There is no customer deployment of SparkBench yet. This page records intentions and the current evidence so that administrators can raise questions early. Prerequisites marked pending will be stated exactly when they are known.
Target hosts and evidence
| Host | Target | Evidence so far |
|---|---|---|
| Excel for Windows | Yes | Bounded synthetic runs: job execution, cancellation, termination, recovery, pause and intervention, edit-while-paused detection. |
| Excel on the web | Yes | Bounded synthetic runs: job execution, cancellation, forced disposal and recovery. |
| Excel for Mac | Yes | None yet. Hardware for actual checks has not been available. |
| Excel on iPad or mobile | Not planned for the first release | None. |
Shape of the deployment
- Add-in origin
The add-in's web assets are planned to be served from a dedicated origin, separate from this website, so that marketing pages never run inside Excel and the workbook-authorised application loads nothing from here. The planned hostname is
addin.sparkbench.io; it is not live.- Where jobs run
On the user's device, inside Excel's add-in environment, in a web worker with brokered workbook access. No workbook contents are sent to a SparkBench server or to the AI.
- Remote AI access
-
Pending decision. Two routes are under consideration and neither is selected:
- AI resources owned by your organisation in your own cloud subscription, so that model usage is billed and governed under your existing agreement. This requires administrator consent and resource permissions in your tenant.
- AI hosted by the publisher, with usage recovered through the publisher's commercial terms.
In both routes, the material sent is limited to the approved task text, data structure and shareable development material. Model choices and regions are also pending.
- Distribution
Microsoft recommends that organisations deploy add-ins centrally through Integrated Apps in the Microsoft 365 admin center; see Microsoft's deployment guidance. Whether SparkBench will be offered through Microsoft's store, by direct deployment package, or both, is pending. No sideload instructions are published.
- Permissions
The add-in's declared Office permissions and any identity or tenant consent requirements are pending and will be stated verbatim from the production manifest when it exists.
- Telemetry
Pending decision. The design rule is that private-derived results, faults and counters are never telemetry. Whether any operational metrics are collected at all, and what they contain, will be decided and documented before release.
- Updates and removal
Pending. Add-in update and removal follow the chosen distribution route; this will be documented with it.
Questions to put to us
If you evaluate SparkBench for an organisation, these are reasonable things to ask before anything is deployed; the answers that are still pending will be given as pending.
- Which AI provider and account route will be used, and under which data-processing terms.
- The exact list of material that crosses to the AI, and how a user reviews it.
- How the local/remote boundary has been tested, and by whom.
- Which Excel versions are supported, and what the add-in manifest declares.
- Whether any telemetry exists, and where it goes.
Contact details are on the support page.