How much does it cost to build an AI agent?

How much does it cost to build an AI agent? I was asked exactly that in an interview recently. The interviewer wanted the rough low end and high end I would quote, then kept pushing on every number: what makes a project cheap, what makes it substantial, where the line sits between a prototype and a real build. My first answer was $4K. A few questions later I revised it, because $4K seems too cheap. This post is that conversation, cleaned up, with the questions kept in, so you can see how I got from one number to the next.

These are starting ranges, not a price list. The final quote always comes after a short scope, because two agents that sound the same in one sentence can be completely different builds once you look at the domain, the data and where the agent has to live.

How much does an AI agent cost? The short answer

On top of the build, plan for maintenance: around $12K a month for ongoing engineering, and around $8K a month once it’s in late-stage maintenance.

What is the low end and high end for building an AI agent?

First, I think the lowest end to build an AI agent I would quote at around $4K. That’s a simple agent with simple automation that uses some AI integration, maybe one AI model or a different one, and maybe some human in the loop. But this is a very basic integration. Think of a pipeline for content validation, or some primitive customer support, things like that. I’ll come back to that number, because I changed my mind on it.

It gets more nuanced as we go into the depth of integration, the quality metrics, the domain, compliance, the size of the data, the volume of the data, different scenarios, or the security measures you need to take. It all depends. If you have some internal team automation, we’re talking about one thing. If you have production customer-facing features, that’s a completely different risk factor and complexity. It requires implementing custom guardrails.

The model is cheap. The harness is the cost.

So I would factor it like this: the level of integration on its own is not expensive at all. It gets progressively, exponentially more expensive as you implement more layers of hardening around the agent. Actually, an agent is an LLM or AI model plus the harness itself: the process around it. If you have a very thin harness layer, the cost can be small, because you’re just paying for AI tokens. You can choose and optimize the model for your use case. You can use OpenRouter to set up cost strategies and use different models.

But the harness layer can still be expensive or cheap. It’s a custom solution for your use case, and it’s where the actual engineering effort goes. I’ve written more about why that layer is where the real engineering lives, and why it’s hard to buy it as a service.

What turns a cheap agent into a substantial project?

The difference between building traditional software and building a hardened agentic system is that a hardened agentic system requires ongoing maintenance. We see changes in traffic, in the volume of data, in the models, in thresholds, and in all kinds of other factors. That requires constant calibration. So I split the cost into different scenarios. There’s the first release to production, which I factor by the complexity of the product and the scope of the product. And I split out the maintenance cost.

It’s important to understand that an agentic system needs ongoing maintenance: monitoring to see the actual inputs, outputs and responses, and collecting a golden dataset of use cases. That could be a golden dataset of typical customer-support inquiries. Or it could be a financial chat where the user wants to see analytics or analysis based on recent SEC filings or stock-market conditions. You collect this behavior data and use it for annotation, for future improvements of the product or incremental improvements of the product. Documenting the harness, and shipping incremental updates as part of the system, are their own line items.

In the interview my first numbers here were a standard retainer starting from about $8K per month for maintenance, and $15K+ for developing an agent system with a harness. The development number held. The maintenance number I adjusted in the next answer.

Is $4K a realistic price for a simple AI agent?

No. For a simple integration I think it starts at $6K. $4K seems too cheap, so I revised that. And for ongoing work, I’d put it at around $12K per month. The $8K per month number I gave first is more like late-stage maintenance, once the system has settled and the changes get smaller. Early on, while you’re still calibrating against real traffic and building out the golden dataset, it’s closer to $12K.

What do you get for $6K? Why not vibe-code it?

A $6K integration is something beyond vibe coding. It could be a pre-production integration or a simple production integration, a simple solution. In some way you can say, “Yeah, I could do it myself at low cost, or I can hire someone cheaper.” But this is always the same claim. If you need a design, you can hire a designer or you can use AI to build your design. The question always remains the same: would a designer with AI produce a better-quality product, or would you produce the better design with AI but without a designer?

I think the former. The designer with AI still makes you a better design. Even if they use AI, it doesn’t matter. They still have internal guidance, taste, an aesthetic style. The same goes for engineers. Yes, you can build simple projects, applications, integrations, or a backend without any background in APIs or stuff like that. But security, knowledge of the integration process, and actual expertise still play a big role. It takes a lot of knowledge that you can’t always get from the sometimes one-sided, slightly biased engineering approach of your vibe-coding LLM.

What I’m trying to say is that this is still an open question for decision-making when people think about the buy-versus-build mentality. AI makes it possible for non-experts to build a working prototype. It doesn’t make the decision for you. It’s the same thing I look for when hiring agentic developers: can they tell when the generated thing is wrong?

Where is the line between a $6K prototype and a $15K agent?

I think $6K is kind of a POC. I’d use the idea of a production-minded prototype. It’s a prototype that would be used by maybe three users, and that’s also a defining line. You can deploy it to production. You can show it to your board of directors, product designers, or developers for internal testing. You can test it in production with your dev users and interact with it.

