What an answer costs, and how the ceilings behave

Every figure on this page comes from one screen and two tables: the Overview dashboard at /admin/ai/support, and the ai_support_turn and ai_support_admin_turn rows behind it. The first records what the assistant spent answering your customers. The second records what your own administrators spent asking it questions. Both count against the same ceilings.

The page itself opens on access.ai.support, but every number on it comes from that route. A role with the page key and not the analytics key gets an empty dashboard and no explanation — see Permissions.

What Overview reports

The masthead rail above them carries the live state — which provider is answering, whether the budget still allows a call, and whether the cache has gone bad.

How the two windows are measured

Both periods are UTC, with explicit UTC boundaries: the day runs from 00:00 UTC, and the month from the 1st at 00:00 UTC. That is deliberate — a window derived from the server's local timezone resets at a different instant on every install, and the admin screen then disagrees with the ledger for a few hours every month.

Spend for a window is one SQL sum over both surfaces:

SELECT COALESCE(SUM(costUsd), 0)
  FROM (
        SELECT costUsd FROM ai_support_turn
         WHERE createdAt >= :windowStart AND costUsd > 0
        UNION ALL
        SELECT costUsd FROM ai_support_admin_turn
         WHERE createdAt >= :windowStart AND costUsd > 0
       ) AS spend

There is no status filter. A row carrying a cost is a row somebody was charged for, whatever happened to it afterwards — and that is the whole rule, so a status added later needs no maintenance here.

Skipped turns contribute nothing, but not because of their status: nothing was generated, so costUsd is zero and the costUsd > 0 clause never sees them.

The two ceilings

Both are Super-Admin only and both live on the Guardrails tab of /admin/ai/support/settings. 0 is unlimited on either — a real escape hatch for an operator running a local model whose per-token cost is genuinely zero, not a way of saying "unset". Setting both to 0 is also the only way to switch off the refusal in the next section; one ceiling left standing keeps it armed.

The shipped $5 a day and $100 a month are deliberately low enough to be safe on an install nobody is watching. They are not production numbers, and leaving them is the single most common cause of the next two sections.

Three things happen the first time a customer's message hits a closed gate, and only the first of them is on a screen:

1. The Overview status line reads "Paused: the spend cap has been reached." It says that for every pause, including the unpriced case below, which is the one sentence on the dashboard that can send you looking for the wrong problem.
2. A notification goes to the whole desk — everyone holding view.support.ticket — linking to /admin/ai/support. It is the only signal that names the reason, and it says one of three things: the daily cap was reached, the monthly cap was reached, or spend cannot be measured. Only the first two end with "Raise the cap in Settings to resume"; the third counts the calls that could not be priced and points at the two per-million-token variables in the install guide rather than spelling them out on screen. It is sent once per cap period — the first conversation the cap stops that day, or that month — so the hour after a ceiling trips is not sixty identical alerts to every administrator while the desk is answering everything by hand.
3. That conversation's session moves to AI_SUSPENDED, which is per-conversation and does not lift on its own. It is lifted on the next message the customer sends once the budget allows again — and only for the two cap reasons, so a conversation suspended because spend could not be measured stays suspended even after you have fixed the pricing.

Admin-side model calls hit the same gate and return a 429 rather than escalating. Four of the seven admin-side spenders listed at the end of this page say The AI budget for this period has been reached; the provider connection test and gap drafting word it themselves, and gap drafting is the only one that tells you when the reason was the unpriced case rather than a ceiling. The seventh — the retrieval preview — consults no budget at all and so cannot report one.

When spend cannot be measured, the gate refuses

A cap is only a cap if the spend under it is a real number. Two providers can record a turn that moved thousands of tokens and cost, as far as the database is concerned, nothing:

• The compatible endpoint adapter reports no cost by design. It cannot know: the thing behind it may be a local model, a metered API or a flat subscription.
• Any model id not in the price table — typically one typed in by hand — prices to nothing rather than being guessed at.

So the budget gate counts those turns. A turn is unpriceable when it moved tokens (inputTokens or outputTokens above zero) and recorded no cost; genuinely free work — a skipped turn, a refusal before the call, a direct answer from one of your own articles — moves no tokens and is not caught.

If either cap is above 0 and the current month contains even one unpriceable turn, the gate refuses everything, with the reason unpriced_model. The assistant escalates every conversation and the admin routes return 429, exactly as if a ceiling had been reached. That is deliberate: the alternative is reporting $0.00 every day while the real bill grows at the provider, under a ceiling that was never going to fire.

There are two ways out, and both are decisions only you can make.

Price the endpoint. Set both of these in .env to what your endpoint actually charges, in USD per million tokens, and restart the backend:

AI_SUPPORT_OPENAI_INPUT_USD_PER_MTOK=0.60
AI_SUPPORT_OPENAI_OUTPUT_USD_PER_MTOK=2.40

Both are required and both must be above zero — half a rate card is worse than none, and a rate of 0 would be indistinguishable from "unknown" all over again. Cached input is billed at the full input rate here, because these endpoints do not agree on whether a cached prefix is discounted and over-counting is the safe direction for a ceiling.

Or set both caps to 0. That has always meant "no ceiling", and it is the honest setting for an install running a local model that really is free.

What drives the bill, in order

1. The answering model. aiSupportModel, on the Model tab. It ships as MashDiv AI Core, and on an install running MashDiv AI those three tiers are the whole vocabulary — the tier is a promise about the job, not a pinned model, so what serves it can improve without you editing a setting.

