AI automation and agents
AI automation and agents in your processes, with permissions and a log of every action
We put AI to work inside YVY with the rules of a colleague: an identity, permissions, a human owner, a log of every action. Three routes, also together: the assistant inside YVY, which you name as you like (at Aldeia, where YVY was born, it's called Ambrogio); your own chatbot (Claude, GPT or another) connected through the MCP server; AI agents that are users of the system and receive work like people do. Sensitive actions are always approved by a person.
Almost every company has tried AI by now. Almost always next to the business software: copy the data, paste it into the chat, copy the answer back. YVY takes the missing step. The AI reads and writes where the real data lives, and it cannot do anything the person using it could not do.
What are the routes for AI to work in YVY?
| Channel | Who acts | What it can do |
|---|---|---|
| The assistant inside YVY, under the name you give it | The person writing to it, with their permissions | Reads and creates on its own; prepares edits, sends and deletions on a confirmation card |
| Your chatbot (Claude, GPT or another), connected through the MCP server, phone included | The person, with their permissions | Reads, creates and edits; sensitive actions are prepared and approved in YVY |
| Claude Code and the command line | The person, with their permissions | Like the chatbot, from the terminal and in development projects |
| AI agent | The agent, with its own identity and a human owner | Reads and creates; edits only what the administrator enables; never sensitive actions |
One set of permissions and one log apply to all four. What a person cannot see in the interface, their AI cannot see through any channel.
What do we actually do?
We map your processes
We look for repetitive, measurable work: sorting email, drafting replies, updating project progress, planning the week, filling in records.
We pick the right channel for each
A quick question goes to the assistant. A long, recurring job goes to an agent. People who live in a chatbot connect it to YVY.
We create the agents
In Settings › Users › New agent: name, Owner, and under Enabled edits what it may change beyond reading and creating.
We connect your people
The MCP connector in each person's chatbot, for web, desktop and phone, and Claude Code with a starter kit already connected to YVY.
We test on real data
Every automation first runs under its owner's eyes, with the log open.
We leave the rules in writing
Who answers for what, what a person approves, how an agent is switched off.
What can an AI agent do, and what not?
An agent in YVY is a user that is not a person. You assign it like a colleague: owner of a project node, a task, a ticket.
- It gets the work with the specs. On assignment YVY sends it a complete package: what to do, context, link, deadline and how to hand the work back. It finds it in its work queue via MCP, on a signed webhook or by email.
- It works wherever it likes. YVY does not decide how the agent works: it can work outside, with its own tools.
- It hands the work back. It closes the task, writes where the results are and mentions whoever continues, person or another agent. Everything stays in the record's conversation.
- It has limits. At most 20 writes a minute, with an intent line in the log before each write. Never sensitive actions.
- It has an owner. If the owner is no longer active, the agent switches off. Suspend stops it in one click.
Who approves sensitive actions?
Always a person, inside YVY. When an AI asks to issue an invoice, record a payment, close an opportunity as won or lost, write to the client, touch an IBAN, remove something or change roles and permissions, it does not execute. It prepares the action and sends a link. The person opens the page, checks Action data and chooses Approve and run or Reject. If nobody decides within the time shown, the action lapses.
What examples already exist in YVY?
- Email. The assistant sorts incoming mail, drafts the reply under each email and suggests the actions to take in YVY, confirmed with an OK.
- The project manager's week. YVY proposes the week's plan; with the assistant you test what happens if you move a deadline.
- Progress from code. A project linked to a repository updates its nodes' progress from the work done there.
- Quick questions. "What is due this week?", "Summarise the unread conversations", "Create a task for tomorrow", "Which opportunities are open?".
How much does it cost and how long does it take?
It depends on how many processes we automate and how many agents you need. We set it after the analysis, in writing.
- Price: on a quote, after the analysis.
- Typical duration: to be defined.
Frequently asked questions
Can an AI agent work safely in business software?
Yes, if it has the rules of a colleague. In YVY it has its own identity, a human owner, permissions set by the administrator, a write limit and no access to sensitive actions. Every action it takes is in the log.
Does the AI see all company data?
No. The assistant, the chatbot and the command line see only what the person using them sees. An agent sees only what its permissions allow.
Which AI do you use?
The YVY assistant runs on Anthropic's Claude. Through the MCP server you connect your own chatbot: Claude, GPT or another. Agents can run on Claude Code, on Anthropic's Agent SDK or in a pipeline, with credentials meant for automated use.
Does data go through Anthropic?
When you use the assistant or your own chatbot, the request text and the data needed for the answer go through its model: Claude for the assistant, the one you choose for your chatbot. The assistant is switched on per company: until an administrator turns it on, nothing runs.
What is the MCP server?
It is the standard connection that lets an external AI, such as Claude or GPT, work inside YVY. We explain it in What is an MCP server.
Can I give voice commands from my phone?
Yes, with your chatbot on your phone connected to YVY. The connector is added first on the web or desktop, then works on the phone too.
Who answers if an agent gets it wrong?
Its owner, who is always a person. The log shows what the agent did, when and with what intent.
Let's talk about what to automate
Bring us the job that eats the most hours every week. On a call we tell you whether it goes to an agent, to the assistant or to your chatbot, and with which rules.
Book a call