From Tool Selection to Intent-Driven AI Orchestration
Independent UX/Product Concept
I have been thinking about the way ChatGPT’s + menu presents files, photos, thinking, research, web access, plugins/apps, visualization, and other capabilities.
My observation is not that these capabilities are unnecessary. Quite the opposite.
The question is whether they should all be presented to the user as though they are the same type of choice.
A file is something I provide.
Web search is a resource.
Thinking is a reasoning capability.
Deep Research is a research workflow.
Visualization is something I want ChatGPT to accomplish.
Those are different concepts.
The basic idea
I would separate the experience into three conceptual layers:
1. What am I providing?
Files
Photos
Camera
Sketch
2. What resources/reasoning/actions may be required?
Thinking
Deep Research
Web access/search
Connected sources/apps
Other tools
3. What do I want?
Visualize
Analyze
Summarize
↔️ Compare
Rewrite
Create a report
The key change is the second category:
ChatGPT should normally determine it automatically from the user’s prompt, inputs, context, constraints, and desired outcome.
The user should not have to understand ChatGPT’s internal machinery before asking it to do something.
The principle
I would describe the intended experience like this:
The user specifies the destination; the AI determines the route.
For example:
“Analyze this PDF, compare it with current information from reliable sources, and prepare a detailed report with citations.”
Instead of asking the user to manually construct:
📎 File
+
🧠 Thinking
+
🌐 Web Search
+
🔬 Deep Research
+
📝 Report
ChatGPT could infer the required workflow.
The interaction becomes:
User
↓
Express goal
↓
ChatGPT understands the goal
↓
ChatGPT determines appropriate workflow
↓
ChatGPT communicates what it is doing
↓
ChatGPT executes
↓
User reviews / intervenes when necessary
↓
Output
↓
User reviews / refines
This does not mean “AI decides everything and the user has no control.”
It means:
Intelligent defaults first; meaningful manual control when needed.
Multi-select and intervention should remain possible
A user may know exactly what they want.
For example:
“Use these two files, search the web, and create charts.”
The interface should allow:
📎 Report.pdf
📎 Data.xlsx
🌐 Web Search
📊 Visualize
It should also allow the user to change the workflow during execution.
For example:
“Stop the research. Use only government and university sources.”
The system should be able to pause, adapt, and continue.
This is especially relevant for complex research workflows. OpenAI’s current Deep Research documentation already describes proposed research plans, user review/modification, visible progress, and interruption/refinement during execution.
Transparency is essential
Automatic orchestration should not become a black box.
Instead of technical labels alone, ChatGPT could communicate in plain language:
Standard transparency could look like:
And after completion:
✓ Completed
Workflow
────────────────────
📎 Analyzed uploaded PDF
🌐 Searched the web
🔬 Conducted multi-step research
🧠 Compared findings
Sources
────────────────────
📎 uploaded PDF
🌐 8 web sources
Research time
────────────────────
4 min 12 sec
Additional interface mechanisms could include concise descriptions, tooltips, live status messages, progress bars, expandable workflow details, and completion summaries.
The goal is:
Explain enough to build understanding without forcing users to manage the machinery.
What should happen when users select something unnecessary?
Suppose the user selects Deep Research but asks:
“Calculate the average from this Excel file.”
ChatGPT should be able to recognize that the selected workflow may not be necessary.
It could say:
Deep Research isn’t needed for this task.
and offer to remove or bypass it.
For ordinary processing, unnecessary capabilities can be optimized away.
For external actions, sensitive data access, or consequential operations, permissions and appropriate user confirmation should continue to apply. OpenAI’s app documentation describes permission-sensitive behavior for connected services and actions.
Why I think this matters
AI interfaces are different from traditional software interfaces.
Traditional software often asks:
“Which tool do you want to use?”
An AI assistant can instead ask:
“What are you trying to accomplish?”
That is a significant shift.
The + menu should therefore be more than a toolbox.
It could become a context-and-control surface:
WHAT AM I PROVIDING?
📎 Files · 🖼 Photos · 📷 Camera · ✏️ Sketch
WHAT MAY BE REQUIRED?
🧠 Thinking · 🔬 Research · 🌐 Web
🔗 Connected sources/apps · ⚙️ Tools
WHAT DO I WANT?
📊 Visualize · 🔎 Analyze · 📋 Summarize
↔️ Compare · ✍️ Rewrite · 📝 Report
But the user should usually only need to answer the third question through natural language.
ChatGPT can work out the second.
Android vs. Web is also interesting
The Android and Web + menus I observed are organized differently.
Web
Upgrade to Plus
Camera
Photos
Files
Create Image
Sketch
Thinking
Deep Research
Web Search
OpenAI Platform
Visualize
Plugins
Android
Camera
Photos
Files
Plugins
├─ Create image
├─ Sketch
├─ Thinking
├─ Deep research
├─ Study
├─ OpenAI Platform
├─ Visualize
└─ Plugins
Think harder
I see this less as a question of “which platform is right” and more as evidence that the product has several different conceptual dimensions being surfaced through the same general entry point.
OpenAI’s documentation confirms that plugin/app capabilities and availability can vary across interfaces and other account/workspace factors, while plugins can package skills, connected apps, extensions, and workflows.
Proposed product principle
Automate the workflow. Explain the workflow. Preserve meaningful user control.
The user should not need to become a tool operator in order to use an AI assistant effectively.
They should be able to say:
“Analyze these documents, research the latest information, compare the evidence, visualize the important findings, and create a report.”
And ChatGPT should be capable of determining the appropriate process.
At the same time, an advanced user should be able to say:
“Don’t search the web.”
“Use only these files.”
“Use Deep Research.”
“Change the visualization.”
“Stop the process.”
“Continue.”
That combination—automatic orchestration plus human intervention—is the part I find most interesting.
Final concept
Instead of treating every ChatGPT capability as another attachment/tool option, distinguish what users provide, what resources/processes are involved, and what users want accomplished—while allowing natural language to remain the primary expression of intent.
The design goal is not to remove capabilities.
It is to reduce the amount of configuration the user must perform manually.
In one sentence:
Let the user describe the destination; let ChatGPT determine the route; let the interface make the route understandable and controllable.
Note: This is an independent product/UX concept inspired by observing ChatGPT’s interfaces and current OpenAI documentation; it is not an official OpenAI product proposal or statement of OpenAI’s roadmap.