How Much Does It Cost to Build an AI Agent?
Sergio JimenezOctober 21, 20266 min read
An AI agent costs between $20,000 and $150,000 to build, and unlike ordinary software it also costs money every time someone uses it. That second number is the one teams forget, and it is the one that decides whether the thing survives its first month in production.
Here is how both numbers are built.
The short answer
| Type | Build cost | Timeline | What it does |
|---|---|---|---|
| Assistant over your docs | $20,000 to $45,000 | 4 to 8 weeks | Answers questions from your own content. Retrieval, no actions |
| Agent that takes actions | $45,000 to $90,000 | 8 to 16 weeks | Calls your systems, writes data, needs approvals and an audit trail |
| Multi agent workflow | $90,000 to $150,000 | 16 to 28 weeks | Several specialised agents, handoffs, evaluation harness, human review |
The jump between the first and second row is the largest in the table and it is not about the model. It is about everything that has to exist before you let software act on your behalf.
Why taking actions costs so much more than answering
An assistant that answers a question wrongly is embarrassing. An agent that acts wrongly issues a refund, emails a customer or deletes a record. The engineering difference is not the prompt. It is permissions, idempotency, approval gates, rollback and a log that shows exactly what happened and why.
Concretely, a read only assistant needs retrieval, a prompt and a chat surface. An acting agent needs all of that plus a tool layer with real authorisation, a way to replay decisions, a human approval path for anything irreversible, and tests that prove the guardrails hold. That is where the second $40,000 goes.
The running cost nobody models
Traditional software has roughly fixed cost per user. An agent does not: every conversation burns tokens, and a badly built one burns many times more than it needs to.
| Driver | Effect on the bill |
|---|---|
| Context you resend every turn | Largest single factor. A 40,000 token system prompt sent 50 times a day per user is the whole budget |
| Model choice | A small model on routine turns and a large one only where reasoning is needed typically cuts spend 60 to 80 percent |
| Retries and loops | An agent that retries a failing tool call ten times costs ten times as much and still fails |
| Caching | Prompt caching on a stable system prompt is the cheapest large win available |
A realistic figure for a working internal agent is $200 to $2,000 a month at modest usage, and it scales with conversations rather than seats. Model this before you build, because a unit economics problem discovered after launch is a rebuild, not a tuning exercise.
What actually drives the build cost
How many systems it touches
One tool is a week. Eight tools across five systems, each with its own authentication, rate limits and failure modes, is most of the project. The model is rarely the hard part; the integrations are.
Whether your data is ready
Retrieval quality is set by the content, not the embedding model. If your documents are current, structured and deduplicated, retrieval works almost immediately. If they are five years of overlapping PDFs with three contradictory versions of the same policy, you are paying for a data cleanup and calling it an AI project. Budget for it honestly and it is fine. Pretend it away and the agent will confidently cite the wrong policy.
How wrong it is allowed to be
"Usually right" is cheap. "Right often enough to act without review" is expensive, because proving it requires an evaluation set, a scoring harness and iteration against measured results. The gap between a convincing demo and a system you would let near a customer is most of the budget, and the demo is the part that takes two weeks.
Whether a human stays in the loop
Counterintuitively, keeping a person in the loop makes the first version cheaper and better. You ship sooner, the failure modes are contained, and every correction becomes training data for the evaluation set. Agents designed to be fully autonomous from day one tend to spend months failing to earn that trust.
What we would actually build first
Narrow it until it is boring. One task, one kind of user, one system of record, with a human approving the output. That version costs $20,000 to $35,000, ships in six weeks, and tells you more about whether the idea works than a year of design would.
Then widen it on evidence: more tools once the first ones are reliable, more autonomy once the evaluation scores justify it. This is how our Studio engagements run on AI work, because the alternative, a large autonomous system specified up front, is the most reliable way we know to spend six figures on a demo.
Frequently asked questions
How much does it cost to build an AI agent?
Between $20,000 and $150,000 depending on what it is allowed to do. An assistant that answers questions from your documents runs $20,000 to $45,000. An agent that takes actions in your systems runs $45,000 to $90,000. A multi agent workflow with evaluation and human review runs $90,000 to $150,000.
What does an AI agent cost to run each month?
Typically $200 to $2,000 a month at modest usage, scaling with conversations rather than with seats. The dominant factor is how much context is resent on every turn, followed by model choice. Prompt caching and routing routine turns to a smaller model usually cut the bill by 60 to 80 percent.
How long does it take to build an AI agent?
Four to eight weeks for a retrieval assistant, eight to sixteen weeks for an agent that acts in your systems, and sixteen to twenty eight weeks for a multi agent workflow. A convincing demo takes about two weeks, which is why timelines get underestimated so consistently.
Why is my AI agent so expensive to run?
Almost always context. If the whole system prompt, the full document set and the entire conversation history are resent on every turn, you pay for all of it every time. The usual fixes are prompt caching, trimming retrieval to the passages that matter, and routing simple turns to a cheaper model.
Do I need a custom agent or will an off the shelf tool work?
Off the shelf is the right answer when the task is generic and your data is already in the tool's ecosystem. Build custom when the agent must act inside your own systems, when the logic encodes something specific to your business, or when the data cannot leave your infrastructure.
What is the difference between a chatbot and an AI agent?
A chatbot answers. An agent acts: it calls tools, changes state in real systems and is accountable for the result. That difference is the entire reason agents cost more, because acting requires permissions, approvals, rollback and an audit trail that answering does not.
Where to start
Name one task, one user and one system. Write down what "right" means for that task in a sentence a non technical colleague would accept. If you cannot write that sentence, the agent cannot be evaluated, and an agent that cannot be evaluated cannot be trusted with anything that matters.
If you can write it, send it to us and we will tell you what it costs to build and what it costs to run, before you commit to either.
