes the task from about $0.16 to about $0.06, a drop of around 60%, as long as the steps run close together.
Why the cheapest model is often the expensive one
Token price is the wrong number to compare. What you're buying is a finished task, and every task the agent can't finish goes to a person.
Say each failed task takes someone four minutes, at a staff cost of $35 an hour. That's about $2.33 a failure. Now suppose that in your testing, Haiku 5.5 finishes 70% of tasks and Sonnet 5.5 finishes 92%. Measure yours on real work.
| Haiku 5.5 | Sonnet 5.5 | |
|---|---|---|
| Model cost per task | $0.008 | $0.16 |
| Share of tasks that fail | 30% | 8% |
| Cost of a person fixing those, per task | $0.70 | $0.19 |
| Cost per task, all in | About $0.71 | About $0.35 |
| At 3,000 tasks a month | About $2,120 | About $1,040 |
The model that costs 20 times more comes out at half the price. The result flips if your tasks are simple enough for Haiku to finish nearly all of them, which is why it's worth testing both on real work. And you don't have to pick just one. Many agents send the routine steps to a small model and save the larger one for the hard decisions.
Mistakes change the math further. A failed task that gets flagged costs a few minutes. A wrong refund or a bad email to a client that nobody catches can cost far more. So the higher the cost of a mistake, the less the model's price should matter.
How to find your break-even volume
Before you ask anyone for a quote, work out how many tasks a month the agent needs to handle. It only takes one calculation.
Fixed cost is the build spread over the months you expect to use it, plus upkeep. Here's an example with a $30,000 build spread over 24 months and $1,500 a month for upkeep, so $2,750 a month. The agent handles 60% of tasks, staff time costs $35 an hour, and each task costs $0.16 in model fees.
| Task | Minutes saved when the agent handles it | Break-even volume |
|---|---|---|
| 4-minute task, no human review | 4 | About 2,200 tasks a month |
| 4-minute task, 1 minute of review | 3 | About 3,100 tasks a month |
| 30-minute task, 5 minutes of review | 25 | About 320 tasks a month |
Two things stand out.
Review time is the biggest hidden cost. One minute of checking on a four-minute task takes a quarter of the saving and pushes break-even up by 40%. So the plan to reduce review matters as much as the agent. Most teams should start by approving everything, then check a sample once results hold up on a given type of task.
Long tasks are usually the better first target. Teams tend to pick whatever happens most often, like inbox triage. But research, document review, and preparing reports can take 30 minutes or more each, and they pay back at a fraction of the volume. Rank your candidates by minutes per task times tasks per month, not by how often they happen.
Only count hours that really change something. If an agent saves everyone 10 minutes a day and nothing else changes, your costs stay the same. Count time that comes off a real workload, or lets you take on work you'd otherwise turn down.
What to include in your monthly budget
Model fees are usually the smallest line once an agent is live. The bigger ones are the hours people spend checking and tuning it, so that's where to plan.
| Cost | What it covers | How to estimate it |
|---|---|---|
| Model usage | Everything the agent reads and writes | Cost per task × tasks a month. $0.16 × 3,000 is $480 on Sonnet 5.5 |
| Search and data tools | Looking up documents, records, or the web | Price per call × calls per task × tasks a month |
| Monitoring | Logs of each step, and alerts when something breaks | Usually part of the platform or the upkeep fee |
| Tuning | Reviewing results and fixing prompts | Hours a month × hourly cost. 10 hours at $93 is about $930 |
| Human review | People checking the agent's work | Minutes per task × tasks × staff cost per minute |
The items to watch are tuning and review, because they don't drop to zero, and search or data tools, because they bill on every call and an agent can call them more than you planned.
Common mistakes when budgeting for an AI agent
- Estimating from demo tasks. Demo tasks finish in a few steps. Real ones hit missing data and retries, and because of how context grows, the extra steps cost more than you'd think. Measure steps on real work.
- Leaving out a spending limit. An agent stuck in a loop keeps paying to reread its context. Cap the steps per task and the monthly spend, and send anything over the limit to a person.
- Comparing quotes with different test plans. A quote with no test set isn't really cheaper. It just moves that work to after launch.
- Giving it full write access on day one. This is the most expensive version to build and test. Start with drafts, then add actions one at a time.
- Not tracking cost per task. Without it, a prompt change that doubles your costs only shows up on next month's invoice.
Conclusion
The token price tells you very little about what an AI agent will cost. What decides it is how many tasks it finishes, how much a person has to check, and how many tasks it gets each month. Work out your break-even volume first, start with a long task, and test more than one model on real work.
We build AI agents in your own accounts, using tools like n8n with Claude or ChatGPT, with limits on what they can do and logs of what they did. Every project is scoped and quoted after a discovery call, around the task with the best payback. If you're still comparing builders, see the best AI agent development agencies.
Frequently asked questions
Most of the price is engineering time. At about $93 an hour, the US median developer cost with benefits, a one-month proof of concept is about $15,000 and a production agent connected to several systems can run $60,000 to $90,000. The integrations it can write to move the price most.
Model fees can be under $100 or several thousand dollars, depending on steps per task and how much each step reads. Add monitoring, tuning, and the time people spend reviewing its work, which is usually the biggest part.
An agent rereads its full context on every step, and that context grows as tools return results. Doubling the steps can nearly triple the cost, so limiting tool output is one of the best ways to cut the bill.
No. If a cheaper model finishes fewer tasks, people have to pick up the rest, and that usually costs far more than the model. Compare cost per completed task, including failures.
Divide your monthly fixed costs by what each task saves after model fees. In our example, a $30,000 build needs about 2,200 four-minute tasks a month, or about 320 thirty-minute tasks with review.
Sources
- Anthropic, Claude pricing (opens in a new tab), read November 7, 2026.
- U.S. Bureau of Labor Statistics, Software developers, quality assurance analysts, and testers (opens in a new tab), Occupational Outlook Handbook, read November 10, 2026.
- U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026 (opens in a new tab), released September 9, 2026, read November 10, 2026.