The three steps
You open a workbook in Excel and tell SparkBench what you need, in your own words. SparkBench shows you exactly what it will send: your description, plus the structure of the table you point at. When you agree, that goes to a remote AI. The AI writes a job, a short program with readable steps, and sends it back. SparkBench runs the job on your computer, inside Excel, on your actual table. The results appear next to your data. The AI does not see them.
That is the entire loop. Everything else on this page is about where the line falls and how you keep control of it.
What the AI receives
Three kinds of material are allowed across the line, and only after you have seen them:
- Your approved task description
The sentence or paragraph you wrote. “Approved” means you read it and chose to send it. It does not mean SparkBench has checked that it contains nothing private; only you can judge that. The documentation shows how to describe a task without putting private facts into it.
- The structure of your data
Table names, column names and column types, so that the job addresses the right columns. Not the values in the cells, and not how many rows there are. A column name can itself be revealing, so SparkBench shows you the structure before it goes.
- Shareable development material
Invented example rows and checks that contain nothing from your workbook. The AI can use them to make sure its job behaves before you run it on real data. These examples are generated independently of your contents.
Nothing else is designed to cross: not cell contents, not results, not error messages, not counts, and not what you learn from a run later in the conversation. The AI cannot interpret observations it never receives; when SparkBench shows you an explanation of a result, that explanation was produced on your computer by the job itself.
What runs here
The job is a small program. SparkBench runs it inside Excel, in its own sealed compartment, with its own clock and budget. It can read the table you pointed at, compute, and prepare an output: a results area, a report, a chart's data. It cannot reach the network. If it fails, the failure is shown to you, in Excel, and nowhere else.
Writing to your workbook is a separate action. A finished job proposes what it wants to write and where; you confirm. Keeping an experiment's result and changing your workbook are different steps, and SparkBench keeps them different.
You stay in charge
- You read what will be sent before it is sent.
- You read the job before it runs. Its steps are written to be read by the person who asked for them.
- You can stop a run. A stopped or failed run says what it did and what it did not do; it does not pretend to have rolled anything back.
- You decide what, if anything, is written into the workbook.
- If you edit a cell while a job is paused, SparkBench says whether the job is still looking at its earlier snapshot or has re-read the sheet. It does not silently mix the two.
An investigation, not a one-shot
The product being built is a small laboratory rather than a formula vending machine. A job can be told to stop at a condition: “pause before the first order we cannot fully fill”, or “stop at the first pupil who falls below the threshold”. At the pause you can ask why, see the state that led there, change an assumption, and continue as a named alternative. Baseline and alternative stay side by side so you can compare them honestly. When you are done you keep what was useful and discard the rest.
The fictional school example on the front page walks through a simplified version of this, entirely inside the web page.
What this does not promise yet
- It is in development. The boundary described here is the requirement SparkBench is built and tested against, in bounded synthetic tests. It is not an audited guarantee of a released product. The status page lists what has been shown.
- Your description is yours. If you type a name, an address or a figure into the task text, it goes with the task. SparkBench shows you what will be sent; it does not promise to catch everything.
- Existing cloud storage is unchanged. A workbook saved in Microsoft 365 or opened in Excel on the web is already stored by Microsoft. SparkBench changes nothing about that, and does not claim that all information stays on one device.
- The AI provider's own terms apply to the material that is sent to it. Which provider, and under whose account, is not yet decided; the administrator notes set out the candidates.
For technical readers
Jobs are small programs in a Lisp-family language, run by SparkBench's own runtime in a web worker inside Excel's add-in environment, with Excel access brokered rather than direct. Source that captures private values is treated as private. The language target, the runtime's limits and the current evidence are in the concepts and administrator pages. The main pitch deliberately stays out of the compiler; the documentation does not.