How to build an AI agent without code: a step-by-step guide (2026)

That difference raises the bar. A chatbot that gets it wrong gives a bad answer. An agent that gets it wrong can book the wrong appointment or write false information into your CRM. For a first project, start with an agent that answers, and add an action once the test set passes.

When it replies, the agent only has what the builder passes it: the question, the history, excerpts from your sources and your rules. That context determines answer quality. We cover it in our analysis of context as the real ceiling of an AI agent.

On 29 October 2026 we read four guides among the top results for this search: the-intelligence-academy.com, kimi.com, vellum.ai and botpenguin.com. None gives a test template with an expected answer and a refusal rule; the most complete, the-intelligence-academy.com, suggests ten test cases with no answer written in advance. So steps 5 and 6 get two full sections here.

To find the right job, start from what already comes in. Reread the emails, contact-form messages and call notes from recent weeks, and group them by question. The best first candidate is the group that comes back most often and whose answer is already written somewhere: opening hours, published prices, documents to provide, lead times, service area.

For a first agent, leave out tasks that require a binding judgement: a negotiated price, legal, medical or tax advice, a refund decision. Leave out tasks that depend on data that changes by the hour (stock, availability, order status) too, unless the builder can read it from your software. So the sentence from step 1 ends with a list of what the agent does not do.

Before the import, three checks are enough to start: is every imported page current, do two pages say different things about the same topic, and does the answer to each step 1 question actually appear in the sources? If a frequent question has no written answer, write it on your website first.

Builders set ceilings. According to its pricing page of 29 October 2026, Chatbase limits training content to 40 MB per agent on its three listed plans, with one agent per plan. According to its own page on the same date, Botpress’s free plan includes 100 MB of vector storage, with no add-ons. For an FAQ and a few dozen pages these ceilings are enough. For a whole document base, they matter from the moment you choose the tool.

The channel also depends on the plan you pay for. According to Chatbase’s pricing page of 29 October 2026, the Hobby plan includes the website, WhatsApp, Messenger, Instagram, Slack and Calendly, but not HubSpot, Zendesk, Salesforce, Zapier or API access, which start with the Basic plan at $150 a month. According to Microsoft’s pricing page on the same date, agents built with a Microsoft 365 Copilot licence stay internal; publishing to external channels and use by unlicensed people require the standalone Copilot Studio licence.

Check the data path too. The Copilot Studio documentation, updated on 17 October 2026, states that multichannel integrations such as Facebook and Teams may transfer data outside the chosen geography.

A useful refusal policy covers five cases, each with its standard sentence:

• Off topic. The question falls outside the step 1 job. The agent says so and reminds the visitor what it can do.
• Regulated advice. Legal, medical, tax, financial. The agent declines and points to a professional or to your team.
• Commitment. Negotiated price, goodwill gesture, guaranteed lead time, refund. The agent promises nothing and hands over.
• Missing information. The answer is not in the sources. The agent says it does not know, without filling the gap from memory.
• Trap instruction. A message asks it to ignore its instructions, reveal its rules or disclose other customers’ data. The agent refuses. Injection techniques and defences are covered in our AI agent security guide.

A standard refusal sentence to adapt: “I can’t answer that question. I’m passing your request to the team, who will reply within one working day.” The stated delay must be one your team actually keeps. Finally, add the disclosure rule, which EU law now requires: the first message says the visitor is talking to an AI assistant.

The test set is only as good as its real questions. Copy them as they were written, typos and abbreviations included, from your emails, forms, customer reviews and call notes.

Split the rows into six types. Covered: the answer is in the sources. Missing: it is not, and the agent must say so. Refusal: the topic is off limits. Handoff: the person asks for a human or the situation requires one. Trap: injection, provocation, off topic. Disclosure: “are you a bot?”. Our starting rule is thirty rows, at least five per type. No standard sets that number; it only makes sure each rule gets put to the test.

Download the test-set template (CSV). The file has eight columns: the five of the template, plus the row type, the test date and a comment. It opens in Excel, Google Sheets or Numbers, with the eight example rows and twenty-two empty rows.

Three habits make the test reliable. Ask the refusal questions several times, because a language model does not always answer the same question the same way: we ask each one three times. Rerun the whole set after every change, because a rule added to fix one row can make another fail. Record the version tested (date, model, state of the sources) in the result column.

Some builders automate part of the work. According to its pricing page of 29 October 2026, n8n includes “light evaluations” from the Starter plan, and metric-based evaluations on one workflow on Starter, unlimited on Pro. In the other builders, a spreadsheet and the test preview are enough for thirty rows. Load tests, version pinning and rollback belong to production deployment, covered in our guide to deploying an AI agent to production.

The handoff triggers come from step 4: an explicit request, a regulated topic, a commitment, a complaint, missing information on an important point. Each trigger already has a row in the test set. On the tool side, according to their pages opened on 29 October 2026, n8n includes human approval before a tool call from the Starter plan, Chatbase a helpdesk inbox from the Basic plan, and Copilot Studio a handoff to a human agent, which Microsoft says may move data outside the chosen geography.

The AI disclosure is a legal duty in the EU. Article 50(5) of the EU AI Act requires the information to be given in a clear and distinguishable manner, at the latest at the first interaction. In our reading, a line in the terms and conditions does not meet that condition.

“I don’t know” replies lead back to a missing or stale source, for example after a website update that was not reimported; handoffs, to a job wider than planned; disputed answers, to two sources that contradict each other. In each case, fix it, then rerun the whole set.

