Ask Greg, the AI Analyst | ATX AnalyticsTable Handbook

Ask Greg — the AI designer and analystDescribe what you want. It proposes objects the project already knows how to build — and every one has been run before you see it. Ask Greg is a ribbon tab and a panel docked beside the grid, where you describe what you want in words and the application answers with objects it can actually place in the project. The idea it is built on is a single sentence: the model never writes code and never writes to the project. It writes proposals — typed objects the application already has a builder, a validator, a renderer and a persister for — and every proposal is executed against your real data before you are shown it. An assistant that emitted JavaScript, or SQL that ran unattended, could not be reviewed, undone cleanly or audited. One that emits the same chart configuration the Insert Chart button produces is a structured suggestion, validated by the same validator and undone by the same Ctrl+Z.

What it does

A docked panel, not a modal — the sheet stays visible and your selection becomes context
Read tools that run immediately against your own database: list and describe tables, profile a column, run SELECT-only SQL, read a sheet range, list existing objects, preview a pipeline
It sees the automation already in the project — pipelines, PDF reports, alert rules and schedules — and can read and explain any of them, including the ones you built by hand
Proposals for the sheet: saved and pinned queries, charts, pivots, KPIs, dashboards, blocks of computed cells and value lists — and a whole new project written to its own folder
Proposals for automation: a one-time Transform on existing cells, a new saved pipeline of up to 20 steps, one step added to any existing pipeline without touching the others, a saved PDF report of up to 40 blocks, an e-mail alert rule, and a schedule
Pipeline steps it can use: sort, de-duplicate, merge columns, value mapping, text cleaning, PII masking, drop or keep columns, replace anomalies, formatting and conditional formatting, read-only SQL and text to columns — plus four that reach outside the table: AI analysis, send e-mail, export a file and import a file
Schedules for a pipeline, an alert rule, a query or pinned-query refresh, or a PDF export — every N minutes (15 at the least) or at a set time on chosen weekdays, in a named time zone, with retries
Linked proposals in one answer — a chart, a report that prints it and a schedule that exports the report every weekday arrive together, and Place all lays out the whole set
Every proposal certified gold, silver or bronze after a dry run against your real data; rejected ones are never shown at all
Everything it placed carries an Ask Greg tag with the question that produced it, and the automation a card placed can be taken back with Remove
Analysis tools that compute rather than guess: anomalies against a series' own trend and season, forecasts from the forecast chart's own model, period-over-period change with each group's share, cluster analysis and key drivers — each answer able to propose the matching chart (chapter 26)
Governed metrics first: search, describe and query certified definitions from the semantic layer, with a Certified, Governed or Unverified badge under every answer worked out from the tools that actually ran (chapter 27)
Ask Greg → Ask → Data story writes a checked narrative of the active dashboard (chapter 27)
Per-session model choice, with the panel showing the model and the data level that endpoint is allowed to see

How it works

Ask in the panel — a fresh session offers starting points such as "Summarise what is in this project" or "Which values in the current sheet look wrong?". Ask for automation the same way: "clean this range every morning and e-mail me the result", "add a step that masks the customer names", "warn me when yesterday's volume drops below 1,000".
It answers from work, not recall: the model does not say revenue is concentrated in one region, it runs the query and reads the number. Every query it ran is listed under the answer so you can check it.
Each proposal arrives as a card with a certification badge. Gold is earned two ways and neither is the model's opinion of itself: either the model stated what the result would be before anything ran and the prediction matched, or a second, independent route recomputed the same numbers and agreed.
Silver means it ran and returned a plausible result but there was no independent route to check it against. Bronze means it was built and validated but could not be executed here — a missing connection, a file that does not exist yet, or a picture that needs a live browser.
Automation cards show evidence rather than a picture. A Transform shows the rows Before and After; a pipeline shows the result of every step; an added step shows the pipeline run with and without it; an alert says whether it would fire now, and why; a schedule lists its next three runs, computed with the scheduler's own arithmetic. Nothing is sent, written or exported during the check — the card says so in as many words: "Would e-mail … when it runs — nothing was sent now."
Every card also has "What was checked" and "What this does not prove", so the limits of the dry run are stated next to its result.
Place… asks only what that object needs — a tab and an anchor cell, a dashboard, a range, a name in a library, or a folder — and says what placing does: a pipeline is saved in Transform → Pipelines and "placing it does not run it"; a report is saved in Export → PDF report and "placing it does not print it". Alerts and schedules show every recipient and an on/off switch before anything is saved, and can be placed switched off.
Sheet objects land in one step on the existing undo stack, so Ctrl+Z removes the whole placement, and a Transform is undone the same way. Pipelines, added steps, reports, alerts and schedules get Remove on their card instead, which takes back only what nobody has edited since.
What it places is saved in the project exactly like something built by hand — so under Git it appears as an ordinary change you can review, carrying the Ask Greg marker with the model and the question.

