The boundary
SparkBench is the remote AI's local representative, acting with your authority. The AI receives what it needs to write a job; the job runs where the data is. In detail:
Stays local
- Cell contents All values, in the table you point at and everywhere else.
- Results derived from them Output rows, summaries, charts' data, and counts. A count of matching rows can reveal values, so counts are local too.
- Faults, reports and traces Error messages and run reports can quote data. They are shown to you only.
- Later chat Observations from a run do not become context for the next message to the AI.
- Source with captured values If a job's source ends up containing a value from your sheet, that source is private, whatever it is called.
Crosses, after your review
- Task text Your description as written.
- Structure Table and column names and types.
- Shareable development material Invented examples and checks with nothing from your workbook.
This boundary is the design requirement SparkBench is built against and tested with synthetic data. It is not yet a verified guarantee of a released product, and SparkBench does not describe it as one.
Describing a task without leaking
The task text goes to the AI word for word. The structure goes name for name. Both can carry private facts if you put them there. Three habits keep them out:
- Name columns, not people
-
Say “pupils whose Term 3 score is below the threshold”, not “pupils like Ola in 4B who scored 45”. The first is a rule the job can apply; the second is a disclosure.
- Give conditions, not values from the sheet
-
A threshold you choose, like 60, is fine. A figure copied out of a cell, like a particular pupil's score or a customer's balance, is not. If a rule depends on a value in the sheet, name the cell or column and let the job read it locally.
- Check your column names
-
Column headers travel as structure. A header that is a person's name, an account number or a diagnosis is private. Rename the column in your sheet before describing the task, or point the task at a copy with neutral headers.
SparkBench shows the exact task text and structure before anything is sent, and the front-page demonstration shows the same view with invented data. Approval means you have looked; it does not make private content safe.
What the AI can and cannot do
- It can write a job, propose revisions and explain the job's logic, from your task text and structure.
- It can test its job on invented example data that contains nothing of yours.
- It cannot see your rows, your results, your error messages or your counts, so it cannot interpret them. When you see an explanation of a result, the running job wrote it here.
- It cannot write into your workbook. A job proposes output; you confirm.
What SparkBench does not change
- Where your workbook already lives. A file in OneDrive or SharePoint, or opened in Excel on the web, is already stored and processed by Microsoft under your organisation's agreement. SparkBench does not move it, copy it or alter that.
- Excel's own behaviour, including any diagnostics Microsoft collects from Office, which SparkBench neither adds to nor removes.
- Your organisation's policies. If your administrator restricts add-ins or network access, SparkBench works within that or not at all. It does not offer a way around tenant controls.
- The AI provider's terms for the material that is sent to it. Which provider, and whose account, is pending; see the administrator notes.
How this is being verified
The boundary is written as a requirement and exercised in synthetic tests: capturing inputs and reports in bounded Excel sessions, checking that private-derived material stays on the local side, and refusing jobs built on stale inputs. These tests use invented workbooks and a stand-in for the AI. They are evidence about the design, not an audit of a shipped product. The status page states what has been shown.
The privacy policy is the formal statement covering this website, support correspondence and the planned add-in and service, with its unresolved facts marked.