AI agent development cost and how a project runs

AI agent development cost is driven by what the agent may do, not by the model behind it: its tools, the systems it reads and writes, the testing that proves it acts correctly, its guardrails, and the model usage you pay for every month. A narrow first agent costs a fraction of a production agent, and an honest price starts from a short brief. This guide covers architecture, every cost line, MVP against production, year-one running costs, the EU AI Act and how a project runs.

Last updated: 21 October 2026.

How much does AI agent development cost?

An AI agent is software that uses a language model to decide which steps to take towards a goal, and then takes them through tools: it searches, reads, fills in, sends, books or updates, and checks the result before it moves on. That is the difference from a chatbot, which only answers. Its cost depends on those actions.

At Odduel, a creative and technology agency in Barcelona and São Paulo, we price AI development from a short brief: the task the agent owns, the people it serves, the systems it touches, what it may do without asking, and how good it must be before anyone relies on it. From that brief we give a fixed quote for a defined first stage. Before it, any figure would be a guess.

In short: the model is rarely the largest line. Integration with your systems, evaluation and guardrails usually take most of the build, because an agent that can act needs proof that it acts correctly. Running costs are real and recurring: model usage grows with every step the agent takes. Start with one narrow agent, measure it, then decide on production.

Why agent quotes differ so much

Two quotes for “an agent that handles our supplier emails” can differ widely and both be fair. One assumes the agent reads emails and drafts replies for a person to send. The other assumes it reads attachments, checks them against the purchasing system, updates orders, replies on its own within set limits and escalates the rest. The second has more tools, write access, approval rules, a larger test suite and an audit log. The cost lines below are how you compare them on equal terms.

Where AI also lowers the cost of the build

AI sits inside the product, and it also carries much of the repetitive production of the build: interface components that follow a design system, boilerplate code for tool connectors, first drafts of test cases, synthetic sample data. That shortens projects and lowers their cost. The decisions that make an agent safe and useful stay human work: what it may do, where a person must approve, how it fails, how it is measured, how personal data is handled. That is where the budget belongs.

How is an AI agent built?

Every agent, however simple, is made of the same parts. Knowing them lets you read a proposal and see what is included and what is missing.

The parts of an agent

• The model: the language model that reads the situation and decides the next step. Usually a hosted model reached through an API; sometimes a model you host yourself.
• Instructions: the written role, rules and limits the agent works within, versioned like code.
• Tools: the actions it can take, each a defined function with inputs, outputs and permissions (search the knowledge base, read an order, create a ticket, send an email).
• Knowledge: the documents and records it can consult, usually through retrieval: the agent finds relevant passages and reads them before deciding.
• Memory and state: what it remembers within a task and, where allowed, between tasks.
• Orchestration: the loop that runs plan, act, check and repeat, with limits on steps, time and spend.
• Guardrails and approvals: checks on what goes in and what comes out, and the points where a person must approve.
• Logging and monitoring: a record of every step, tool call and decision, for debugging, cost control and accountability.

One agent or several

Many tasks need one agent with a handful of tools. Some benefit from several specialised agents handing work to each other, for example one that sorts incoming requests and others that handle each kind. Multi-agent designs add orchestration, more testing and more model usage, so they should be justified by the task, not chosen because they sound advanced. Our comparison of agentic AI and generative AI explains when acting, rather than answering, is worth the extra cost.

Workflow or autonomy

The most important design choice is how much freedom the agent has. A fixed workflow with the model making judgements at set points is cheaper to build, easier to test and more predictable. An open agent that plans its own steps can handle messier tasks but needs more evaluation and stronger limits. Most good business agents sit near the workflow end.

What are the cost lines of an AI agent?

Eight lines explain nearly every AI agent budget. Each one is a decision you can influence before the build starts.

Discovery and design

This line defines the task precisely: what the agent owns, what success looks like, which cases it must refuse or escalate, and what the people around it will see. It includes mapping the current process, writing the first test cases from real examples, and designing the screens where people review, approve or correct the agent. It is a small share of the budget and the line that protects every other one.