Practices that keep it reliable

Read the badge before you place. Gold means a second route agreed on the numbers — not that the analysis was the one worth doing
Read "What this does not prove" on an automation card: a dry run shows what would happen with today's data, not that the process is the right one
Place a new alert or schedule switched off, run it once by hand, then switch it on
Open the Context disclosure the first few times: it prints exactly what was sent, after masking, and which chunks were retrieved
Treat a picture as never gold — certification proves the numbers behind a chart, never the drawing
Ask Greg never runs anything by itself. A schedule it proposes runs a pipeline, an alert, a refresh or a PDF export — never the assistant. Scheduled briefs and saved prompts are on the ribbon but not yet available

In a regulated environment

The model cannot act. It emits proposals and every write goes through a human pressing Place, so a prompt injected into a data cell can at best produce a card nobody asked for — sitting next to the others, with a certification badge.
E-mail recipients are decided by the application, not the model: an address must have been typed by you in the session, be the mail profile's own address, or already receive an existing alert. Addresses found in the project's data are never used, and the model only ever sees how many recipients there are.
Steps that reach outside the table are fenced: an export writes only files inside the project folder, an import reads only files inside it, an e-mail step attaches only the pipeline's own table, and reading a mailbox is not available to Ask Greg at all.
Every piece of SQL in a proposal — a pipeline step, a report block, an alert source — is held to the same SELECT-only contract as the assistant's own queries.
Each automation kind is licensed on its own: placing needs ai-agent-actions plus the module it creates — the Transform module for transforms and pipelines, Export for PDF reports, email-alerts for alert rules, scheduler for schedules. A missing one is refused before anything is written.
Egress policy is enforced in one server-side function, set by an administrator and seeded from the installer so a fleet is configured once. Users may narrow it; users may never widen it. Public-cloud endpoints ship capped at schema and aggregates.
Masking runs before anything leaves the machine: column rules first, then value detectors, then deterministic pseudonymisation — the same name becomes the same token everywhere, so joins and group-bys still work, and tokens are restored locally in the answer. You read real names; the model never saw them.
One append-only audit line per model call and per placement: user, model, egress level, mask counts, prompt hash, tool calls, tokens, proposal tiers and what was placed — and, for alerts and e-mail steps, each recipient with the rule that allowed it. Prompts and completions are not written by default.
The action catalog has no delete, no shell, no arbitrary fetch, no credential access and no DDL. It cannot edit or remove existing steps, rules, reports or schedules, or create connections, mail profiles or credentials — there is nothing for an injected instruction to reach.
With an on-prem endpoint and the shipped manual and capability vectors it works with no internet at all — the configuration the design is optimised for.

Screens

Ask Greg — asking, grounding, models and governance, plus an Automation group whose scheduled briefs and saved prompts are shown but not yet available.
The panel docks beside the grid so the sheet stays visible. The header names the model and the level of data it is allowed to see.
Model providers — managed endpoints ship with the installer and cannot be repointed from inside the app.
AI audit — what was asked, what was sent, what came back, and what was applied.

Related guides

Glossary terms on this page

ATX Institute on LinkedIn

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