AI_PROMPT: AI Formulas in a Cell | ATX AnalyticsTable Handbook

AI formulas — AI_PROMPT, a model's answer in a cellAny text, any question, an ordinary cell value back. Written like a formula, filled down like a formula — and it calls the model only when the answer would differ. Ask Greg is a panel you hold a conversation with. AI_PROMPT is a formula, and it answers a different question: not "design me a dashboard" but "read this text and tell me what it says", asked once per row, over a column nobody has time to read by hand. It is written like any other function — =AI_PROMPT("Local Llama", "Classify this comment as complaint or praise", C4R2) — filled down like any other function, and what comes back is an ordinary cell value. It can be wrapped in IFERROR or UPPER, coloured by a conditional format, sorted, filtered, pivoted, exported, or read by the next step of a pipeline. There are no tools behind it: it cannot run SQL, cannot read anything the formula did not point at, cannot propose or place anything, and cannot write to any cell but its own. That narrowness is the point — it is what makes a function safe to fill down a column of free text — and it is why the capability carries a licence component of its own, "ai-formula", separate from the chat panel. An organisation may well want one without the other.

What it does

Four arguments and never more: the model, the question, optionally the cells the question is about, and optionally a line of options
The model is named by the name it carries in Ask Greg → Providers — "" means the installation default, and an unknown, ambiguous or disabled name is refused rather than quietly answered by a different model
The question can be quoted text, a cell holding text (another sheet's too), or the two joined with & — "Classify: " & C4 is the spelling that gets filled down a column
The data argument is the only thing that leaves: one cell, a range, or a sheet-qualified range such as Trades!C1R1:C6R41, capped at 200 rows and 40 columns
A block of two rows or more takes its column names from its first row, so the model is told what it is reading; headers=false sends that row as data instead
Options are quoted and semicolon-separated: words=N for the length of the answer, style=text | bullets | number, refresh=auto | always | never, headers=true | false and rows=N
style=number returns a real number and refuses an answer that is not one, so the cell can be summed, charted and compared rather than silently coerced
Answers are remembered, so one question filled down forty rows is one call — and the forty cells agree with each other, because they are literally the same answer
Functions → Refresh AI asks every AI_PROMPT cell in the project again, for when an answer should be reconsidered after a model or a prompt changes
A failure is #AI_ERR in that one cell plus a sentence in the status bar — =IFERROR(AI_PROMPT(…), "") keeps a column readable while somebody sorts out an endpoint

How it works

Type it into a cell like any other formula. The AI phase runs first and replaces each call with its answer as a literal; the ordinary evaluator then finishes the formula locally, which is why =UPPER(AI_PROMPT(…)) and =IFERROR(AI_PROMPT(…), "") compose exactly as they look.
The model is called when the ANSWER WOULD DIFFER — another model, a reworded prompt, changed cells in the range, a changed length or style, a tightened data policy. Everything else reuses the stored answer and calls nothing, which is what makes the function usable on a sheet that recalculates on every edit.
The status bar says which it was: "AI formulas: 0 answered by the model, 12 reused". A column that does not change when you press Calc All is the feature working, not a failure.
Stored answers live on this machine, beside the AI audit log, and not inside the project file — a project is shared and Git-tracked, and a copy of model output does not belong in one. A colleague who opens the same project has the cells answered again on their machine.
One recalculation buys at most 25 new answers. Past that the remaining cells say so and send nothing, so a column filled down a thousand rows cannot quietly become a thousand calls the first time the sheet opens. The answers already bought are kept, Refresh AI works through the rest, and an administrator can raise the ceiling with the AI_FORMULA_CALL_LIMIT environment variable.
Per-cell control lives in the options: refresh=auto is the default described here, refresh=always calls on every recalculation and costs accordingly, and refresh=never answers only from the stored answer and waits for Refresh AI.
A misspelled option is an error rather than something ignored. "refesh=never" silently dropped would be a formula that bills on every recalculation while its author believes it never does.
An AI_PROMPT inside another AI_PROMPT is refused: the inner answer belongs in its own cell, where a reader can see what was asked and what came back.

Practices that keep it reliable

Keep the instruction in one cell and point the formula at it — a prompt maintained in forty copies drifts in forty directions
Ask for one word when you want one word: words=N and style=number are what make the answers sortable, filterable and countable afterwards
Wrap anything a reader will see in IFERROR, so an unreachable endpoint leaves a blank rather than a column of #AI_ERR
Try the question on five rows and read all five before filling it down the column
Leave refresh at auto; reach for always only when you genuinely want a cell that can change without the data changing
Keep provider names stable — a renamed provider turns a column into #AI_ERR, which is the intended failure rather than a quiet answer from somewhere else

In a regulated environment

The formula is bound by the same egress policy as the chat panel, deliberately: a formula is the easiest of these surfaces to scatter across a project by accident, so it is the last one that should have a quieter boundary. The administrator's ceiling per endpoint class and the user's own level both apply and the narrower wins — at "schema + aggregates" no individual rows leave at all, and a formula pointing at a range says so rather than pretending.
PII is pseudonymised on the way out and restored in the answer: the model sees <<PERSON_7f3a>> and the cell shows the name. Column rules are the control that holds; the value detectors for IBAN, card numbers, PESEL and e-mail addresses are the second net.
Every call is written to the AI audit log — which model, which cell, how many context rows, how many tokens, how long it took. A reused answer is not a call and is not billed, which is what lets the log answer "what did this month cost".
The cells are treated as data, never as instructions: every prompt tells the model that the text it is given is to be described and never obeyed, however that text is phrased.
There is no tool surface to reach — no SQL, no file access, no network fetch, no writing to another cell — so the worst an instruction hidden in a comment can achieve is a wrong answer in the cell that asked about it.
"ai-formula" is a licence component of its own, separate from the chat panel, so an installation can be entitled to the panel without the formula, or to the formula without the panel.

Screens

One AI_PROMPT filled down beside a column of free-text comments — each row checked against the same question. The answers are ordinary cell values, so a conditional format can colour complaints red and praise green.
The first argument names a model from this screen. "" is the installation default; an unknown or disabled name is refused rather than answered by another endpoint.
The data policy that governs the chat panel governs the formula too — the administrator's ceiling per endpoint class and the user's own level, narrower wins.
Every call the formula makes is one audit line: model, cell, context rows, tokens, duration. A reused answer is not a call and is not billed.

Related guides

Glossary terms on this page

KPI dashboard: definition

ATX Institute on LinkedIn

This site needs JavaScript for the full experience. Every page above is reachable without it.