How AI credits are counted

Every AI feature is paid for in AI credits, from your workspace’s balance. This page explains which credits a request uses, how DataSquares counts them and how the numbers you see are rounded. What credits cost, the allowances and the packs are on Seats, plans & AI credits.

Which credits a request uses

A request uses one kind of credit after another, by who asks:

An editor uses the workspace’s included credits, then the credits the company bought.
A viewer uses their own 1,000 a month, which every viewer seat brings, then the credits the company bought while the company lets viewers use them. The included credits come with the editor seats, so they are for editors.
A viewer with no seat’s credits (when a company has more viewers than viewer seats), where the company lets viewers use the credits it bought, uses those.
Ask AI on a shared link or an embed uses the credits the company bought, and stops at zero: a question that uses the last of them finishes, and what it goes past zero is not billed.
Upkeep that runs for no one (building and updating the search index and the AI search catalog, storing the catalog, a scheduled data-quality scan) uses the included credits, then the bought ones, and stops at zero the same way.

Past those, people signed in to DataSquares go below zero while the company has room under its limit: what is below zero when the month closes is billed on the next invoice at the pack price, and buying a pack pays it back first. A billing administrator turns it on or off and sets the limit; see Below zero.

When the credits run out

When nothing is left that a request may use, the request is refused, and the message says what comes back on its own, for the person who asked:

An editor is told the workspace has used its AI credits, and that the monthly allowance comes back at the start of the next month.
A viewer is told You have used your own AI credits for this month, and the credits your company bought are used up too: their own come back next month.
A viewer with no seat’s credits (more viewers than viewer seats) is told You have no AI credits of your own this month, and the credits your company bought are used up: nothing comes back on its own for them.
Where viewers may not use the credits the company bought, a viewer asking for more than they have of their own left is told This needs more AI credits than you have of your own left this month, and your company hasn’t turned on AI for viewers from the credits it bought, and one with none of their own left is told You have no AI credits of your own left this month, and AI for viewers from the credits your company bought isn’t turned on; their own come back next month.
When the company’s limit below zero is what ran out, everyone is told so, and that AI is back when the month turns.

Each is asked to ask an administrator to add more; a billing administrator reads where to buy credits instead.

Two kinds of question still answer, because they need no AI at all, and the answer ends by saying the credits are used up:

A verified question asked word for word (or in one of the phrasings your admins added): its governed metric or card runs again on fresh data (Questions answered at once).
Asking for the last quick answer again (again, refresh, in English): the same governed metric or card, run again as it was. Asking it again for one product, customer or place is a new question, and is refused like any other.

Neither uses any AI credits.

Running low

When few AI credits are left in a workspace, one banner under the page’s header says so, to everyone in it who uses AI:

Members read that credits are running low, with no number, and Ask an admin: one click tells the billing administrators (the company’s admins, and anyone whose company role may buy credits) in a notification that says who asked, in which workspace, and how many credits are left for the person who asked. Asking again from that workspace the same day tells no one again, and the banner says you already asked. Each workspace has its own credits, so asking from another workspace tells them about that one.
People who can buy credits get Buy credits, which opens AI credits on Administration → Billing. While the company is below zero, it says AI keeps working up to the company’s limit and what the next invoice will carry.
When the credits run out, the banner says AI pauses and what keeps working.

Each person’s banner counts the credits their own requests can use (Which credits a request uses): for a viewer, their own credits and then the credits the company bought. So a viewer can read that the credits are used up while editors still have the included ones.

“Few” is a tenth of the workspace’s monthly allowance, and at least 50 credits. Dismiss hides the banner until the credits run out, or until next month; it is remembered in this browser. For people who can buy credits it also comes back when the company goes below zero, as its words change then. Someone who uses AI but can’t see the balance can also ask from Administration → Usage, with Request more credits, in a workspace whose AI runs on AI credits; a workspace whose AI doesn’t has no credits to ask for, so the button isn’t there.

Counted to the thousandth of a credit

DataSquares counts AI credits to the thousandth of a credit. Each AI run is charged for the AI work it did: how much it read and wrote, and how deeply it reasoned at the effort you chose. A small run is charged exactly for its work, never rounded up to a whole credit, and there is no minimum charge per run.

The balance, the monthly allowance, purchased credits, budgets and the history behind the Usage page are all kept to the thousandth too, so small runs add up exactly. What a run is charged doesn’t depend on how it is shown.

