Skip to content
DT Nova

AI agent development cost: what an AI agent costs to build and run

AI agent development cost is set by what the agent is allowed to do. An AI agent does not only answer: it plans steps, uses tools and changes things in your systems. Each tool it can use needs a connection, a permission boundary, rules for human approval, logging and tests. Running costs follow the number of tasks and the steps per task. This guide explains each part of the budget, the oversight work that makes agents safe to use, the hidden costs, and how to compare quotes.

It does not quote our own prices. It explains how an agent budget is built, so any quote you receive, from us or from anyone else, can be compared on the same basis. For what we build, see AI agent development; for AI budgets in general, see artificial intelligence cost; for answering-only assistants, see chatbot development cost.

The short version

  1. Design: pick one task, define what done looks like, and decide what the agent may and may not do.
  2. Tools: secure connections to each system the agent reads from or writes to.
  3. Trust: a test set of real cases, approvals for risky actions, logging and monitoring.
  4. Running: model usage per step, hosting and the people who review the agent's work.

What drives AI agent development cost

DriverWhy it adds workHow to keep it in check
Number of toolsEach system the agent uses needs a connection, error handling and testsStart with the two or three tools the task really needs
Read or writeWriting to systems needs permissions, approvals and a way to undoRead-only first; add actions with a person approving
Task complexityMore steps mean more places to go wrong and more model usageSplit long tasks into shorter, checkable ones
Test setTrust needs real cases with expected outcomes, rerun after every changeCollect cases from the work people already do
OversightApproval screens, logs and alerts let people see and stop what the agent doesApprove costly or irreversible actions only
Data and privacyPersonal or confidential data in tasks needs access control and agreementsGive the agent only the data each step needs
Risk classUses the AI Act lists as high-risk carry documentation, oversight and logging dutiesCheck the use case against the Act before design

Running costs: measure per completed task

Agents usually make several model requests per task: to plan, to use each tool and to check results. Hosted models are priced by the amount of text processed, so cost per completed task is the measure that matters, together with the share of tasks done correctly. Ask any supplier to measure both on your test set before launch, and to set a step limit and a monthly budget alert. Providers publish current model prices on their own pricing pages; check them on the day you compare.

Oversight is part of the cost

An agent needs people around it: someone to approve risky actions, someone to review a sample of completed tasks, and someone to update the test set when the business changes. Budget that time from the start. In the EU, AI systems that interact directly with people must be designed so they know they are dealing with an AI, unless that is obvious (Regulation (EU) 2024/1689, Article 50(1)), and the regulation sets 2 August 2026 as its general date of application (Article 113); both read 12 November 2026. For governance, see our guide to AI governance.

Build, buy or combine

OptionWhere the money goesFits when
Agent feature in a product you useA licence or add-onThe task lives inside that product
Agent platformA subscription plus configuration and connectorsCommon tasks and supported systems
Custom agentBuild, tools, tests, then hosting and model usageYour own systems, rules or data location matter

A ready-made feature is often the cheaper start. Compare over two to three years at your expected task volume, including review time, which is needed in every option.

Hidden costs of AI agents

  • Error handling. Systems time out, return odd data or change their interfaces; the agent must cope.
  • Retesting. Every change to prompts, tools or models needs the test set run again.
  • Model retirement. Providers retire models, and switching needs retesting.
  • Integration upkeep. See API integration cost.
  • Review time that grows with volume until accuracy is proven.

People's time

Most of the build is labour, and labour costs follow location. Eurostat's release of 31 March 2026 puts average hourly labour costs across the whole economy in 2025 at €34.9 in the EU and €38.2 in the euro area, from €12.0 in Bulgaria to €56.8 in Luxembourg (read 12 November 2026). Your own experts' time is needed too: they know what a correct outcome looks like.

How to compare AI agent quotes

  • Which task, tools and systems are in scope, and which actions can the agent take?
  • Which actions need a person to approve?
  • How large is the test set, and what accuracy is promised before launch?
  • Estimated cost per completed task at your volume?
  • Which model, and can it be changed?
  • Logging: what is recorded, where, and for how long?
  • Are code, prompts and test sets in your company's repository?
  • What is excluded, in writing?

All our services are on the services page, and every guide is in the guides library.

Questions people ask

How much does AI agent development cost?

It depends on what the agent is allowed to do. An agent that drafts replies for a person to approve is a much smaller project than one that updates records, places orders or moves money in several systems. The number of tools it uses, the permissions and approvals around them, the test set needed to trust it, and the volume of tasks set most of the budget.

Why do AI agents cost more than chatbots?

Because they act. Every action an agent can take needs a secure connection to a system, a permission boundary, rules for when a person must approve, logging, and tests for what happens when the model makes a mistake. A chatbot that only answers needs far less of that.

What does an AI agent cost to run?

Model usage, which grows with every step an agent takes, plus hosting, logging, monitoring and the time people spend reviewing its work. Agents often use several model requests per task, so cost per completed task is the right measure. Ask for an estimate at your expected volume.

How do you keep AI agent costs under control?

Start with one narrow task, measure cost and accuracy per task, use smaller models for simple steps, limit the number of steps an agent may take, set budget alerts, and keep a person in the loop for actions that are costly to undo.

Should we buy an agent platform or build our own?

Buy when the task is common and the platform connects to your systems. Build when the task is specific to your business, needs your own rules, or must keep data where you decide. Many projects combine both: a platform or framework with custom tools and tests.

Sources

Your AI agent scope checklist

  • One task, described as the steps a person takes today.
  • What a correct outcome looks like, with real examples.
  • The systems involved, and whether the agent reads or writes.
  • Actions that need a person to approve.
  • Tasks per month, now and in a year.
  • Personal or confidential data involved.
  • Who reviews the agent's work, and how often.
  • A cost-per-task target and a monthly budget.
  • Whether the use case could be high-risk under the AI Act.
  • Ownership: code, prompts and test sets in your company's accounts.