AI agents and virtual office

An agent is a digital employee: it has a name, a job title and a Nori role, like any person. It uses the assistant under the hood, but on its own: when you assign it something, at a set time or when something happens in the company. What matters it prepares and then waits for someone to approve.

What to know first

• It takes a license of the type its role needs, like a user. If there is no free license, it cannot be created. When you let it go, the license is freed.
• It works with its role's permissions, in its warehouses and, if you choose a salesperson, only with that salesperson's documents and customers. It sees nothing a person with that role would not. It cannot be given the Administrator role.
• It never signs in: not on the web, the point of sale or the app. It does not show in the users catalog.
• It cannot change its role, limits or instructions. Only a person with the permission to set up agents can.
• It uses your company's Anthropic account, the same as the assistant. Anthropic bills the usage directly to that account and the company's monthly AI budget applies to agents too. If you have not connected it yet, see How to connect your company's Anthropic account.

1. Create an agent

You need the Create and set up AI agents permission, which only the administrator license has.

1. Go to AI › Virtual office and press New agent.
2. Who it is:
   • Name and Job title. That is how it shows in the office, the activity log and what it prepares, for example "Sofía · Collections".
   • Job description: what it does, with which criteria and what to avoid. The agent follows it in every task.
   • The Color of its avatar.
   • Supervisor: gets its notices and approves what the agent cannot do by itself.
3. Role and license: the Role, the License (empty uses the cheapest one that covers the role), the Warehouses (empty: all) and the Salesperson.
4. Autonomy and limits (below).
5. Press Create agent.

2. Autonomy and limits

You can also set:

• Monthly cap (US$) and Monthly token cap. When reached, the agent stops until next month. With 0 it has no cap of its own.
• Model: Standard fits almost everything, Reasoning costs more and suits accounting, Light is the cheapest.
• Limit the modules and Limit the actions: if you do not check them, it may use everything its role reaches.

An agent can start only a certain number of tasks per hour, and each task has a maximum number of steps.

3. Give it work

Open the agent from its desk in the office.

• Assign a task: write what you would put in a note for a coworker, for example "review customers more than 30 days overdue and prepare a follow-up for them", and press Assign.
• New task: for what repeats. Choose When:
  • On request: you run it with Run now.
  • Scheduled: daily, weekly or monthly, at a set time.
  • When something happens: a sales order comes in, a home delivery to schedule comes in, there are overdue invoices or items are below their minimum. The last two are checked once a day, in the morning.

For each task you also choose Who does it:

• The agent (AI): it reads the task, looks up what it needs and leaves its summary. In Model you can keep the agent's or pick another one just for that task.
• A script: it runs one of the company's published scripts, always with the same steps and without using AI. It runs with the agent's permissions and its result stays in the history like any other task. If the script changes data, it waits for your approval before running.

Each task can be turned on or off at any time. The agent's supervisor or whoever sets up agents can assign and run tasks.

4. The virtual office

AI › Virtual office shows one desk per agent, live: its avatar, name, job title and role, whether it is Free, Working, Waiting for approval, Paused or in Error, the step it is on, the tasks it finished today, what it cost today and its last steps.

Each agent's detail shows its usage this month against its cap, its tasks and the History: every task with its timeline, tokens, cost and the summary it left. From there:

• Stop a running task. What was waiting for approval is cancelled.
• Pause and Resume the agent.
• Edit its profile or Let go. Letting it go frees its license and keeps its history.

Stop everyone stops every agent in the company at once: none starts or continues a task until you press Resume everyone.

5. Approve what they ask for

When an agent prepares something it cannot do by itself, its supervisor gets the An AI agent asks for approval notice in Nori and in the mobile app. What waits is in AI › Agent approvals:

• Each proposal says which agent prepared it, for which task and what it will do.
• Write a Comment if you want and press Approve or Reject.
• Approving runs it with the agent's permissions, not yours, and the agent continues its task with the result. Rejecting cancels it and the agent is told.
• A proposal nobody decides expires after 7 days.

Only the agent's supervisor or whoever sets up agents can decide. Everything stays in the activity log.

What it costs

• The license: the licenses screen shows how many are taken by agents.
• AI usage: Anthropic bills it to your company's account. In Administration › Integrations › Assistant usage it shows separately as AI agents, and each agent shows its usage this month and today.
• Each task: next to each task you see what it cost on average its last few times.

How to make a task cheaper

• If the steps are always the same, let a script do it. For example, counting today's orders or listing the invoices due tomorrow. A script uses no AI: it costs nothing.
• If it is simple, use the Light model. Summaries, reviews and lists rarely need more. It costs much less than Standard, and the specialists the agent asks use it too.
• Keep Standard or Reasoning for work that needs judgment: preparing quotations, purchase orders or journal entries.
• Write the task precisely: what to review, for which period and what to deliver. The less the agent has to look for, the less it spends.

Security

• The agent only has the assistant's tools, with its role's permissions. It has no access to server files, the database or Nori's code.
• What it reads in your data (a note, an email, the text of an order) is treated as data, never as instructions.
• It is never triggered by what another agent did, so agents do not chain each other.

Permissions

• See the virtual office and approve what agents ask for: to see the office and each agent's detail.
• Create and set up AI agents: to add, edit, pause and let go agents, their tasks and the Stop everyone button.

Any user can open the approvals inbox, but it only shows what that user may decide.