Proposals and limits

The assistant proposes changes. You approve them. Nothing in between writes anything.

The approval card

When the assistant works out what needs changing, it does not change it. It shows you a card listing every record it wants to touch and what it wants to do to each one.

• Apply runs the whole batch and reports what happened to each row.
• Discard throws it away and nothing is written.

Until you press Apply, your data has not changed. Closing the panel leaves the proposal unapplied, and it expires on its own.

A proposal expires 30 minutes after it appears. If you come back to a stale one, ask again and you get a fresh proposal against your data as it is now, rather than as it was half an hour ago.

If a batch is partly applied, the card says which rows went through. That is the honest answer when something fails halfway, and it is the one that tells you what to retry.

What it can reach, and what it cannot

The assistant does not have the run of the app. It has a fixed list of things it can do, and the list is what it is rather than what it can talk its way into:

It can read your projects, contacts, companies, people, CVs, tasks, activities, tracked time, travel and files.

It can propose changes to projects, contacts, companies, people, CVs, tasks and time entries.

It can answer questions about how Steerd works, from this help site, for every part of the product. That includes the parts missing from the two lists above: it cannot touch an invoice and it can still explain invoicing. Answering costs no approval, because nothing is written.

There is nothing else. No tool sends mail or a message on your behalf. None touches billing, plans or payment details. None changes team membership, roles or permissions, and none can read or write a credential. Those are not blocked by a rule you could argue with; there is simply no way to ask for them.

Within that list the assistant has exactly your permissions and never more. If your role cannot edit a colleague's tracked time, neither can the assistant while you are using it, and it says to ask an admin rather than quietly doing nothing.

The daily budget

Each team has a daily allowance of AI usage, counted over a rolling 24 hours rather than reset at midnight. It exists because the key is yours: a job that loops without anyone watching spends real money on an account we do not administer, and the first you would hear of it is an invoice.

If you reach it, the panel says so. The allowance comes back gradually as the last day's usage ages out.

Your provider has its own limits, and hitting one of those reads differently: it clears in a moment rather than over a day.

One case is worth knowing about. If your endpoint does not report token counts, Steerd charges the allowance a worst-case estimate rather than nothing, so it runs down faster than your provider's own dashboard suggests. The named providers all report them; this only comes up on a custom endpoint.

Long jobs stop rather than guess

A single request gets a bounded amount of work. If the assistant runs out before it finishes, it says so and stops instead of pretending it got there.

That is a signal to ask for a smaller piece: "archive the lost projects from last quarter" rather than "tidy up my whole pipeline".