AI usage and models
What this workspace spends on AI models, the monthly ceiling on it, the credit balance it spends past the plan, and which model each kind of work is sent to.
What it is
Every model call the product makes for this workspace is recorded — which feature asked, which model answered, how many tokens it took and what it cost. This screen is that record, plus the three controls over it: a monthly ceiling, a credit balance, and a per-task-class model policy.
The money here is what the AI gateway charges, in US dollars, whatever this workspace bills its own clients in.
How to get there
Money → AI spend in the admin console rail, at /<workspace>/admin/ai.
Reading everything on the page needs billing.view. Changing the ceiling, the model policy and the credit rules needs billing.manage — the same permission as changing the plan. Without it the fields render read-only and say so.
The fourth AI control, switching a named person's AI off, is not here. It is a pane on Security behind security.policy, because whoever configures single sign-on is not necessarily whoever signs for the bill.
How to use it
Read the month — the Spend tab
- Four tiles: Spent this month, Calls, Tokens and the Budget state.
- The spend tile reads Spent this month (at least) when the month contains calls on a model no price is held for. Those calls are absent from every total on the page, because a zero would read as free.
- The meter shows spend against the ceiling. A workspace with no ceiling gets a sentence instead of a bar — a bar at 0% would read as good news.
- Where it went slices the month by feature, by model, by agent or by kind of work. The bars are drawn against the widest slice, not against the total.
- Failovers lists answers that came from a model other than the one first chosen. Refusals lists why calls did not happen. Both render at zero on purpose — an empty list is the difference between "none happened" and "we do not record that".
Set the ceiling — the Budget tab
- Ceiling (USD) is what this workspace may spend in a calendar month. Leave it empty for no ceiling.
- Warn at is the percentage at which the holders of
billing.manageare notified, once per period per threshold. Setting it to 0 switches the warning off, and the screen says so rather than showing "warn at 0%". - Refuse calls past the ceiling is the hard stop. With it off, the ceiling is a warning and spending continues.
- Save budget. If the workspace has already spent past the number you typed, the page says so — the ceiling takes effect on the next call rather than undoing what has been spent.
- Remove the ceiling clears the number and keeps the warning threshold and the hard-stop setting, which is why it is not called "remove the budget".
Buy and spend credit — the Credits tab
- The plan allowance is denominated in tokens because it is a package limit. The balance is denominated in dollars because it is money. They are not converted into each other — a million tokens costs wildly different amounts depending on the model.
- One sentence says exactly where the workspace stands: inside the allowance, running on credit, nearly out, out, or set not to use credit at all.
- Buy credit lists the packs, the bonus on each, and the price. Credit does not expire and is only used once the plan allowance runs out. Buying opens checkout in the same tab and returns here.
- Use credit when the allowance runs out is on by default — a balance somebody paid for that then refuses to be spent is not a feature.
- Buy credit automatically is off by default, because a charge nobody clicked is the one thing here that can surprise a customer. It is bounded three ways — a pack, a monthly cap and a saved card — and the switch stays disabled with the reason next to it until all three are in place.
- Statement lists every movement on the balance, most recent first.
Choose a model — the Model policy tab
- Routing is by kind of work, not by feature: classification, extraction, summarising, conversation, reasoning, client-facing writing, embeddings and transcription. A feature is not one kind of work, and a firm's opinion is about the class.
- Each card shows What this uses — the ordered chain, first choice highlighted — and What we would choose underneath.
- Pin to sends that class to a specific model. Only models that class can actually be pointed at are offered.
- Allow a fallback appears once a class is pinned. Turning it off means that kind of work is down rather than answered by another model, which is the right choice for a firm that pinned for a residency or compliance reason.
- Two notes are not warnings. Embeddings and transcription use one model by design and have no fallback to configure. A pin that names the model we would have chosen anyway freezes that class on it, and the card says so.
Check the plumbing — the Provider tab
Whether the gateway is connected, what it can do, and the voice catalogue. It distinguishes "never configured" from "configured and unreachable", because those are different people's jobs.
What it affects
- Guards run in a fixed order. A person's AI seat being switched off comes first and holds regardless of money — including for agents that person owns. The workspace's own budget comes second. The plan's
ai.tokensallowance comes last, because it is the softest and its refusal is an upgrade prompt. - Every refusal is written to the ledger before it is thrown, so "why did nothing happen" has an answer.
- A ceiling with the hard stop on refuses calls on unpriced models too. A cap that stops what it can count and waves through what it cannot is not a cap. On that page the same count means the opposite thing, and the wording changes with it.
- Reaching the allowance is not the end if the workspace holds credit and draw-on-credit is on. Calls then debit the balance at what they actually cost, with the platform markup applied, and the markup percentage is stated on the balance card.
- Agent runs are on the same meter. A run stops when the workspace's money runs out, not only when the agent's own per-run token cap does, and the budget is checked before every model call.
- Model policy is resolved per call and never cached, so a policy changed after an incident takes effect immediately.
- A stream fails over before the first byte and never after. Once the answer is partly on screen, splicing in a second model's prose is not a recovery.
On mobile
Admin → AI carries Spend, Budget and Models. The ceiling is editable there, because the budget warning arrives as a push and the ceiling is what the person holding the phone came for. Pinning a model is web only — it is a spend decision made against a price table, and what belongs on a phone is knowing what is pinned when somebody asks.
Limits and gotchas
- The month resets on the calendar month. Only priced calls are counted.
- A ceiling of zero is refused, by the form and by the database. That is an off switch, and a screen labelled "budget" is not where anybody looks for one.
- Prices are held for the handful of models this product actually calls. A call on anything else records no cost rather than a zero — and under a hard cap it is refused instead.
- There is no per-person usage view. The ledger holds it so that a specific investigation can run; the breakdown deliberately does not turn that into a leaderboard.
- A workspace cannot cap tokens, calls or runs separately. The only budget is money.
- There is no per-agent model policy beyond an agent's own model preference.
Related
- Billing — the plan, the seats and the invoice.
- Agents — the unattended actors that spend this meter.
- Security — switching a named person's AI off.
- Plans and entitlements — what
ai.tokensis. - Nia — the assistant most of this spend belongs to.
