How to build an AI agent for your business, and what it costs in 2026

What is an AI agent, in plain words?

An AI agent is software built around a language model that chooses its own next step towards a goal: it reads a request, fetches data, takes a permitted action in your systems and checks the result. For a business the limits matter most: one job, a few tools, a cap on every action and a named person to turn to.

A chatbot answers questions, an agent also changes something in your systems, and an automation workflow follows steps a person wrote in advance. Our separate comparison guide covers how to choose between them; this one is about building an agent and pricing it.

What can an AI agent do for a business?

A good first job repeats dozens of times a week, relies on data already in your systems and ends in an action you can undo. Keep money, refusals and promises to customers with people.

How do you build an AI agent yourself, step by step?

A first agent can be built without a developer over a few evenings, as long as choosing a model is not where you start. Work in this order:

1. Pick one jobA process that recurs daily and costs someone time, such as sorting enquiries. One job, one agent.
2. Write the rulesOne page: the goal, the data the agent may read, what it must never do and when it calls a person.
3. Collect test cases20–30 real requests with the correct outcome, including odd ones and cases where the right answer is to stop.
4. Choose a buildern8n or Copilot Studio without code, the Agents SDK with a developer; keep personal data where your law requires.
5. Connect toolsNo more than three actions with the narrowest rights: read and draft rather than delete and send.
6. Add approvalA person approves risky tool calls; in n8n the approval sits on a specific tool.
7. Test and launchRun every test case, give the agent part of the traffic and read its action log daily for a week.

Most failed first attempts skip steps two and three: with no rules and no examples there is nothing to check the agent against, and the review shrinks to “it seems to answer”.

Which AI agent builders are worth a look in 2026?

No single builder suits everyone: your data, your budget and whether anyone on the team writes code decide it. This is what the vendors publish as of 16 October 2026.

OpenAI announced on 3 June 2026 that Agent Builder closes on 30 November and points users to the Agents SDK or Workspace Agents in ChatGPT; its ChatKit interface stays. An agent drawn on that canvas today would need rebuilding within weeks.

Can you build an AI agent for free?

A prototype can be free, a working agent cannot: somebody pays for the model, the server or the set-up hours.

• n8n Community Edition: free to self-host, with almost every n8n feature, agents included; you pay for the server and the model.
• OpenAI Agents SDK: free MIT-licensed code; it needs a developer, and tokens are billed.
• n8n Cloud trial: Starter and Pro can be tried without a credit card.
• Copilot Studio inside Microsoft 365 Copilot: internal agents at no extra cost; outside channels need a Copilot Studio plan.

A free agent with no tests and no log becomes the most expensive kind the day it gets something wrong with a customer.

How much does it cost to build an AI agent?

Build prices follow how much the agent may do on its own. Apptunix, a development agency, published these ranges on 28 May 2026; read them as one seller's list, not a market average.

What does an AI agent cost to run each month?

The monthly bill has four lines: the model, paid tools, the platform plan and the time of whoever handles hand-offs. Take an agent that completes 1,000 tasks a month, each using about 8,000 input and 1,000 output tokens, since an agent calls the model several times per task.

At OpenAI's standard prices on 16 October 2026, gpt-5.4-mini charges $0.75 per million input tokens and $4.50 per million output tokens, so the example comes to $10.50; gpt-5.5, at $5 and $30, comes to $70. Web search on every task adds $10 per 1,000 searches, and the platform plan comes on top.

What can go wrong, and how do you keep a person in control?

• Too much power. OWASP names three root causes of harmful agent actions: excessive functionality, permissions and autonomy. Give the agent few tools, read-only access where possible and approval before it sends or deletes anything.
• Hidden instructions. OWASP describes indirect prompt injection: text in an email, file or web page the agent reads changes what it does. An agent reading your inbox should not be able to move money.
• Personal data. In the EU, GDPR applies, and Article 50 of the AI Act, applicable since 2 August 2026, requires telling people they are dealing with an AI system unless that is obvious.
• Hand-off. In n8n the workflow pauses until a person approves or denies the tool call through Slack, Telegram, Gmail, Microsoft Teams or n8n's chat.

In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls. It also warned of “agent washing”: by its estimate, only about 130 of the thousands of vendors selling agents offer the real thing.

Is your business ready for an AI agent?

Tick what is already true for your company and read the verdict: preparation first, a pilot, or an agent straight away.

How do you order an AI agent from VITON¹³⁠, and what does it cost?

An order moves faster with a one-page job description, 20–30 examples and a list of systems with access details. Tick what your agent needs to see starting prices.

The agent takes 5–8 working days and the audit 5–8 working days. An agent that needs a CRM connection and team onboarding becomes a Production Pilot, scoped in the estimate.