Notion AI credits are a shared workspace pool, not a per-person allowance

Credits sit at the workspace level. There is no per-person cap, no departmental split, and no built-in way to say "marketing gets 40 percent". Any Custom Agent in the workspace draws from the same balance, regardless of who created it or who triggered it.

For a five-person team that is fine. For a forty-person workspace with three departments, it means your finance agent can be starved by somebody's content experiment, and the first signal you get is the agent failing to run.

Credits reset at the start of each billing cycle. If you are on a monthly subscription, credits are billed alongside it. If you are on an annual plan, credits are still billed monthly and separately, which trips up finance teams who expect one annual line item and get twelve small ones.

Practically: your usable balance is a monthly figure even if everything else about your Notion contract is annual. Budget it as opex, not as part of the seat renewal.

Unused credits expire at the end of the cycle. They do not bank, and there is no credit-note equivalent. Buy 3,000 credits for a month where you only use 900 and you have simply spent the difference.

The correct response is not to buy generously "to be safe". It is to buy close to your real usage and use the admin control that lets you adjust the purchased amount up or down each cycle. Under-buy slightly, watch what breaks, and raise it. That costs you a little friction once. Over-buying costs you every single month, quietly, forever.

List them. For each: what triggers it, how often it fires, what it reads, and whose name is on it. Turn off anything nobody can justify in a sentence. In most workspaces this alone removes a third of the load, because experiments never get switched off.

An agent that runs on every page edit will cost many times more than the same agent running once a day on a filtered view. Triggers are where credit bills are made and unmade. Ask of every agent: does this need to react instantly, or is a scheduled batch fine? Almost always, batch is fine.

Point agents at a specific database or filtered view, never at a whole workspace or a sprawling parent page. The more context an agent has to chew through to find the answer, the more steps it takes. This is also why database properties beat free text: a property is a lookup, a paragraph is a search.

If a step is a rule, it does not need a model. "When status changes to Won, create the project page" is a native Notion automation, not agent work. Reserve the agent for the judgement calls in the middle, and let plain automation handle the parts either side of it. This is the single biggest structural saving available, and it makes the workflow more reliable at the same time.

Because credits do not roll over, the correct buying posture is slightly under your expected need, adjusted upward when you actually hit the ceiling. Put a 10-minute review in the calendar for the day after each billing date. Look at what you consumed and set next cycle to match.

If you want to estimate before you commit, we built a free Notion AI credit calculator that turns your agent count and run frequency into a monthly credit figure.