AI Drop
Moritz Wallbrecher  ·  UX / Interaction Design  ·  Kingston University 2026

AI at the Point
of Interaction

A native macOS application that embeds AI directly into drag-and-drop workflows — reducing context switching and interaction overhead.


17–41s
of overhead. Per interaction. → Study

Every time you want AI to help with a file, you leave your work, switch to a browser, navigate to a chat tool, upload the file, construct a prompt, wait, and copy the result back. This is not a minor inconvenience — it's a structural friction that compounds across every knowledge work session and quietly suppresses AI adoption.

— Without AI Drop
1Open browser or app ~4s
2Navigate to AI tool ~6s
3Upload or paste file ~8s
4Construct prompt ~12s
5Wait for response ~6s
6Copy result back ~5s
— With AI Drop
1 Drag file  ·  Drop  ·  Get result ~3s

The trigger is
the drag itself.

AI Drop doesn't add a new step to your workflow — it activates through one you already perform. The moment you start dragging a file, the pill emerges from the MacBook notch. Drop the file, tap an action, get the answer. The mental model is drag-and-drop. You already know it.

01
Drag

Any file, anywhere on screen. The pill appears automatically from the notch — no activation step.

02
Drop

Release onto the pill. Contextual action chips appear — adapted to the file type instantly.

03
Done

AI result renders inline. Full Markdown. No switching tabs. No copying back.

No new interaction paradigm required. Drag-and-drop is one of the most deeply learned behaviors in desktop computing. AI Drop borrows that trust entirely. When AI activates through familiar behavior, adoption increases without onboarding.


This is
shipped.

AI Drop is a fully functional native macOS application — not a prototype, not a concept. Open source, downloadable, 1.2 MB. Built in Swift with zero third-party dependencies.

Platform
Native macOS  SwiftUI + AppKit, no Electron
Download
1.2 MB DMG  Zero dependencies
AI Providers
Groq, OpenAI, Anthropic, Ollama  Bring your own key
Key Storage
macOS Keychain  Never on disk
Files
PDF, DOCX, TXT, code, images  Contextual chips per type
License
Open Source  Non-commercial

The prototype serves as an experimental vehicle to investigate how embedding AI at the point of interaction affects task efficiency, cognitive load, and usage frequency — not a product evaluation, but a study of how interaction placement shapes AI usage behavior.

User study conducted at Kingston University, 2026. Participants completed controlled tasks comparing a standard AI workflow with the AI Drop prototype across three task types.

0s

Maximum interaction overhead eliminated per task

0%

Participants who said they'd use AI more frequently

n≈0

Participants across controlled task conditions

The prototype was not just faster — it showed significantly reduced variance. Users completed tasks consistently regardless of technical background. The system removed the steps that introduced inconsistency, not only the ones that introduced delay.

These findings suggest that AI adoption is not only limited by model capability, but significantly by interaction placement within existing workflows. The results indicate that reducing friction at the point of intent may be a meaningful driver for increasing real-world AI usage frequency.

Full methodology & results — On request

Full preprint paper available upon request.

File upload task Summarisation task Follow-up prompt task Qualitative debrief per participant

Accepted · Forthcoming 2026
AI at the Point of Interaction: Reducing Cognitive Switching Cost through Embedded, Drag-Triggered AI Interfaces in macOS

This work contributes to Human–AI Interaction research by empirically examining how interaction placement within established GUI paradigms affects task efficiency and user behavior. The paper frames AI Drop as an experimental system — arguing that where AI sits in the interaction stack has measurable effects on task completion time, adoption rate, and usage consistency.

Show Abstract
"Generative AI systems are predominantly accessed through detached, prompt-driven interfaces that operate outside users’ primary task environ-ments. While this model lowers entry barriers by leveraging familiar text-input paradigms, it structurally separates AI invocation from established desktop interaction workflows. This separation introduces procedural fric-tion and reinforces a search-centric interaction pattern that does not fully align with file-based productivity practices. This paper investigates an al-ternative interaction model that embeds generative AI functionality within the interaction architecture of the operating‑system‑level workflows through inherent graphical drag‑and‑drop mechanisms rather than treating generative AI as an external conversational tool. Building on decades-established graphical user interface (GUI) conventions—particularly drag-and-drop and direct manipulation—we present an adaptive operating-system-level interface that enables users to trigger contextually relevant AI transformations through spatial file interaction. Through an experimental case study, this structural repositioning of AI in terms of task efficiency and perceived usability was evaluated. Findings indicate that the drag‑and‑drop‑based interaction significantly improves information extrac-tion efficiency compared with traditional prompt‑driven interfaces, while participants reported greater intuitiveness and reduced need for prompt formulation. Participants further reported reduced cognitive effort com-pared to conventional prompt-driven workflows. These results suggest that the integration point of Generative AI within GUI ecosystems plays a criti-cal role in shaping workflow adoption, highlighting interaction placement as a key design dimension for next-generation AI systems."

Full preprint paper available upon request — moritz@wallbrecher.net

Referenced works
Amershi et al. (2019)
Guidelines for Human-AI Interaction
Shneiderman (2022)
Human-Centered AI
Norman (2013)
The Design of Everyday Things

Where AI lives
changes how it's used.

01

Interaction placement has a measurable effect on adoption. A capable model behind three UI steps may see lower usage than a simpler model that is immediately accessible. Friction compounds across repeated interactions.

02

Friction may be a stronger predictor of AI adoption than raw capability. This study observed that participants interacted with AI more frequently when it required fewer steps — regardless of the model used. Proximity to the workflow appears to matter.

03

Mental models reduce learning cost. When AI activates through behavior people already know — dragging a file — adoption increases without a single tutorial. You can borrow trust from familiar gestures.

04

AI should integrate into workflows, not sit beside them. The notch is a statement about where AI belongs in the interaction stack. Not in a browser, not in a sidebar. Here, now, in context.


Open source.

No account. No subscription. No telemetry. Download the DMG and you're running in under 60 seconds.

⚠ Right-click → Open on first launch (not yet notarized)
Requires macOS 13 Ventura or later

1
Download the DMG

Get AIDropv0.7.dmg from GitHub Releases. 1.2 MB.

2
Drag to Applications

Standard macOS install. No system modifications.

3
Pick provider, paste key

Groq, OpenAI, Anthropic, or Ollama. Key goes to Keychain immediately.

4
Drag any file upward

The pill appears. The interaction is designed to be immediately understandable.

LanguageSwift 5.9
FrameworkSwiftUI + AppKit
DependenciesNone
Key storagemacOS Keychain
Concurrencyasync/await + structured tasks
Min. macOS13 Ventura