Model usage

Model usage is paid per amount of text processed, in and out. Agents use far more of it than chat, because each task can run several steps, each step re-reads instructions, history and retrieved documents, and tool results come back into the conversation. The model chosen matters: larger models cost more per step, smaller ones are cheaper and often good enough for routine steps. During the build this line is modest; in production it is the main recurring cost.

Tools and integration

Integration is usually the largest build line. Every tool is a connection to a real system (your CRM, order system, help desk, calendar, content system or database) with authentication, permissions, error handling and tests. Read-only tools are simpler; tools that write or send need validation, limits and often a human approval step. Old systems without a clean API, or several systems that disagree, add work.

Knowledge and data preparation

If the agent consults documents or records, they must be found, cleaned, split into sensible pieces and indexed for search, then kept current when the sources change. Clean, well-kept sources make this a modest line; scattered knowledge, scanned files and conflicting versions make it a large one.

Evaluation

Evaluation is how you know the agent works. It means a set of realistic tasks with the expected outcome, run automatically on every change to the instructions, tools or model, plus human review of a sample. For an agent, the test checks actions as well as words: did it call the right tool, with the right data, and stop when it should? This line is often under-quoted. It is also what lets you change models later without fear.

Guardrails and safety

Guardrails limit what can go wrong: checks on inputs (including instructions hidden in emails or documents the agent reads), checks on outputs, limits on which tools can be used for which users, spend and step limits, and approval steps before irreversible actions. The more an agent can do on its own, the larger this line.

Interface and hand-over to people

People need to see what the agent did, approve what it proposes, correct it and take over. That can be a panel inside an existing tool or a dedicated screen. A good hand-over design is what makes people trust the agent and keep using it.

Hosting, monitoring and logging

The agent needs hosting for its orchestration, a store for its logs and state, monitoring for errors, latency and spend, and alerts when something drifts. Logs that contain personal data must be kept in line with GDPR: limited, protected and deleted on schedule.

Cost line	What drives it up	Relative share of the build
Discovery and design	Unclear task, many edge cases	Small
Model usage (build phase)	Long test runs, large models	Small
Tools and integration	Write access, old systems, many tools	Usually the largest
Knowledge and data	Scattered or messy sources	Small to large
Evaluation	Many task types, high accuracy needs	Large
Guardrails and safety	More autonomy, sensitive actions	Medium to large
Interface and hand-over	Approval flows, new screens	Medium
Hosting and monitoring	Volume, audit requirements	Small to medium

What separates an MVP agent from a production agent?

The most useful decision in an agent budget is to separate the first agent from the production agent. They answer different questions and cost very different amounts.

What an MVP agent is for

An MVP agent proves one thing: can an agent do this task well enough, with our systems and our data, to be worth building properly? It owns one narrow task, uses two or three tools, often with read-only access or a person approving every action, serves a small internal group and is measured against a written test set. It leaves out scale, full permissions and polish. It ends with evidence: measured accuracy, observed cost per task, the failure cases, and what production would need.

What production adds

Production turns that evidence into something people depend on. It adds the remaining tools and write access, proper sign-in and permissions, approval rules by action and by user, the full evaluation suite running on every change, guardrails against misuse, monitoring and spend limits, logging that meets GDPR, documentation, training for the people who work with it, and support. In our experience this is several times the work of the MVP agent, and it is where most of the budget belongs.

Aspect	MVP agent	Production agent
Question answered	Can an agent do this task with our systems?	Can people rely on it every day?
Scope	One narrow task	The full task and its exceptions
Tools	Two or three, often read-only	All needed, with write access where justified
Human approval	On every action	Only where risk requires it
Users	A small internal group	Customers or whole departments
Evaluation	A written test set, partly by hand	Automated on every change, plus review
Guardrails	Basic limits	Full input, output, permission and spend controls
Monitoring	Logs read by the build team	Dashboards, alerts, audit trail
Relative build cost	Low	Several times the MVP

How does an AI agent project run?

A well-run agent project moves in short stages, each ending with something you can test and a decision.

