SparkBench

Describe the task. AI writes the job. SparkBench runs it here, on your data.

An Excel add-in being built for people who know their spreadsheet but do not program.

You say what you need in plain words. A remote AI writes a small, readable job. The job runs on your computer, inside Excel, on your workbook. The AI gets your description and your table's structure, not your rows.

See how it works Read the documentation
How a SparkBench task flows Your computer on the left holds the workbook. A note with the task description crosses a dashed seam to the remote AI on the right. A job sheet comes back across the seam, runs on the workbook, and result rows appear on the left. The right side stays empty apart from the note. Your computer Pupil ID Class Term 1 Term 2 Term 3 Results, written here Remote AI
Illustrative scene with a synthetic workbook. The description crosses; the job returns; the results stay here.

SparkBench is in development. There is nothing to install yet. This website explains what is being built, what has been shown to work in bounded tests, and what has not. Read the current status.

Three steps, and one line drawn by design

The whole idea fits in a sentence. The detail that matters is which side of the line each thing is designed to live on.

  1. Describe the need

    Write what you want in your own words, such as “find pupils whose reading score fell two terms in a row”. The description and the names of your columns are what leaves. You read exactly what will be sent before it goes.

  2. AI writes the job

    A remote AI turns that into a small job: a few readable steps saying what will be read, compared and written. It arrives on your computer for you to look at before anything runs. It has not seen a single row.

  3. SparkBench runs it here

    The job runs inside Excel on your machine, on the real table. Results, explanations and any errors are produced and shown locally. They do not go back to the AI, and they do not become part of a later chat.

What crosses the line, and what stays

SparkBench acts as the AI's local representative, with your authority. The AI gets what it needs to write a job and nothing it does not.

Stays on your computer

  • Cell contents Every value in your workbook, in every sheet.
  • Results and reports Whatever the job produces, including counts and summaries. A count can reveal values too, so counts stay here.
  • Faults and error messages A failing job can expose data in its error text, so errors are shown to you, not sent on.
  • Later conversation What you learn from a run does not flow back into the chat with the AI.

Goes to the remote AI

  • Your approved task description The words you wrote, after you have seen them and agreed to send them.
  • The structure of your table Table and column names and their types. Not the values in them.
  • Shareable development material Invented example data and tests that contain nothing private, so the AI can check that its job works before you run it.

This is the design requirement SparkBench is built and tested against. It is not yet an audited guarantee of a finished product. Two honest limits: your description can carry private facts if you type them in, so SparkBench shows you what will be sent and the documentation explains how to describe a task without them; and a workbook already stored in Microsoft 365 or opened in Excel on the web is already held by Microsoft, which SparkBench does not change.

A fictional school, worked through

Illustrative demonstration with synthetic data. Everything runs inside this page: no AI is called, Excel is not involved, and nothing is sent anywhere.

Hazelmere Primary is invented, and so are its twelve pupils, identified only by made-up IDs. A teacher wants to know who is slipping. Walk through the three steps and then change an assumption to see how an investigation feels when the data stays on the bench.

The interactive demonstration needs JavaScript. In short: the description and the five column names cross to the remote side; a pre-written job returns; it runs on twelve synthetic rows and flags four pupils whose score fell in both terms and is now below 60. The rows, the results and the counts stay on the local side.

Not a formula generator. A place to investigate.

A job is a starting point, not an answer to accept. SparkBench is being built so that you can stop a job on a condition, ask why, change something and compare, all without the data leaving the room.

  1. Ask

    Put the question in your words.

  2. Model

    The AI writes a job you can read.

  3. Run

    It runs here, on the real table.

  4. Observe

    Pause on a condition; see the state.

  5. Intervene

    Change an assumption or a cell.

  6. Compare

    Baseline and alternative, side by side.

  7. Keep

    Write back only what you choose.

Where this stands today

Dated 7 November 2026. The status page keeps the longer list.

Shown to work in bounded tests

  • Small jobs written as source run inside Excel on Windows and in Excel on the web, with cancellation and clean recovery, on synthetic workbooks.
  • A job can pause on a condition, explain its state, accept a named change and compare two scenarios, in a synthetic test story.
  • Editing a cell while a job is paused is detected; the job either keeps its snapshot or re-reads explicitly, and says which.

Not yet

  • A live AI connection in the product. Development so far uses a stand-in that writes nothing on its own.
  • Anything to install, a store listing, prices or dates.
  • Evidence on Excel for Mac, and general workbook workflows beyond the tested stories.
Read the documentation Support and contact