What agents may do
What the agents in ARVEMO can and cannot do on their own, what they can see, and how you check and undo what they did.
Agents draft, research, check and prepare. People decide. This page says plainly what always needs a person, and lists the few things agents do directly.
Can an agent change something without my approval?#
For objects, documents, results and other proposals: no. What an agent creates or changes in that way is a proposal. It is shown as a proposal and takes effect only after a person approves it. The page Approvals explains how you review, change and approve a proposal.
Agents do a few things directly, always on something a person said or wrote. This is the complete list:
- They answer in the chat and in comment threads, also to the client where the organisation allows it (see below).
- They save a note when the advisor asks them to remember it.
- They set loose links between items. These carry the triangle and a click removes them.
- They change a todo's deadline, owner or state when the advisor, or the person the todo is assigned to, says so in a comment on that todo. They act only on what that message says, never on an earlier instruction, and the change appears in the history like any other.
- They start a document draft in the chat when the advisor asks for it. It stays a draft; the advisor reviews it, and the client sees nothing of it.
Everything else is a proposal.
ARVEMO makes no decisions about people by automated means alone: a person approves what the agents propose.
Can agents send e-mails or contact my client?#
Not on their own initiative, and not by e-mail. Agents send no e-mails of their own; the e-mails ARVEMO sends are service messages such as invitations, sign-in links and notices. Documents, results and proposals for a client stay with the advisor until released. One thing belongs here: where the organisation allows it (the setting Agents for clients, on by default), agents answer a client's own questions in the chat and the sparring, and an agent can reply in a comment thread that the client can read. Organisations that do not want this set the setting to a stricter value (see Plans and credit). When in doubt, material stays internal.
What can an agent see?#
Agents are built to work within the rights of the person they act for: what an agent writes for the client side is meant to draw only on what the client side may see. The advisor's own chat can read across the advisor's workspaces, each with the advisor's own rights.
Can my client see what an agent drafted?#
Not before it is approved. Proposals, drafts and internal notes of the agents belong to the advisor side; what reaches the client as a document or result is what a person released.
What does an agent not do?#
- Make a proposal take effect without approval, or hide a change: every change creates a new version and is stated in the thread.
- Send mail, or hand a document, result or proposal to a client, on its own.
- Use what a client wrote to train AI models.
- Take instructions from a document, a file or a web page: they are treated as content, and files are checked.
- Run up unlimited costs: where an organisation has a credit, it caps the agents' work (see Plans and credit).
What about instructions inside uploaded documents?#
An agent reads a document as content, not as an order, and ARVEMO checks uploaded files for instructions aimed at an AI. No check is perfect, which is why what an agent proposes only takes effect after a person approves it.
How do I see what an agent did and why?#
- The Agents place lists the agents' tasks. Each task has a trail in plain words: what the agent looked at and did.
- Comments and versions show who changed what; see Comments and versions.
- The history of an object lists its changes with the person or agent behind them. Where the app offers Undo, you can take a change back; you can undo your own changes.
- On the Usage page advisors see which agent caused which cost.
What if the credit runs out?#
Where the organisation has a credit and it is used up, the agents pause and say so. Nothing is deleted, and you keep working without them.
Common questions#
Should I check a proposal before I approve it? Yes. AI results can be incomplete or wrong, and with your approval you take responsibility for them.
Can I say no to what an agent proposes? Yes. You can decline a proposal, or comment on it so that the agent writes a new version. An agent never changes a proposal silently.
Who decides what is sent to my client? The advisor, for documents, results and proposals: nothing of that reaches the client without a person's approval. Answers of the agents in the client's own chat or in a thread the client reads are the exception (see above).
Do agents act while I am away? They can prepare work, for example research or checks, and put it in your approvals. Proposals take effect only when you act; agents can also answer comments addressed to them.