Those are the gateway's sell side, so the local budget gate and the figure on your wallet statement agree. Core is currently on an introductory rate and returns to $5.40 / $27 on 1 September 2026; the schedule is in the price table, so the gate switches over on the day with nothing for you to change.

An install running its own provider key picks from a different list, priced at the vendor's own rates:

claude-sonnet-5 carries a dated change of its own — $3 / $15 from 1 September 2026 — read from the same schedule, on the same UTC day boundary as the caps. Which model serves a MashDiv tier is a routing decision made at the gateway, and nothing on your install asserts one.

A model outside both lists prices to nothing. The agent editor refuses one with a 400, but aiSupportModel written through the settings endpoint is not checked against the list, so a typo there is accepted, answers normally, and records $0.00 a turn — which is what the refusal in the previous section exists to catch. If the assistant stops the day after somebody edited the model, look there first.

2. The two internal models. aiSupportRerankModel and aiSupportClassifierModel both ship as MashDiv AI Lite, the cheapest tier. Reranking runs on every question that retrieves anything and classification runs on every message: they are high-volume and structurally simple, and pointing either at a larger model is the easiest way to multiply your bill for no measurable gain. Both controls appear on the Model tab only on a self-managed install — on MashDiv AI the gateway routes those calls down to the cheap tier itself, so there is nothing to choose.

3. Reasoning effort and answer length. aiSupportEffort defaults to low and aiSupportMaxTokens to 4000. Thinking tokens are billed and count against the ceiling on length, so effort is a real cost lever rather than a quality dial — a two-sentence answer measured on a live install at high effort took 42 seconds, 1,026 output tokens and $0.25. On MashDiv AI each tier has its own effort ceiling and a setting above it is served at the highest the tier allows; on a self-managed install the adapter strips the parameter for models that reject it — claude-haiku-4-5 among them — so raising it there changes nothing but your expectations.

4. Prompt cache reads. A cache read costs a tenth of a fresh input token; a cache write costs 1.25× for the five-minute lifetime, or 2× for the one-hour one. When the cache is working, most of your input is being billed at a tenth.

Prompt-cache health

The Prompt cache tile is not a setting — it is a measurement. The ratio is cacheRead ÷ (cacheRead + input) over the last 24 hours of successful customer turns; unlike the money, it does not read the admin table. Below 0.5 it is called unhealthy. Fewer than ten turns in the window is not enough evidence either way, so a quiet install does not alarm.

Two things watch it for you:

• The Prompt cache tile and its alert banner on Overview.
• An hourly job, aiSupportCacheHealthCheck, described in Scheduled jobs. When the ratio is unhealthy over at least ten turns it writes a warning to the backend log and raises an admin notification — "AI support is costing more than it should" — to everyone holding view.support.ticket, linking to /admin/ai/support.

It runs hourly rather than daily precisely because a day of an invalidated cache is a day of multiplied bills.

The two levers that cost nothing

Neither calls a model, so neither can spend a cent. Deflection is a search over articles you wrote and published, offered while the customer is still typing their subject line, with the file-anyway button never hidden. A direct answer arrives in about half a second, and only on the opening question of a conversation written in English — a verbatim English article is a worse answer, not a faster one, for a customer who wrote in Spanish. In COPILOT it is filed as a draft like everything else, because an install where a person reads every reply does not stop being one just because this reply is your own prose.

Direct answers ship off because they change the voice — the customer reads your prose rather than the assistant's persona — not because they are expensive. On an install with a real knowledge base they are the largest reduction in cost per ticket available, and they are the reason chasing coverage on Gaps pays for itself.

The admin-side spenders

Seven things in the admin area call a model. Five of them now record what they spent, so the ceiling is computed from the whole bill rather than from your customers' half of it.

The Handbook screen is two routes wearing one face: it fires the console route when the operator assistant is switched on, and the documentation-only handbook route when it is not. Both are billed and both are recorded, but only the first keeps a conversation you can reopen.

All but the last check the budget before spending and refuse with a 429 when it is reached. The operator assistant books its cost onto the conversation you were having, so you can read it turn by turn. The other four have no conversation to book against, so each files its turn under a per-administrator, per-surface ledger row — "Handbook questions — 41 turns, $3.12" — which is what makes their spend readable rather than merely counted. Those ledger rows are listed apart from real conversations in the console's history, and deleting a conversation clears its text without giving its cost back. See Your own assistant.

Generating a draft from the Live Inbox spends too, but it runs the same engine a customer's message runs and writes an ordinary customer turn — so it is already in both the spend tiles and the turn counts, and it is not in the table above.

Two more worth knowing before you hand out the permissions:

• Teach from this ticket is the most expensive single call an admin can make in this addon: the answering model over a whole transcript at 1,500 output tokens, reachable by anyone with create.ai.support.knowledge, on as many tickets as they care to click.
• Test connection is not a reachability ping. It runs the whole path a customer's question takes, ending in a real generation, and bills a fraction of a cent for it — on the one screen an operator opens repeatedly because the assistant has gone quiet.

Where to look when the bill moves

1. Prompt cache on Overview. If it has fallen, nothing else matters until it is fixed.
2. The average per answer under the month tile. A jump here with flat volume means a model, an effort level or a max-tokens change.
3. Spent this month against that average × the turn count. The gap is what your own administrators spent. If it is large, the console is your bill.
4. What it did (30 days). Spend rising while Answered is flat means you are paying for refusals.
5. The Model tab, for the two internal models. They are the change nobody remembers making.

More failure shapes are in Troubleshooting.