How much does it cost to build an AI agent in 2026
What an AI agent costs to build and to run: published agency estimates, our starting prices, operating costs and the factors that move a quote.
Building an AI agent costs anywhere from a few thousand dollars for a simple FAQ bot to well over $100,000 for an enterprise agent that reads documents and acts across several systems. Published agency estimates put an FAQ chatbot at $3,000 to $25,000 and a retrieval-based (RAG) chatbot at $30,000 to $80,000, plus running costs. Our own build prices start at $18,000 for an agent integrated with your helpdesk or CRM. The rest of this guide explains what moves a project between those numbers.
We build AI agents for customer support and operations from our studio in Armenia, Colombia, so we price this work every week. We would rather you walk into a vendor conversation knowing the ranges than discover them in a proposal. Two caveats first. The market ranges below are vendor-published and vary widely, so treat them as orientation, not as a quote. And the build price is only part of the cost: the agent also has to be run, measured and kept accurate after launch.
What the market publishes
Agencies that publish AI build prices tend to group projects into three tiers. The figures below are published agency estimates gathered from vendor pages, not our prices and not a benchmark we have verified.
| Type of build | Published agency estimate (build) | Running cost (published) |
|---|---|---|
| FAQ chatbot | $3,000 to $25,000 | Not published |
| RAG chatbot over your documents | $30,000 to $80,000 | $400 to $6,000 per month to operate |
| Enterprise RAG agent | $80,000 to $180,000 | Not published |
| Maintenance across all tiers | 15% to 25% of the build cost per year | |
The spread inside each tier is large for a reason: the word "agent" covers very different amounts of work. A bot that answers from a fixed list of questions is a different object from one that searches your policies, looks up an order in your system and hands the conversation to a person with context.
What we charge to start
Our engineering team is in Colombia, which affects what we can charge. For context, FullStack's 2026 nearshore economics report puts agency hourly rates in the US at $150 to $300 and in Latin America at $40 to $100. The effect on a fixed price is real, but it is not the only factor, and we do not claim a fixed percentage saving.
| Starting point | Price | What it covers |
|---|---|---|
| AI Readiness Sprint | $4,500 fixed | 2 to 3 weeks: prioritized use cases, architecture and a fixed estimate. 100% credited to the build if you sign within 60 days. |
| Workflow automation | From $8,000 per workflow | A defined process automated end to end, with a human review step where needed. |
| AI agent development | From $18,000 | An agent for support or operations integrated with your helpdesk, CRM or WhatsApp, with evals and human handoff. |
| AI integration and RAG | From $15,000 | Language models and retrieval over your documents wired into software you already run. |
| AgentOps | $1,500 per month | Monitoring, evals, guardrails, model cost control and monthly improvements. |
"From" means the entry scope. A single-channel support agent with a handful of integrations sits near the bottom. More channels, more systems or stricter review requirements push it up, and we fix the price after the readiness sprint rather than quoting blind.
What moves the cost
When two quotes for "an AI agent" differ by a factor of four, the difference is almost always in these places.
- Number of systems it touches. Reading from a knowledge base is cheap. Writing to a CRM, issuing a refund or updating an order is where permissions, error handling and testing take time.
- Quality of your source material. A retrieval system is only as good as the documents behind it. If policies are outdated, contradictory or locked in scanned PDFs, cleaning them is part of the project.
- Channels. Web chat, email and WhatsApp each have their own formatting, rate limits and expectations. Voice adds latency and speech handling.
- Risk of a wrong answer. An agent that suggests a reply to a human is easier to ship than one that answers customers directly about money or health. Higher risk means more guardrails and more evaluation.
- Evaluation effort. Writing a test set from real conversations and scoring the agent against it is real work. It is also the part most often skipped by low quotes.
- Handoff design. Deciding when the agent stops, what context goes to the person and how they take over is product design, not a setting.
Why cheap and fast is risky
Two public data points explain why we spend time on evaluation. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027. And in the LangChain State of Agent Engineering report (June 2026), quality was the top barrier to putting agents in production, even though 57.3% of respondents already had agents running. Customer service was the most common use case at 26.5%.
Neither figure says a particular project will fail. They tell you where the risk sits: not in getting a demo to work, but in keeping answers correct once real users arrive. A price that omits evaluation and monitoring is not cheaper, it moves the cost to later.
What it costs to run
Operating costs are the part that surprises people most. Budget for four lines.
| Running cost | What it is | How it behaves |
|---|---|---|
| Model usage | Per-token fees to the model provider | Grows with volume and with how much context each answer needs |
| Infrastructure | Hosting, vector store, queues, logging | Mostly steady; grows with document volume |
| Monitoring and evals | Reviewing conversations, catching regressions | Needs a person or a service, not only a dashboard |
| Maintenance | Updating documents, prompts, integrations and tests | Published estimates are 15% to 25% of the build per year |
Published estimates for RAG chatbots put total operating cost at $400 to $6,000 per month, which is a wide range because volume and integrations vary. We separate this into an AgentOps retainer at $1,500 per month so the cost is visible and the work is defined: evals, monitoring, guardrails, model cost control and a monthly improvement cycle. You can also run it in-house if you have the people.
Ways to keep the cost sensible
- Start with one use case. One queue of tickets or one document process, not "AI for the company".
- Suggest before you automate. Let the agent draft replies for staff first. You collect real evaluation data and reduce risk before it speaks to customers alone.
- Use existing tools where they fit. If a platform does the job, buy it. We compare the options in build vs buy for AI.
- Fix the price after discovery. A short, paid readiness sprint costs less than discovering scope halfway through a build.
- Own your code. Code lives in your repositories from day one, so you can change vendors without rebuilding. The terms are on our trust page.
How to compare quotes
Ask every vendor the same five questions: What exactly does the agent do, and what does it not do? How will accuracy be measured before launch, and who owns the test set? What happens after launch, and what does it cost per month? Who owns the code and prompts? What is the plan when the agent is wrong? A vendor who answers these concretely is quoting a real project. A vendor who answers with a number and a demo is quoting a prototype.
If you want us to look at a quote you already have, book a scoping call. We will tell you what we would change, even if you stay with the other vendor.
Frequently asked questions
How much does a simple AI chatbot cost?
Published agency estimates for an FAQ chatbot run from $3,000 to $25,000. The low end is a bot on a fixed set of questions with little integration. If it must search your documents or connect to your systems, it moves into the next tier.
Is the monthly cost bigger than the build?
Not usually in the first year, but it adds up. Published estimates put maintenance at 15% to 25% of the build cost per year, and RAG chatbots at $400 to $6,000 per month to operate. Plan for both when you set a budget.
Can we start smaller than $18,000?
Yes. Our AI Readiness Sprint is $4,500 and produces prioritized use cases, an architecture and a fixed estimate. It is credited in full to the build if you sign within 60 days. Individual workflow automations start at $8,000 per workflow.
Why do quotes differ so much?
Because "agent" is not a fixed product. Integrations, document quality, channels, risk level and evaluation work account for most of the difference. Compare scope before price.
Do we own the agent once it is built?
Yes. Contracts and invoicing are in USD through Cloudsoft Industries LLC (Delaware). You own 100% of the code and IP once each milestone is paid, and the code lives in your repositories from day one. A mutual NDA is available.