In your company's own accounts
The tool uses Microsoft 365 or Google Workspace and AI services set up under your company's account, such as Claude or ChatGPT.
Your company's accounts
- Email and documents
- The agent
- AI service under your account
AI tools for business, and the technology behind them
We choose the AI for each task and explain how it works before it goes into your company.
What goes into our AI tools
Each part below does one plain job, and a tool uses only the parts its task needs.
Agents and connections
- Agents
- They take on tasks with several steps: reading the request, checking systems, drafting, then handing the result to a person. Each agent works with a set of tools, such as reading email or looking up a price.
- Voice agents
- They answer calls in European Portuguese to take bookings, give the status of an order or pass the caller to the right person.
- Connections to your systems
- We connect email, documents and the ERP through MCP, the open standard the main vendors have adopted. MCP stands for Model Context Protocol.
- Screen and browser use
- For older systems with no direct connection, the agent works on screen as a person would.
Models
- Reasoning models
- We use them where a task needs analysis or calculation, such as building a proposal or finding anomalies in data.
- Open models on your servers
- We run Llama, Mistral, Qwen or Gemma through Ollama or vLLM on a company server, and the data never leaves it. We also use DeepSeek, gpt-oss and other open models from Hugging Face, run with Ollama, vLLM or llama.cpp.
- Models tuned to your company
- When one task runs at high volume, we adjust a small model to the company's documents and tone. Done with LoRA fine-tuning or distillation, in which a small model learns from a larger one.
Reading and finding
- Document reading
- Invoices, forms and scans become structured data in the right system.
- Search across company knowledge
- The team asks a question and gets an answer drawn from procedures, price lists and past emails, with the source always shown.
Marketing material
- Image, video and audio
- Marketing material generated with AI and reviewed by your team before it is used.
Private AI: where your company's data goes
We set up each tool in one of two places, and agree which with you before the build starts.
On your company's own server (local AI)
Open models such as Llama, Mistral or Qwen run on a server that belongs to your company, and the data never leaves it.
Your company's server
- Your documents
- The agent
- Open model
- Your data stays in your accounts
- The tools run in your company's own accounts, and we use your data only for your project. Before we start building, we sign a GDPR data-processing agreement with you.
- Legal support
- When a project needs legal support (GDPR, the AI Act, contracts), we work with lawyers who provide it.
How we choose a model for a task
The choice depends on what the task involves, how often it happens and what data it touches. Some tasks need a large model, used through your company's account; others run well on a small open model on your own server. We explain the choice before anything goes into your company.
Testing and monitoring
Automated evaluations and logs that show, month by month, whether everything still works. Before a tool goes live we test it on examples from your company's own work, and we run the same tests again after every change.
Model choice for one task (illustrative sample)
- Task
- Reading supplier invoices into the ERP
- Volume
- About 300 a month
- Data
- Supplier names, amounts, tax numbers
- Choice
- A small open model on the company's server
- Why
- Fixed layout, steady volume, financial data kept in-house
Start with a 30-minute call
Bring one task that takes your team too long, ideally one that comes round every week.
Book a 30-minute call