AI Agent Development and Implementation

AI agent development at Metanow

Metanow designs, builds and runs AI agents that take over defined parts of a workflow: answering and qualifying enquiries, finding information in your documents, or creating and updating records in Odoo and other business systems. Each agent is built for one clear job, connected to the tools it needs and limited to what it is allowed to do.

The service covers the whole path from scoping and model configuration to system integration and deployment. It sits next to our Intelligent Automation service and our AI-Native Operations programme, where we rebuild a complete high-margin workflow around agents.

Why many AI pilots never reach production

Many AI projects stop at a chatbot demo. The agent has no access to real data, cannot act in the systems where the work happens, and nobody can check what it did. Without guardrails, logs and a person for the difficult cases, it stays a pilot.

How we build an AI agent

Every agent project starts from the same scoping template, so you know upfront what the agent is, what it may do and when it counts as done.

Role and persona

We define the agent's role, for example customer support, lead qualification or technical assistant, its tone of voice and the languages it works in, and write the system prompt that sets its identity.

Model and memory

We choose the base model for the task, such as GPT-4o or Claude, and decide whether the agent needs only session memory or long-term memory through a vector database and retrieval (RAG).

Tools and integrations

We list exactly what the agent is allowed to do: read a knowledge base, PDFs or your website; create leads, update tasks or check inventory in Odoo; call external APIs such as Stripe, Slack or Google Calendar; and hand over to a person when a defined condition is met.

Knowledge sources

The agent answers from your own material: internal documents, selected web pages prepared for retrieval, and read-only access to the database tables and fields it needs.

Guardrails and approvals

Scope limits define what the agent must not discuss, such as pricing or competitors. You decide between fully autonomous actions and human-in-the-loop approval, and data privacy rules such as PII masking and GDPR requirements are part of the setup.

Definition of done

Before go-live we test the agent against real test cases for intent recognition, verify the integration with your main system, check that the fallback message appears for out-of-scope questions, and ask you to approve the tone of voice.

Maintenance after launch

With an optional maintenance plan we review the logs for unanswered or wrong answers, tune the prompt, sync new documents into the knowledge base, manage API keys and usage limits, update n8n nodes and libraries, rerun regression tests after updates and keep token costs under control.

Tools we work with

• OpenAI GPT-4o, Anthropic Claude or open models such as Llama
• Retrieval (RAG) with a vector database
• Odoo ERP and CRM
• n8n workflows
• APIs such as Stripe, Slack, Google Calendar and WhatsApp

Where we already use AI in production

• Our own website: the German, Italian and Albanian versions of metanow.com are translated by an AI agent inside Odoo, and each result is checked for intact HTML before it is saved.
• Client SEO projects: an n8n autoblogger publishes AI-assisted blog posts on client WordPress websites, based on the keywords from our SEO research.

Frequently asked questions