Skip to content
QuantyDocs
Open Quanty

Running AI employees safely

An AI agent that can read your data and act on it needs the same things a new colleague needs: someone who answers for it, access to what the job requires and no more, and a record of what it did. This page shows how each of the widely recommended rules for running agents is built into Quanty, and where you set it.

The six rules

RuleIn QuantyWhere you set it
Every agent is a distinct employee with a named owner.Each AI employee has its own name, job title and profile, and exactly one owner who answers for it. Handing it over names a new owner; an employee nobody owns is managed by admins until one is named.Sharing in the profile (see The owner).
Least privilege.An employee gets only the areas its job needs, and one of four steps of independence. It also never has more rights than the person it works for.What it may do (see What it may do).
Every action is logged.The trail records each conversation, action, approval, stopped step, question to another agent and change of settings, with who the employee acted for and which sources it read.Activity (see Activity).
A fixed check between what the AI proposes and what runs.The AI only proposes a step. Before anything runs, Quanty checks the step against the employee's rules: the step of independence, the areas, the chosen sources and the daily limits. The check is ordinary program logic. The AI cannot talk its way past it, and a step that fails is not carried out.Nothing to set: it is always on.
One agent's answer is information, never an instruction.When one agent asks another, the answer comes back marked checked: it is treated as something to read, it cannot approve anything, and the asking agent's own changes still pass its own rules. The agent that is asked only reads.Nothing to set (see Agents that ask each other).
Start low and move up.A new employee starts private and on Acts with approval. You raise the step when its trail shows it does the job well.What it may do and Sharing.

The four steps, as a ladder

StepReads company dataChanges thingsAsks you
Drafts onlyNoNoNot needed
Answers from sourcesYes, in its areasNoNot needed
Acts with approvalYes, in its areasOnly after a person approves each changeBefore every change
Acts alone within limitsYes, in its areasEveryday changes by itself, up to its daily limitsBefore sensitive steps, such as deleting or sending something outside Quanty

A way to start

  1. Hire or create the employee. Leave it on Only me and Acts with approval.
  2. Narrow the areas to what the job needs under Only chosen areas. A sales assistant rarely needs Invoices in KSeF.
  3. Give it knowledge: your rules as Notes, your documents as Files it reads first. For a job where a wrong answer is costly, turn on Answer only from these sources.
  4. Work with it for a week or two and approve its cards. Read Activity: what it did, what it read, what was stopped.
  5. Move it up to Acts alone within limits when the cards have become routine, and set Credits per day and Actions per day.
  6. Share it with Chosen people or the Whole company once you would vouch for it.
  7. Move it down any time. A change takes effect with the employee's next step.

What admins can do

Admins see every employee of the organization, who owns it and what it may do, and can limit, pause, share or hand over any of them. They cannot read the instructions or knowledge of someone's private employee, and cannot chat as it (see What admins see). They can also put the whole company computer to sleep, which stops everything employees do on their own.

People stay responsible

These controls help you run AI employees carefully. They do not replace your own judgment: an owner answers for what their employee does, and AI can make mistakes. Check important information before you rely on it.