AI Desktop Operation

Many business systems have no API, so the work can only be done through the interface. This is a laboratory question: under what conditions is handing interface operation to an AI both safe and worthwhile? Boundaries, authorisation and the ability to stop are the substance of the investigation, not a footnote to it.

Where there is no interface to call, automation must act through the UI — and UI actions come with no argument validation and no permission model. Without a boundary, a single misjudgement can be irreversible.

• 01

  Explicit authorisation scope

  The applications and windows it may touch are declared in advance; anything outside is refused.

• 02

  Human confirmation

  Irreversible actions — submit, pay, delete — must be confirmed by a person.

• 03

  Full action log

  Every click and keystroke is recorded and can be audited afterwards.

• 04

  Stop at any time

  A stop control is always available and halts execution immediately.

• 05

  Restricted credential handling

  Passwords, payment details and identity document numbers sit on a never-type list; the moment such a field appears, control returns to the person.

1. 01

   Data movement without APIs

   Transcribes information between legacy systems that expose no API.

2. 02

   Repetitive UI work

   Runs structurally fixed, low-judgement interface tasks in volume.

3. 03

   Process verification

   Tests in a controlled environment whether a process can be automated at all.

• Browsers
• Desktop office software
• Legacy systems without APIs
• Workflow engine

This is not a system with unlimited authority. Scope must be declared in advance, irreversible actions require human confirmation, a stop control is always available, and all actions are logged. The system does not enter passwords, payment details or identity document numbers. It runs only in controlled environments.

An experimental project, not open for general use. Current work is focused on boundary definition, failure modes and audit mechanisms — whether this becomes a product depends on those answers.

• 04AI Workflow AutomationFlow EngineOrchestrating triggers, models, knowledge, tools, approvals and business systems into a process that keeps running.Private Preview
• 02Enterprise AI Agent StudioMoShun Agent StudioDesigning AI agents around real roles, processes, knowledge and system interfaces — agents that step into the flow of work rather than sit beside it.Enterprise Delivery

• 2026-05-06Lab NoteDesktop operation: defining what it must not do, firstThe substance of this experiment is the boundary, not the capability. Passwords, payments and identity documents are explicitly excluded.