Shown in whole credits

Everywhere you read credits — the balance on Administration → Billing and in its AI credits sheet, Administration → Usage, the credit activity, the AI ledger and its CSV export, budgets and the emails about your credits — the number is a whole credit:

Rounded to the nearest whole credit. A half rounds up: 2.5 shows as 3, and 2.499 shows as 2.
“under 1” for an amount above zero that rounds to nothing: a run, a day or a balance with less than half a credit in it reads under 1, so AI work that was done never reads as 0.
0 only when there is nothing there: a run that did no AI work, or a day with no AI use.

A balance that reads under 1, or 1 with less than a whole credit in it, still has credits to use: AI chat keeps answering on what is left, and the request after the balance reaches zero is refused, unless your company lets AI go below zero. How far below zero the company is reads in whole credits too, under 1 included.

Totals are rounded once

A total — a day, a feature, a person, a month, the balance — adds up the exact thousandths first and rounds once, at the end. It isn’t the sum of the rounded numbers listed above it: each of those can be up to half a credit away from its exact amount, so the two can differ. Three runs that each read under 1 can come to a total of 1, and runs that each read 1 can add up to more or less than how many there are. The total is the exact one.

One charge for each chat answer

An answer in AI chat is one charge, however many steps the assistant took to give it: reading your data, running queries, checking its work, summing up its earlier steps in a long build and writing the reply all add to that answer’s single charge. The AI ledger lists the answer once, with the whole amount it used. An answer that was stopped or broke off before all its AI work was reported can have a second, later line with the rest of its credits, a little later: it is part of the same answer, not another run.

Every piece of AI work an answer used is charged once, whatever happened to it. When the AI service had to try a step again, each try is charged for the work it did. When the AI declined a request and another AI model took it over, both are charged, each for the work it did, at the rate of the AI that did it. Work that was done before an answer was interrupted and picked up again is charged once, and the rest of the answer adds to the same charge. Waiting for the AI to be reachable again uses no AI credits.

AI search and its catalog

On a server that turns on AI search, the assistant searches a catalog of a large model’s tables and fields to find the ones a question needs. What it does during an answer is counted in that answer’s line; the catalog’s own lines are listed as AI search on the Usage page and in the AI ledger:

Searches during an answer are part of that answer’s one charge, with the quick AI steps that rewrite the question and pick the tables.
The schema finder, where the server turns it on: its steps are part of the answer’s one charge too, counted in the answer’s own line rather than as AI search. It needs no catalog, so no catalog is stored or charged for.
Building and updating the catalog runs for no one in particular, so it has its own lines: building a workspace’s catalog uses AI credits once, and after that only what changed is sent again.
Storing the catalog is included up to 100,000 documents a workspace; above that it is 1.5 AI credits per 1,000 documents a month, charged a day at a time, one line a day. Administration → AI shows the documents and what they use a month, in whole credits.

Turning AI search off deletes the catalog, and nothing more is charged for it.

Draft with AI in Prep for AI uses AI credits too, one quick AI step for each table it drafts for, listed as Description drafts, and so does Describe with AI in the Catalog, one quick AI step for each dashboard or metric it drafts for, on the same line. A draft that was already waiting, or an asset that already has a description, uses none. Suggest terms to define in the glossary uses no AI.

Explore this model in Prep for AI, where the server turns it on, shows its estimate in whole credits before it starts, and is charged as one line, Explore this model, for the AI work it used, however many steps it took, a stopped one included.

Presentations

Draft from data runs in two steps you start yourself, the outline and then the slides. On a slide, Suggest a headline and Write speaker notes are one step each. Each step shows its estimate in whole credits, with how many credits you can use, before it runs, and runs nothing when its estimate has gone up since you saw it, or is more than what is left of a monthly AI credit budget. Each step is charged for the AI work it used, whether or not you keep what it wrote and when DataSquares leaves out words it refuses, and is listed as Presentations on the Usage page and in the AI ledger.

In the credit activity

Activity, in the AI credits sheet on Administration → Billing, lists credits coming in (the monthly allowance, packs and adjustments) one by one. AI use is listed as one line a day for each AI feature, with the number of AI runs it adds up, rounded once like any other total. A chat answer is one run, however many steps it took, and the credits charged later for an answer that was cut off add to it, not another run. The days are UTC days: credits charged later after midnight (UTC) are on the next day’s line, and the answer still counts as one run, on its own day. The AI ledger has every run.

The list shows the last 90 days. To see more, choose the last 12 months or all time above the list, or choose the last 30 days to see less.

AI in DataSquares: what uses AI credits and what happens at zero.
Seats, plans & AI credits: allowances, packs, auto top-up and below zero.
AI controls: budgets, the AI ledger and its export.
AI search on large models: the catalog behind it, and what it uses.