Agents
An agent is a program that works on its own inside your tools. While the office sleeps, it reads the email that came in, finds what it needs in your documents and gets every reply ready. In the morning, you say yes or no.
What we build agents with
Agents by department
Development
Prepares changes to internal tools in a test environment; a person approves going live.
See CaesarProduct
Sales
From radar to contract: dossier, deck, proposal and contract, with the CRM up to date; a person approves what goes out.
See MadisonProduct
Marketing
From topic to publication: ideas, scripts, carousels, b-roll, editing and social media; it publishes when you approve.
Customer service
Finds the answer in authorised documents and prepares it with the citation; a person decides what is answered.
Administration
Looks up a piece of data in one tool and updates the record in another; a person approves what is sensitive.
Human resources
Answers the team's questions on internal rules, with the citation; the human-resources person decides what is answered.
Operations
Prepares the day's work from the team's mail, calendar and documents; a person approves sensitive actions.
Accounting
Checks that every received invoice is readable and has the agreed fields, and prepares the pending list; the reviewer validates each item.
Custom agents
The agent no box covers, or several agents connected; a person approves everything that leaves the company.
The contract covers only the scope that is selected and validated, not the sheet.
Portraits generated with AI.
We work with agentic AI
The agent finds context where you already work, delegates to subagents and acts in your tools.
Where it gets context
- Calendar
- CRM
- ERP
- Documents
- Drive
- Web
The agent
Finds contextExecutes
Delegates to 4 subagents
- Reads the request
- Looks the client up in the CRM
- Works out the quote
- Drafts the reply
What it executes
- Sends the reply
- Alerts the team
- Books the meeting
- Updates the CRM
- Issues the invoice
- Prepares the report
- Asks for your yes
- Posts on LinkedIn
Where a person decides
- Before anything leaves the company
- Per action, never by habit
- Never delegated
How an agent works
Five steps, with the data and tools you authorise. Each step has its control.
- Diagram: three sources reach the agent, which weighs two options with their reasons and doubts and asks a person a question.
Reasoning and asking
Compares options, with the reasons and the doubts. If it lacks context, it asks before acting.
ControlExplicit criteria and sources: a good argument does not prove it is right.
- Diagram: the perimeter agreed in writing, two steps done, the next one under way and one step left outside the perimeter.
Choosing steps
Decides the next step, within a perimeter.
ControlWhat is delegated, in writing.
- Diagram: the agent acts with permission in two tools (it updates a record and schedules a task), the third stays locked and everything goes into the log.
Acting with tools
Updates a record, schedules a task.
ControlMinimum permissions and a log of every action.
- Diagram: the draft is ready, but sending stays pending until a person says yes or no.
Waiting for approval
Leaves the action prepared, pending a yes or a no.
ControlApproval before sending, deleting, buying or changing permissions.
- Diagram: the retries reach the error threshold and the agent alerts a person with what is done and what is left.
Handing over to a person
If it cannot finish, it tells someone what it did and what is left.
ControlAn error threshold and a retry limit.
Who decides
The company decides what is allowed and Qvasar proposes the minimum permissions. Being able to do something does not mean being allowed to, or doing it well: tests on your task check it.
Controls catalogue
- Perimeter and minimum permissions
- Prior approval
- Log
- Evaluation
- Error threshold
- Test environment
Chosen per system; not every system needs all of them.
Glossary
- Permission
- What it may do, with which data and tools.
- Approval point
- Where a person says yes or no before a sensitive action.
- Log
- What it did, kept so it can be reviewed.
- Evaluation
- Tests with known answers, before and during use.
- Test environment
- A separate place to test without touching real work.
AI and data
The EU AI Act, in plain words
Artificial intelligence
What we deliver takes the EU AI Act into account.
European UnionEU AI ActRegulation (EU) 2024/1689
Every agent comes with:
- A written record of what it does, with which data and permissions
- A person who approves what goes out and can stop it
- A reviewable log
- Tests on your task
Data protection
Data processing takes the laws of Andorra, the EU and Spain into account.
AndorraLaw 29/2021Qualified Law on the protection of personal data
European Union and SpainGDPR + LOPDGDDRegulation (EU) 2016/679 and Organic Law 3/2018
We look at which personal data the system processes, what for, who can access them and how long they are kept. When there are clients or data in the EU, we also check the legal basis of each processing and that only the data it needs are used.
