Feature Request: Secure Cross-Application Handoff Between ChatGPT Agents in Excel, PowerPoint, and Word

ChatGPT Feature requests
, , , ,
1

I would like to propose a secure, permission-aware interoperability layer allowing ChatGPT agents embedded in Microsoft Excel, PowerPoint, Word, and potentially Outlook to communicate and hand off work to one another.

Today, a user may perform a complete analysis with a ChatGPT agent in Excel, but the PowerPoint agent cannot directly receive the resulting tables, charts, assumptions, cell references, analytical context, or presentation instructions.

The user must manually copy and paste information between applications, which breaks provenance, increases the risk of inconsistencies, and forces the user to explain the same context repeatedly.

Primary user story

A user completes a financial or operational analysis in Excel and asks:

“Send the selected analysis to the ChatGPT agent in PowerPoint and create a 10-slide management presentation.”

With the user’s explicit approval, the Excel agent should be able to transfer:

  • selected tables and cell ranges;

  • Excel charts or chart specifications;

  • calculated KPIs and model outputs;

  • source worksheet and cell references;

  • formulas, values, units and number formats where authorized;

  • assumptions and explanatory notes;

  • a structured executive summary;

  • presentation instructions, audience and desired tone;

  • links back to the source workbook;

  • a version identifier or timestamp for synchronization.

The PowerPoint agent could then generate a proposed slide structure, create the presentation and maintain traceability between each slide element and its Excel source.

Current workflow and limitations

The current workaround is to copy tables and charts manually, export images, or write a separate prompt describing the Excel analysis.

This approach has several limitations:

  1. Context is lost between applications.

  2. Cell-level source references are not preserved.

  3. PowerPoint content can become outdated after the Excel model changes.

  4. Users must repeat instructions already understood by the Excel agent.

  5. Manual copying creates formatting and transcription risks.

  6. There is no visible audit trail showing what information was transferred.

  7. Sensitive workbook content may be copied more broadly than intended.

Proposed product behavior

Each Office application could expose an explicit action such as:

  • Send to PowerPoint

  • Continue in PowerPoint

  • Create presentation from selection

  • Share with another ChatGPT agent

Before any transfer, ChatGPT should display a confirmation dialog showing:

  • the destination application or agent;

  • the workbook, worksheets, ranges and objects being shared;

  • whether formulas, values, charts, comments or metadata are included;

  • whether the transfer creates a static snapshot or a live connection;

  • the permissions granted to the destination agent;

  • the duration and revocation options for those permissions.

Nothing should be transferred without explicit user authorization.

Suggested technical architecture

One possible architecture would be an OpenAI Workspace Context or Agent Handoff Protocol that sits between the Office-hosted agents.

Rather than letting agents communicate through unrestricted natural-language messages, the platform could exchange typed and permission-scoped artifacts.

Example artifact envelope:

{
“handoff_id”: “handoff_123”,
“source_host”: “excel”,
“destination_host”: “powerpoint”,
“workspace_id”: “workspace_456”,
“artifact_type”: “financial_analysis”,
“created_at”: “2026-09-03T08:24:00Z”,
“source_document”: {
“document_id”: “workbook_789”,
“version”: “v17”
},
“artifacts”: [
{
“type”: “range”,
“sheet”: “Filiales PNB”,
“address”: “B1:H16”,
“representation”: “values_and_formats”
},
{
“type”: “chart_specification”,
“source_range”: “Filiales PNB!B2:H16”
},
{
“type”: “narrative”,
“purpose”: “executive_summary”
}
],
“permissions”: {
“read_values”: true,
“read_formulas”: false,
“read_comments”: false,
“create_slides”: true,
“modify_source_workbook”: false
},
“sync_mode”: “snapshot”
}

This payload is only an illustrative concept. The actual implementation could use internal structured objects, the Apps SDK, Microsoft Graph, Office.js or another secure interoperability layer.

Core technical components

1. Shared workspace identity

Excel and PowerPoint agents should know when their documents belong to the same user-approved workspace or project.

The workspace should contain references to documents and artifacts, but should not automatically expose the full content of every document.

2. Typed artifacts

Agents should exchange typed objects instead of relying exclusively on unstructured prompts.

Possible artifact types:

  • spreadsheet range;

  • table;

  • chart;

  • image;

  • KPI collection;

  • analytical narrative;

  • assumption set;

  • source citation;

  • presentation outline;

  • document excerpt;

  • action request.

3. Source provenance

Every transferred value or chart should retain its origin:

  • workbook identifier;

  • worksheet;

  • cell or range;

  • document version;

  • timestamp;

  • value type and unit;

  • static or dynamic status.

In PowerPoint, the user could select a chart and choose View source in Excel.

