Most quotes for building an AI agent are wrong in the same direction, because they price the build and ignore the running. An agent is not a project that ships and stops costing money. It is a system with a per-request bill, an evaluation habit, and a maintenance tail that starts the day it goes live.
Here is what the four cost layers actually contain, with working ranges for each, and the two questions that decide the number long before any code is written.
How much does it cost to build an AI agent?
A narrow internal agent built by an existing engineering team typically costs $15,000 to $60,000 to reach production. A customer-facing agent with integrations, evaluation and support commonly runs $75,000 to $300,000 in the first year. The spread is not mostly about model choice. It is about how many systems the agent has to touch and how expensive a wrong answer is.
| Agent type | First-year total | What sets the price |
|---|---|---|
| Internal assistant over your own documents | $15,000–$60,000 | Retrieval quality; low blast radius if it is wrong |
| Workflow agent with two or three tool integrations | $50,000–$150,000 | Integration surface, permissions, failure handling |
| Customer-facing agent | $75,000–$300,000 | Evaluation, guardrails, escalation path, support load |
| Agent in a regulated or high-liability process | $250,000+ | Audit trail, human review, compliance sign-off |
| Off-the-shelf platform plus configuration | $10,000–$50,000 plus licence | Configuration effort; you inherit the platform's ceiling |
Treat these as scoping bands, not quotes. Their purpose is to tell you whether a proposal is in a plausible range, and to make the omissions in a cheap one visible.
The four cost layers
Every agent budget has the same four layers. Proposals go wrong by pricing the first and mentioning the rest.
1. Build. Engineering time to first production release: prompt and tool design, retrieval, the integrations, the interface, and the plumbing that makes it observable. For a small team this is usually four to twelve weeks of work, and it is the only layer most estimates contain.
2. Inference. The per-request model bill. Small in absolute terms for internal tools, and the layer that scales with success rather than with effort. This is where a cost-per-successful-outcome model earns its keep, because the number that matters is not cost per API call but cost per outcome the user actually accepted.
3. Evaluation. A test set, a way to run it, and someone who looks at regressions. Teams budget for inference and forget this one entirely, which is why quality drifts silently after launch. Plan for this as a recurring line, not a one-off.
4. Operations and maintenance. Model deprecations, prompt regressions when a provider updates a model, integration breakage when an upstream API changes, and the human review queue if one exists. A working planning figure is 20 to 30 per cent of the build cost per year, higher than conventional software because the substrate underneath you moves.
What does an AI agent cost to run per month?
Running cost is dominated by volume and by how much context you resend, not by which model you picked. A rough per-request model needs four numbers: input tokens, output tokens, model price, and calls per user-visible outcome.
| Scenario | Monthly inference | Dominant driver |
|---|---|---|
| Internal tool, ~50 users, light use | $50–$500 | Usually rounding error against the engineering time |
| Team workflow agent, thousands of runs | $500–$5,000 | Context size per run; tool-use loops multiplying calls |
| Customer-facing, high volume | $5,000–$50,000+ | Volume, retries, and how much of the traffic hits the large model |
Two levers move this more than model shopping does. Routing sends the easy majority of requests to a smaller model and escalates only the hard ones, which routinely halves the bill. Prompt caching and context discipline stop you re-sending the same 6,000-token system prompt on every call. The cheapest token remains the one you do not send.
Are AI agents free to build?
The frameworks are free; the agent is not. Open-source orchestration libraries, vector databases and model SDKs cost nothing to download, and a capable engineer can have something demo-worthy running in a weekend. That is genuinely new and genuinely useful.
What is not free is everything between the demo and production: the evaluation set that tells you it works, the error handling for the fifteen ways a tool call fails, the permissions model that stops it reading data the user should not see, the logging that lets you debug a bad answer three weeks later, and the inference bill once real people use it. The demo is perhaps 10 per cent of the work, which is the single most reliable source of over-optimistic agent budgets.
Can I build my own AI agent?
Yes, and for a well-chosen first use case that is often the right call. The question is which use case, not whether you are capable.
Build it yourself when the agent operates on data you already hold, when a wrong answer is inconvenient rather than expensive, and when the users are colleagues who will tell you plainly that it is not working. Those three conditions describe most good first agents, and they keep the cost of being wrong low while your team learns what "working" even means here.
Bring in help when the agent touches customers or money, when a wrong answer creates liability, or when it must integrate with systems whose failure modes you do not control. The expensive part of these projects is never the model. It is the workflow fit, the data access and the definition of correct, which is exactly the ground a forward-deployed engineer exists to cover.
Build, buy, or configure a platform?
Agent platforms have improved enough that the default answer is no longer "build". The comparison is the ordinary build versus buy one, with one AI-specific wrinkle: platforms move fast, so a capability gap you are building around today may close within two quarters, and you will own the workaround forever.
- Configure a platform when the workflow is common, integrations exist, and your differentiation is elsewhere.
- Build when the agent's behaviour is the product, when your data or workflow is genuinely unusual, or when per-seat platform pricing becomes worse than engineering cost at your volume.
- Do neither yet when you cannot say numerically what a good output is. That is not a budget problem; it is a scoping problem, and spending money before solving it is how pilots die.
How to get a number you can defend
An agent estimate becomes defensible when it stops being a single figure. Five things make it so:
- Name the outcome the agent produces, and how you will know it produced it correctly.
- Prototype ten representative requests and measure their real token counts. This takes an afternoon and beats any estimate from intuition.
- Multiply by a retry factor of 1.2 to 1.5 for anything with tool use or structured output.
- Add the three layers everyone omits: evaluation, maintenance at 20 to 30 per cent per year, and human review if the process needs it.
- Present a range with its largest assumption named, plus what would move the number. A point estimate on a system this new is a marketing claim, not an analysis.
The teams that stay on budget are not the ones that estimated most accurately. They are the ones that scoped the narrowest useful agent, shipped it, and priced the second one using their own numbers instead of somebody else's ranges.
Frequently asked questions
How much does it cost to build an AI agent?
A narrow internal agent built by an existing engineering team typically costs $15,000 to $60,000 to reach production. A workflow agent with two or three tool integrations runs $50,000 to $150,000, and a customer-facing agent with evaluation, guardrails and a support path commonly runs $75,000 to $300,000 in the first year. Agents in regulated or high-liability processes start around $250,000. The spread is driven by how many systems the agent must touch and how expensive a wrong answer is, not by model choice.
Are AI agents free to build?
The frameworks are free; the agent is not. Open-source orchestration libraries, vector databases and model SDKs cost nothing, and a capable engineer can have a demo running in a weekend. What is not free is everything between the demo and production: the evaluation set that proves it works, error handling for the many ways a tool call fails, a permissions model, logging, and the inference bill once real people use it. The demo is roughly 10% of the work.
How much does an AI agent cost to run per month?
An internal tool serving around fifty users typically costs $50 to $500 a month in inference, which is rounding error against the engineering time. A team workflow agent running thousands of times a month lands between $500 and $5,000. High-volume customer-facing agents run $5,000 to $50,000 and above. Two levers matter more than model shopping: routing easy requests to a smaller model and escalating only hard ones, and prompt caching so you stop re-sending the same system prompt on every call.
Can I build my own AI agent?
Yes, and for a well-chosen first use case it is often the right call. Build it yourself when the agent works on data you already hold, when a wrong answer is inconvenient rather than expensive, and when the users are colleagues who will tell you plainly that it is not working. Bring in help when the agent touches customers or money, when a wrong answer creates liability, or when it must integrate with systems whose failure modes you do not control.