When assistants can edit the file, the review gate becomes part of the skill
Apple says new productivity assistants and MCP access will let AI act on document context; the capability shifts training from prompting alone to scoped edits, review and recovery.
What happened
Apple announced on 27 October 2026 that beta assistants in Pages, Numbers and Keynote will understand document context, while MCP support on macOS will let third-party agents read content, make edits and execute automated actions.
Why it matters
Once an assistant can change a working file, competence includes defining edit scope, recognising high-risk changes, reviewing diffs and restoring a trusted version.
The capability shift
Apple says beta assistants in Pages, Numbers and Keynote will understand document context. It also says MCP support on macOS will allow third-party agents to read content, make edits and execute automated actions. The Verge independently reports announced writing, calculation, organisation, design and animation actions. Availability is planned for November and depends on compatible products and systems.
This is an announced capability, not evidence of accuracy or productivity in a particular workplace. But it changes the training object: a response in a chat box can be discarded; a mutation to a working file must be scoped, reviewed and reversible.
Train the edit cycle
Design exercises with four observable steps. The user first states which objects may change and which must remain fixed. The assistant then produces an edit or proposed action. The user reviews a diff or another evidence view, checks formulas, sources, layout and downstream links, and either accepts, narrows or rejects. Finally, the user demonstrates undo or restoration from a trusted version.
Separate low-risk formatting from consequential content, formulas, permissions and external actions. A learner who writes an elegant prompt but cannot spot a changed denominator or recover a file has not demonstrated safe competence. Records should show who approved the final state, not merely who initiated the assistant.
Decision rule
Do not enable a write-capable integration for a role because staff completed generic prompting training. Require a role-specific exercise covering an allowed edit, a forbidden edit, a misleading but plausible change, a clean rollback and escalation. This turns the earlier agent write-boundary evaluation into a user-level skill gate.