4. Permission-scoped handoffs

Permissions should be granular and revocable.

For example:

  • read selected values only;

  • read values and formatting;

  • read formulas;

  • create a PowerPoint draft;

  • update existing slides;

  • never modify the Excel source;

  • allow one-time access only;

  • allow access until the project is closed.

5. Snapshot and synchronized modes

Two transfer modes would be useful:

Snapshot mode: PowerPoint receives a fixed, versioned copy of the approved artifacts.

Synchronized mode: PowerPoint maintains authorized links to Excel artifacts and can notify the user when the underlying ranges have changed.

Updates should never silently overwrite manually edited slides. The user should receive a comparison and approve the refresh.

6. Conflict and version management

If a source range changes after the presentation is created, ChatGPT could display:

  • previous value;

  • current value;

  • affected slides;

  • proposed slide updates;

  • last synchronization time.

The user could approve all changes, approve them individually or preserve the existing presentation version.

7. Audit trail

The workspace should record:

  • which agent initiated the transfer;

  • which document received the information;

  • which artifacts were shared;

  • which permissions were granted;

  • which user approved the operation;

  • when the transfer and subsequent updates occurred.

8. Security boundaries

The implementation should respect Microsoft 365 tenant policies, document permissions, data residency requirements and organizational controls.

Important safeguards could include:

  • explicit consent before each new connection;

  • least-privilege access;

  • no automatic access to unrelated documents;

  • tenant and user boundary enforcement;

  • administrator-controlled cross-application sharing policies;

  • revocable access tokens;

  • retention controls;

  • prevention of indirect prompt-injection propagation between documents;

  • clear warnings when externally sourced or untrusted document content is transferred.

Suggested agent actions

A standardized handoff interface might support actions such as:

create_handoff()
list_shareable_artifacts()
request_artifact_access()
accept_handoff()
reject_handoff()
create_presentation_from_artifacts()
check_source_changes()
propose_artifact_updates()
revoke_handoff()

The destination agent should receive only the minimum context required to complete the authorized task.

Minimum viable product

A useful first version would not require fully autonomous communication between agents.

An MVP could provide:

  1. A Send selection to PowerPoint button in Excel.

  2. A preview of the ranges, charts and summaries to be transferred.

  3. Explicit user confirmation.

  4. A structured handoff package sent to PowerPoint.

  5. A PowerPoint agent that proposes an outline before creating slides.

  6. Cell-level source references stored with generated slide content.

  7. A manual Check for updates from Excel command.

This would already remove a large amount of manual copying while keeping the user in control.

Example end-to-end workflow

  1. The user selects an Excel table and two charts.

  2. The user asks the Excel agent to prepare a management presentation.

  3. The Excel agent identifies the selected artifacts and creates a summary.

  4. A permission dialog displays exactly what will be shared.

  5. The user approves the handoff to PowerPoint.

  6. The PowerPoint agent receives the structured artifacts.

  7. It proposes slide titles, narratives and visual layouts.

  8. The user approves the outline.

  9. The agent creates the slides and records links to the Excel sources.

  10. When Excel changes, the presentation can check for affected content and propose updates.

Why this matters

This would transform separate Office agents into a coordinated productivity system while preserving human control.

It would be particularly valuable for:

  • financial analysts;

  • consultants;

  • auditors;

  • corporate finance teams;

  • sales and marketing teams;

  • researchers;

  • project managers;

  • students and educators.

The same architecture could later support:

  • Excel to Word reporting;

  • Word to PowerPoint conversion;

  • Outlook to Excel action tracking;

  • PowerPoint to Excel source validation;

  • multi-document project workspaces.

Questions for the OpenAI team and community

  1. Is cross-application agent handoff already being considered for ChatGPT integrations?

  2. Could the Apps SDK support typed artifacts and permission-scoped handoffs between host applications?

  3. Would OpenAI consider a standardized Agent Handoff Protocol for user-authorized context transfer?

  4. How could provenance and document-level permissions be preserved across Microsoft Office applications?

  5. Would other users find snapshot mode sufficient for an MVP, or is live synchronization essential?

Desired outcome

If this were implemented successfully, users could begin an analysis in Excel, continue it in PowerPoint and finalize it in Word without repeatedly copying data or rebuilding context.

The agents would remain separate and permission-bounded, but they could collaborate through structured, auditable and user-approved handoffs.

Thank you for considering this feature request.

2

Hey, thanks for sharing this feature request!

I agree that a secure way to hand off context and work between ChatGPT agents in Excel, PowerPoint, and Word could make cross-application workflows much smoother while keeping the user in control of what gets shared.

I'll surface this feedback to the product team for consideration.

To keep feature requests consolidated, I'll be closing this topic. Thanks again for sharing!

Avinash