But things like security, custom integrations, and the long tail of cases wouldn’t be covered in this prototype. You also wouldn’t have in-depth analytics or observability for it, or extensive error handling and things like that. That’s what separates the two tiers. At $6K the thing works. At $15K it has the reliability, risk handling and operational support to stay working when real users show up.

How much does a production AI agent cost? $15K+

You can say $15K is the bare minimum for a project. It could be something simple, like an internal integration, a simple app integration, or maybe an iOS app where you want to add agents on the backend that interact with customer data. For example, a mental-health app that uses data from customers to give them feedback or create dynamic scenarios. That’s a simple idea, but it touches customer data, so the harness has to be there.

It can get much more complicated when you start dealing with multimodal cases: audio data, video data, and so on. Each new modality brings its own processing, its own failure modes and its own cost, so a multimodal version of the same app is a bigger build.

What does production ready actually mean for an AI agent?

Sometimes it’s not that complicated to launch a product, but it’s hard to conclude that the actual product is ready to be served to customers. You can say, “I built an app within the last hour, and it’s ready for production.” We don’t know that. You need to review the code. You need to test it yourself. You need to roll it out gradually to customers, probably starting with a few customers at the beginning and gradually raising that number to something like 20%, 50%, 80%, and then all customers.

You need to see how they actually interact with your system. You probably need constant review, and you need to keep an eye on the logs and observability. That’s a different phase of the project, and it costs time and money of its own. Testing, gradual rollout, error handling and monitoring real behavior are what production readiness is made of.

What makes an AI agent cost $50K+?

It’s the domain. It’s the volume of traffic. It’s the latency of the integration. It’s the sensitivity of the information you need to process, security guardrails, tool calling, and the sophistication of the harness. It’s the risk factors for the brand, the domain, the company, or the developers working on it. Obviously, all these things together are considered. Maybe the development requires multiple steps, or there are difficulties with the integration.

Again, some companies have very strict policies for how they deploy something to production. In some cases it can be done within one hour: set up your CI/CD, get credentials to push the build to your Kubernetes cluster in production, and you’re good to go and integrate with all the services. In other cases it takes months, or at least a two-week quarantine on internal dev and staging services while your internal QA team goes through the build, tests every scenario, tests integrations, runs regression testing and covers the other factors needed to actually deploy to production. I’ve seen those scenarios as well.

Likewise, things like instrumenting metrics in production take time. Provisioning certificates takes time. This is operational stuff that might be a nuance of your company, and we cannot predict it up front. It isn’t the agent, but it’s part of what it costs to get the agent live.

If we’re talking about generic infrastructure, meaning how to deploy the agent code, take a cloud provider like AWS. Again, it’s nuanced. A separate AWS account with semi-public data is one complexity and one cost. Integrating deeply into the company’s infrastructure, depending on its internal architecture design, is a completely different price. That also requires due diligence: understanding the business, the data, the architecture, and the points of integration. It’s not like we ship a standalone solution that would fit any design.

Some people have very exotic technologies. Some have a legacy stack like Java. Some use closed cloud providers or very complex setups. So designing the harness for the system can be much more sophisticated and costly than “I just have a generic cloud provider, I just host my application”, or using something very standardized like Cloudflare. Those are different scenarios and different pricing factors, because they require domain knowledge, expertise, and an understanding of how the system works at its integration points.

How much does it cost to maintain an AI agent?

Around $12K a month for ongoing engineering while the system is still being calibrated, and around $8K a month for late-stage maintenance. The reason it doesn’t go to zero is the same reason I gave earlier: models, traffic, data, thresholds and user behavior all change, and an agentic system has to be recalibrated against them. The work is monitoring real inputs and outputs, growing the golden dataset, annotating behavior data, documenting the harness, and shipping incremental updates. Budget for the first release and for this as two separate things.

How my estimate changed during the interview

I think the way the numbers moved is useful on its own, so here it is plainly. I first estimated $4K as the lowest-end simple agent. Later I reconsidered and said $4K seems too cheap, and revised the practical floor to around $6K, which I described as a POC, a production-minded prototype used by around three internal or dev users. I put $15K+ as the starting point for a more substantial agentic system with a meaningful harness layer. For maintenance I first said around $8K a month, then clarified that $8K is more like late-stage maintenance, and put ongoing engineering at around $12K a month.

What actually drives AI agent cost

So how much does it cost to build an AI agent?

Around $6K for a production-minded prototype, $15K+ for an agent system with a real harness, $50K+ when the domain, the risk and the infrastructure demand it, and $8K-12K a month to keep it calibrated. Tokens are a small line on that bill. You pay for the harness, and for the rollout and monitoring that prove it’s ready for your customers. If you want a number for your own project, the engagement shapes and how I scope them are on the rates page.