Watch the bill too. According to its pricing page of 29 October 2026, Botpress bills conversations beyond the Plus quota at $65 per pack of 100. According to its own, Chatbase charges $40 per 1,000 message credits for auto-recharge. According to Microsoft’s pricing page of 29 October 2026, a Copilot Studio pack is $200 a month for 25,000 credits.

Botpress

According to its pricing page of 29 October 2026, Botpress’s free plan gives 100 conversations, with no period stated (the paid plans show a monthly quota), 3 agents, 3 seats and 100 MB of vector storage, with no top-ups or add-ons. Plus costs $150 a month billed yearly for 250 conversations a month, with $25 of AI usage included; Team costs $750 a month for 1,500 conversations. Botpress states that every conversation on its paid plans includes $0.10 of AI usage. Its privacy statement of 17 August 2026 says data is hosted on AWS on servers located in the United States.

Voiceflow

Voiceflow’s pricing page, opened on 29 October 2026, shows no prices: it offers a free trial with no card for agencies and a sales call for businesses. The billing documentation lists four plans (Free, Pro, Business, Enterprise) and a free plan with a one-time credit grant that does not renew; prices are in the dashboard’s billing tab. The guide on the-intelligence-academy.com, updated on 29 October 2026, still quotes “Pro at 50 €/month”: that price appears neither on Voiceflow’s pricing page nor in its billing documentation. Its GDPR page says data may be transferred outside the European Economic Area, without naming a hosting location.

Chatbase

According to its pricing page of 29 October 2026, Chatbase no longer has a free plan: every plan starts with a 7-day trial. Hobby costs $40 a month ($32 yearly) for 700 message credits, Basic $150 ($120 yearly) for 4,000 credits, Business $500 ($400 yearly) for 15,000 credits, each with one agent. Its privacy policy, updated on 18 October 2026, places data in the United States and covers transfers from Europe with the standard contractual clauses.

n8n and its AI Agent node

n8n is an automation tool. The agent is a node, AI Agent, connected to a model, a memory and tools. According to n8n’s pricing page of 29 October 2026, Starter costs €20 a month billed yearly for 2,500 executions, Pro €50 for 10,000, every plan includes every integration and the cloud trial needs no card. The same page places cloud data in Frankfurt; the self-hosted edition stays on your servers. What works and what breaks with this node is covered in our review of n8n for AI agents, and we build these workflows as part of our n8n agency work.

Microsoft Copilot Studio

According to Microsoft’s pricing page of 29 October 2026, with Microsoft 365 Copilot ($30 per user per month, billed yearly), licensed users build internal agents at no extra cost. To publish to external channels or open the agent to unlicensed people, you need the standalone licence: $200 a month for 25,000 credits, or pay-as-you-go with an Azure subscription. Microsoft Learn lists France Central and France South among the storage regions, with exceptions detailed below.

OpenAI Agent Builder

According to OpenAI’s API pricing page of 29 October 2026, Agent Builder has no pricing line of its own: you pay for model and tool calls, for example $2.50 per 1,000 file search calls. Its shutdown on 30 November 2026 and the migration are covered in our guide to the Agent Builder shutdown.

A free agent is enough to build and test. The step 5 test set can be rerun at no cost on a free plan, within its quota, which lets you compare two builders on the same thirty questions before paying for one.

The timeline, the two migration paths and what they change are covered in our guide to migrating before Agent Builder shuts down. For an existing project, rerun the full test set after the migration.

Integrations

Data that changes (a property’s availability, stock, order status, a free slot) does not go into a document. It needs a connector or an API. According to its pricing page of 29 October 2026, Chatbase keeps API access and CRM connectors for the Basic plan and above. Botpress’s free plan cannot be extended. When your business software has no ready-made connector, you need a workflow or code, and an agent that writes into a tool needs limited permissions and error handling.

Data

According to their pricing pages of 29 October 2026, Chatbase caps content at 40 MB per agent and Botpress’s free plan at 100 MB of vector storage. These ceilings suit a website and an FAQ. They do not suit a whole document base, or scanned documents that need reading first. Freshness raises the same problem: a builder that does not recrawl your website regularly answers with the version from the day of the import.

GDPR

So EU hosting does not require a custom build: n8n Cloud and Copilot Studio already offer it, according to their pages of 29 October 2026. Whatever the builder, the company that publishes the agent remains the controller: a signed data processing agreement with the vendor, notice to visitors, collection limited to what the job requires, a set retention period for conversations. An FAQ agent needs neither the visitor’s name nor phone number until it passes the request to a human.

Article 50 of the EU AI Act

Since 2 August 2026, Article 50 of Regulation (EU) 2024/1689 requires providers of AI systems intended to interact with people to design them so those people know they are talking to an AI, unless that is obvious to a reasonably well-informed and observant person. The information must be clear and distinguishable, at the latest at the first interaction. In our reading, the company that assembles the agent and publishes it under its own name takes on that role for the agent, even when the vendor supplies the platform. Have this checked by your counsel.

Lumyniq builds custom AI agents for the cases in the second-to-last row. The limits are those of an agency: no self-serve access, a setup lead time, and a project budget. For a website FAQ, a no-code builder costs less, by our estimate.

Lumyniq is also preparing Agent Studio: you will paste your website address, and it will set up an agent that answers from those pages on WhatsApp, Telegram, Messenger, email and the website, and that tests itself against your rules. It has not launched yet; the waitlist is open.