What AI Agents Actually Cost to Build

$15K to $400K depending on scope. Here’s what drives costs and what you get at each level.

Photo by Immo Wegmann on Unsplash

You asked three vendors to quote your AI agent project. The answers came back: $18,000, $95,000, and $220,000.

Same brief. Three wildly different numbers. This happens on nearly every AI agent project. The confusion comes from a fundamental problem: “AI agent” means different things to different people. It could be a simple rule-based bot that calls one API. Or a multi-step autonomous system that reasons across ten tools, manages memory, handles exceptions, and hands off to humans at the right moment.

The cost difference is real. But it’s not because one vendor is trying to rip you off. It’s because they’re building different things.

The short answer: AI agent development costs between $15,000 and $400,000 depending on what you’re building. A single-workflow agent costs $15,000 to $40,000. A production-ready agent with multiple integrations and monitoring costs $80,000 to $200,000. Multi-agent systems cost $150,000 to $400,000. Then you have ongoing costs that most founders don’t budget for: $1,500 to $8,000 per month post-launch.

Understanding the spread means understanding what’s actually inside an AI agent and why it costs more than a chatbot.

What’s Inside an AI Agent

A chatbot is simple. Take input. Generate response. Done. No memory between sessions. No tool use. No decision-making about what to do next.

An AI agent is different in every dimension.

Orchestration. The agent receives a goal, not just a question. It decides step by step how to achieve that goal. It selects which tools to call, in what order, and what to do if a tool returns an error. This decision-making logic is custom code. It doesn’t come with the language model. It has to be designed, built, and tested.

Tool integrations. The agent needs to act in the world. Reading from databases. Writing to CRMs. Calling APIs. Triggering workflows. Every integration is a separate engineering project. Authentication. Rate limiting. Error handling. Data mapping. All custom work per system.

Memory management. A good agent remembers what happened before. Short-term memory within a session is manageable. Long-term memory — knowing what a customer ordered six months ago or that this invoice was disputed twice — requires a vector database, retrieval logic, and careful design to avoid pulling stale context.

Evaluation. You can’t ship an agent like you ship software. You’re not just checking for bugs. You’re checking whether it makes the right decisions in edge cases. Does it escalate when uncertain? Call the right tool? Avoid hallucinating? Evaluation requires test scenarios, a scoring method, and an engineer who knows how to grade agent outputs.

Each of these four layers adds real cost. None of them exist in a chatbot. That’s the gap.

The Five Cost Components

A well-scoped agent project breaks down into five pieces.

Discovery and architecture: $5,000 to $20,000. Before any code is written, a competent team maps your workflow, assesses your data sources, and selects an architecture. Discovery identifies which systems the agent touches, where humans should step in, what success looks like, and which failure modes are unacceptable. A quote delivered in 48 hours with no discovery is based on a template, not your problem.

Tool and integration engineering: $8,000 to $30,000 per integration. Each system the agent connects to requires authentication, data schema mapping, error handling, write-back logic, and testing against real production-like conditions. A clean REST API integration costs $8,000 to $15,000. A legacy SOAP system with inconsistent data formats costs $20,000 to $30,000. Most production agents touch three to five systems. That’s $40,000 to $150,000 in integration work before the agent logic is written.

Agent logic and orchestration: $15,000 to $60,000. This is the core. Simple linear workflows cost $15,000 to $25,000. Agents with conditional branching, multi-step reasoning, and complex exception handling cost $35,000 to $60,000. Multi-agent systems where one coordinator delegates tasks to specialists add $50,000 to $100,000 on top.

Evaluation and testing: $5,000 to $20,000. This is the most skipped component. It’s also the one that causes public failures. Shipping without evaluation is how agents confidently tell customers their order shipped when it hasn’t. It’s how an agent misclassifies records because an edge case was never tested. Good evaluation means a curated test set, automated scoring for accuracy, human review of flagged outputs, and monitoring that catches accuracy drift post-launch. This takes two to four weeks.

Deployment, monitoring, and observability: $5,000 to $15,000. Getting to production requires cloud infrastructure, logging every decision and tool call, alerting when error rates spike, cost controls so an agent loop doesn’t burn $5,000 in API costs overnight, and human review queues for escalated tasks. This is not optional.

