Why we build our agents on Claude, the Anthropic model

When an AI agent talks to your customers on the phone, replies on WhatsApp on your company’s behalf or reads your invoices, the question “which model does it run on?” is not a technical detail. It is a business decision. The quality of the conversations, the security of your data and how far you can rely on the system all depend on it.

At Neurotech we made this decision once and we apply it everywhere: every agent we build runs on Claude, the model developed by Anthropic. In this article we explain what Claude is, who Anthropic are and why we believe it is the right choice for businesses that want to use AI seriously.

Who Anthropic are

Anthropic describes itself as an AI safety and research company. On the company’s official page, its mission is framed as building AI systems that are “reliable, interpretable, and steerable”.

A few things that matter for a business client:

• It is a public benefit corporation. Anthropic is organised as a company with a public benefit purpose, and its stated purpose is the responsible development and maintenance of advanced AI for the long-term benefit of humanity.
• It has a dedicated governance structure. Alongside the shareholders, part of the board of directors is elected by a separate body called the Long-Term Benefit Trust.
• It treats safety as a science. The company publishes research on the safety and inner workings of its models, and has a public policy on the responsible development of increasingly capable models.

For you as a business owner, this means a provider that explicitly puts safety and predictability at the heart of its product. Exactly what you want when an automated system is talking to your customers.

What Claude is

Claude is the family of AI models developed by Anthropic. It can be used directly, in a chat app, or integrated into your own systems via the API, meaning a technical connection through which your applications send it requests and receive responses.

We use the API. This means your agent is not a generic chatbot but a system built to measure: it has your company’s instructions, it only has access to the data it needs and it only takes the actions you allow.

Claude’s constitution: written values, not guesswork

One of the things that sets Anthropic’s way of working apart is the approach called “Constitutional AI”. In short, the model is trained on an explicit set of principles, a “constitution”, rather than only on implicit preferences.

Anthropic published the principles of this approach in May 2023, and in January 2026 it published a new version of Claude’s constitution. The company explains that, because the principles are written down, they are easier to specify, check and understand.

Why does this matter for a business? Because an agent that talks to customers has to behave predictably. To stay polite even when the customer isn’t. Not to make up things it doesn’t know. To turn down inappropriate requests. A model built on explicit principles gives you a more solid foundation for this behaviour, on top of which we add the rules specific to your business.

What happens to your data

This is probably the question we are asked most often. Here is what Anthropic says in its Privacy Center: by default, the company does not use inputs and outputs from its commercial products, which include the API, to train its models. The exception is where the customer explicitly chooses to allow it, for example by sending feedback.

In practice: your agent’s conversations with your customers do not become training material for the model. On our side, business data sits in Supabase, on servers in the European Union, and the agent’s access to it is limited to what it needs. The details are on our data security page.

MCP: how Claude connects to your applications

An AI model, however good, is only useful if it has access to the right information. An agent that takes orders needs to see stock levels. An agent that books appointments needs to see the calendar.

In November 2024, Anthropic released the Model Context Protocol (MCP) as an open-source project: an open standard for connecting AI assistants to the systems where data lives. We use MCP connectors, together with n8n, to link agents to the client’s database, calendar, telephony and other applications.

The advantage of an open standard is simple: connections are built to clear, documented rules, not improvised in ways that are hard to maintain.

Why one model, not several

We chose to work with a single model for practical reasons:

1. We know it in depth. We know how it responds, where it excels and what instructions it needs to behave exactly as the client wants.
2. We test once, and test well. Every agent goes through tests on real scenarios. When you work with one model, the tests are more relevant and easier to repeat.
3. We build with it too. Our systems are developed with Claude Code, Anthropic’s coding tool. The same ecosystem, from build to day-to-day running.
4. One set of data rules. The client has a single privacy policy to check, not several.

What this means for you in practice

On Claude we build, among other things:

• AI voice agents that answer the phone, take orders and book appointments, with voice via ElevenLabs and telephony via Twilio;
• chatbots for WhatsApp and your website, which reply to customers around the clock;
• automations that read emails, documents and messages and turn them into actions;
• ERP and CRM applications that the AI agent is connected to directly.

A few real, anonymised results: a chain of patisseries in Sibiu recovers around 20 calls a day that used to be lost, with a voice agent and a custom-built ERP. A hospitality client in Cluj takes direct bookings through its own AI agent, instead of paying around 20% commission to booking platforms.

What a project built on Claude looks like

So this doesn’t stay at the level of principles, here are the steps an agent usually goes through, from idea to going live:

1. Process analysis. Together we decide what the agent does, what it doesn’t do and when it hands the conversation over to a person.
2. Instructions and knowledge. We write the company’s rules: tone, prices, opening hours, exceptions. The agent only receives the information it needs.
3. Connections. We link the agent to the database, calendar or ERP, through n8n and MCP connectors.
4. Testing. We run real scenarios, including customers in a hurry, odd questions and requests the agent has to turn down.
5. Launch and fine-tuning. In the first few weeks we read the conversations and adjust the instructions wherever the agent can do better.

Questions to ask any AI agent provider

Whoever you work with, it is worth getting answers to the questions below before you sign:

• Which model does the agent run on and who develops it?
• Are my customers’ conversations used to train the model?
• Where is the company’s data stored and who has access to it?
• What does the agent do when it doesn’t know an answer?
• How is it tested before launch and how is it monitored afterwards?
• Can I see the conversation history and correct the agent’s behaviour?

A serious provider will give you a clear answer to each. If the answers are vague, that tells you something.

The limits we openly acknowledge

No AI model is perfect. Claude can make mistakes, like any system of its kind. That is why our agents have clear rules: they hand the conversation to a person when the situation calls for it, they don’t promise things the company hasn’t approved and they leave a history you can check. A good agent doesn’t replace your judgement; it takes on the repetitive work so you have time for the decisions that matter.

Want to see what it would look like in your business?

If you want to find out more about how we work, read our about us page. And if you already have a process in mind that you want to hand over to an AI agent, book a free 30-minute call. We will show you exactly what Claude can do for your business, and what it can’t.