Custom agents and MCP
An agent does not just answer, it runs a multi step process inside your systems. Anything involving money stays with a person, and every step is written to a log.
What people usually arrive with
What we do
Custom agent for a process
Takes the request from email or a messenger, checks stock and prices in the accounting system, creates the order, the invoice and the waybill, and notifies the customer and the manager.
MCP server for your systems
The CRM, accounting system, warehouse, knowledge base and internal APIs are exposed to AI assistants through one server: roles, logging and limits.
Agent testing and control
Test scenarios, quality metrics for answers and actions, protection against injected instructions, limits on actions, and a regular run after model updates.
A person in the loop
Anything that costs money goes for confirmation. The agent closes the routine and hands the doubtful cases over with their context.
What this looks like in practice
Typical scenarios in this area: what people arrive with and what we do about it. These are not named client cases, the confirmed work is collected in the portfolio.
- Situation
- An email enquiry passed through three people before it became an order and an invoice.
- What we do
- The agent reads the email, checks stock and prepares the order and documents. The manager confirms the dispatch.
- Situation
- Every integration had its own connector, and maintaining them cost more than using them.
- What we do
- One connection layer with role based access. New assistants and agents connect in days.
- Situation
- The team worked with ChatGPT, but company data never reached it, so the answers stayed generic.
- What we do
- Access to the knowledge base and the systems through MCP, with a log of who asked what and what was done.
The questions we get most
Book a thirty minute conversation
Pick a time that suits you. We will talk about how the process works today and where it breaks