What Different Budgets Actually Deliver

$30,000 to $50,000: Prove the concept. One agent. One workflow. One integration. A real system that handles 80% of the use case under normal conditions. Not a demo. Not production-grade. Think: a support triage agent that routes tickets to the right queue, or an internal research agent that searches your knowledge base and drafts answers. Timeline: four to eight weeks. Risk: underscoping the integration. The target system might have an undocumented API or inconsistent data. Spend $5,000 on discovery first to avoid finding that out mid-build.

$80,000 to $150,000: Production-ready agent. Three to five integrations. Human handoff. Real monitoring. The level where most serious projects live. The agent handles edge cases and escalates when uncertain. It logs every decision and has token budgets to prevent runaway costs. It was tested before launch. Examples: a customer support agent integrated with Zendesk, Shopify, and your shipping API resolving Tier 1 tickets. A sales development agent connected to your CRM enriching leads and drafting outreach. Timeline: twelve to twenty weeks. The jump from POC to production is real. Most buyers underestimate the compliance, integration stability, edge case handling, and monitoring work involved. Start with two integrations in the first phase.

$200,000 and up: Enterprise system. Multi-agent or single agent with enterprise requirements. Multi-tenant architecture. Compliance documentation. Full observability. Capacity to serve hundreds of users simultaneously. Examples: a compliance-grade agent for healthcare with HIPAA controls and audit logging, or an orchestration system connecting six to ten internal systems. Timeline: sixteen to thirty weeks. Governance and security approvals are the biggest timeline risk. Build in four to eight weeks for approvals.

The Ongoing Costs That Kill Budgets

Most founders stop thinking after the build is done. That’s a mistake.

LLM API costs. Every query runs through a language model. GPT-4o and Claude Sonnet cost roughly $2 to $15 per million input tokens and $8 to $60 per million output tokens. At low usage (internal tool, small team), expect $50 to $300 per month. At medium usage (customer-facing, 10,000 to 50,000 queries per month), expect $200 to $1,500. At high usage (customer-facing, growth stage), expect $1,000 to $8,000. Complex multi-step agents with long memory cost more per query than simple single-turn agents. Every production agent needs hard limits on context length and tool-call iterations. Without limits, a runaway session can cost more than a week of normal usage.

Hosting and infrastructure. Vector databases for memory, cloud compute, API gateways add $100 to $1,000 per month depending on load.

Maintenance and monitoring. Budget $1,000 to $5,000 per month for an engineer reviewing performance, tuning prompts, updating integrations when APIs change, and fixing edge cases that surface in production. Models update. APIs change. User behavior reveals scenarios the test set never covered.

Total monthly cost at medium scale (customer-facing, typical growth company): $2,000 to $5,300. At large scale: $3,800 to $16,000.

These costs are not optional. They’re the cost of running production software. Any build estimate without an operational cost estimate is incomplete.

Three Questions Before You Get a Quote

Question 1: What is the exact task the agent will own?

Not “improve customer service.” That’s a goal. The task is: “read incoming support tickets, look up order status in Shopify, determine if it’s a shipping delay or product defect, send resolution email or flag for review.” Specificity drives accuracy.

Question 2: Which systems does the agent need to access?

List every system. Note whether each has a documented API, whether write access needs security review, and whether the data is clean and consistent. This list directly determines integration cost. It’s also the most common source of scope surprises.

Question 3: What does a bad output cost?

An agent that routes a ticket wrong is annoying. An agent that gives incorrect refund information is a liability. An agent that processes a payment twice is a financial error. Understanding failure modes tells your team how much evaluation rigor is needed.

Red Flags in AI Agent Quotes

No discovery included. Quotes without a discovery phase are based on assumptions about your systems and data. Those assumptions will be wrong. When they are, you get scope creep, cut features, or change orders.

No mention of evaluation or testing. If there’s no line item for testing the agent’s decision-making before launch, it won’t happen. Evaluation is unglamorous. It doesn’t appear in demos. Vendors who skip it ship faster, until the agent tells a customer something false in front of your whole team.

Build cost only, no operational cost. A quote that gives you a build price without estimating monthly costs is hiding information. Ask for a 12-month total: build, LLM API at your expected usage, hosting, and maintenance. If they can’t give you that number, they haven’t thought about your architecture carefully enough.

The core principle: AI agents are not just chatbots with more features. They’re infrastructure. They need discovery, integrations, evaluation, monitoring, and ongoing maintenance. The build cost is one number. The total cost of ownership is what actually matters.

Originally published at https://www.raftlabs.com/blog/ai-agent-development-cost