1. Brief. One page: the task, the people, the systems, what the agent may do alone, what needs approval, how success is measured. This is what a fixed quote is based on.
2. Discovery. Map the current process, gather real examples, write the first test cases, check data access and the legal position.
3. Architecture. Choose the model, the tools, the degree of autonomy, the approval points and where the agent is hosted.
4. MVP build. Build the narrow agent, its tools and a simple review screen; run the test set daily.
5. Pilot with real users. A small group uses it on real work, with a person approving actions; measure accuracy, cost per task and time saved.
6. Decision. Stop, adjust or go to production, based on the evidence, not on enthusiasm.
7. Production build. Full tools, permissions, guardrails, evaluation suite, monitoring, documentation and training.
8. Launch and tuning. Gradual roll-out, close watch on logs and spend, approval steps relaxed only where the data supports it.

The same short-stage logic applies whoever builds it. Our article on hiring an AI development company or a freelancer looks at which kind of partner suits which stage.

What does an AI agent cost to run in year one?

An agent is not paid for once. Plan these recurring lines before the build, not after launch.

What recurs every month

• Model usage. Grows with the number of tasks and the steps per task. Usually the largest recurring line. Caching repeated instructions, trimming what the agent re-reads and sending routine steps to a smaller model all reduce it.
• Hosting and storage. The orchestration, logs, state and any search index.
• Third-party services. Paid APIs the tools call, search, document reading, email or messaging services.
• Monitoring. Tools that track errors, latency, quality and spend.

What recurs in work

• Evaluation runs. Every change, and every new model version from the provider, needs the test suite run and reviewed.
• Model changes. Providers retire and replace models; moving to a new one is work, made safe by a good test suite.
• Knowledge upkeep. Indexes refreshed as documents change.
• Tool upkeep. Connected systems change their APIs, fields and permissions; the agent’s tools follow.
• Review and tuning. A person reads a sample of the agent’s work, adds new test cases from real failures and adjusts instructions.

In our experience the running costs of an agent in its first year are a significant share of the build, larger than for ordinary software, because model usage and evaluation never stop. Ask every supplier for an estimate of cost per task at your expected volume, and for the levers that lower it.

What does the EU AI Act mean for an AI agent?

The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in stages. Its bans on certain practices and the duty to give staff AI literacy have applied since 2 February 2025; obligations for providers of general-purpose AI models since 2 August 2025.

For most business agents, the practical points are these:

• Transparency. People must know when they are dealing with an AI system rather than a person, unless it is obvious. An agent that writes to customers or answers them should say so; this is cheap to build in from the start.
• Risk category. Most agents (support, sorting, drafting, internal operations) are not high-risk. An agent used for decisions on hiring, credit, access to education or essential services may be, and high-risk systems carry heavy duties on documentation, human oversight and quality. Check this in discovery: it changes the budget more than any technical choice.
• Your role. If you build or commission the agent and put it into service under your name, you may be its provider; if you use a vendor’s agent, you are a deployer. The duties differ.
• GDPR still applies. An agent that reads emails, records or documents processes personal data: you need a lawful basis, data minimisation, a processing agreement with the model provider, care with transfers outside the EU and a retention rule for logs.

Check the dates of the later stages, high-risk rules in particular, against the current official text when you plan; this is general information, not legal advice.

How do you read an AI agent quote?

Check a quote for these:

1. The task in one sentence, and what the agent will not do.
2. The tools listed one by one, with read or write access and where approval is needed.
3. An evaluation plan: how many test tasks, what counts as a pass, who reviews.
4. Guardrails named, not implied: input checks, output checks, spend and step limits, approvals.
5. Separate prices for the MVP agent and for production, with the decision point between them.
6. An estimate of running costs, model usage per task at your volume and the recurring work.
7. Ownership: who owns the instructions, tool code, test suite and logs, and whether you can change model provider.
8. The EU position: transparency, risk category and data processing addressed.

A quote with a single total and no test plan is the one to question. If you are also weighing an AI feature inside an app rather than an agent, our app development guide covers how such a feature is built.

Questions people ask

If you have a task in mind for an agent, write the one-page brief: the task, the people, the systems, what it may do alone and how you will measure it. We will turn it into an architecture, a plan for an MVP agent and a fixed quote for that first stage, with running costs estimated. See how we approach AI development, or talk to us about your project.