How Much Work Can You Delegate to AI? Five Levels from Answer to Finished Work

AI can do more than answer questions, but delegation should grow with the task, permissions, and cost of a mistake. Use five levels: ask for an answer, transform supplied material, produce a deliverable, act in a tool, or repeat a monitored workflow. Choose the lowest level that finishes the job, then define the evidence, approval point, and acceptance test before the AI begins.

Recommendation

Move beyond search by starting at Level 2 or Level 3 on a low-risk task you already understand. Give the AI real source material and ask for a reviewable artifact, such as a comparison table, meeting action list, cleaned spreadsheet, or draft report. Keep actions in external systems manual until the output is consistently accurate.

The five-level model is SikiT’s editorial framework, not an official classification from OpenAI, Anthropic, Google, or another provider. It tracks how much authority leaves your hands. A higher level calls for narrower access, an explicit approval boundary, and a recovery path.

Decision table

The table is a control ladder, not a claim that every product supports every capability. A good Level 3 result may be more valuable than a weak Level 5 automation.

Important differences

Level 1: Answer

This is the familiar search-box pattern. You ask what something means, request options, or find sources. It is fast and useful, but the work still begins after the answer. You must interpret it, move information into the destination, and check whether the result solves the task.

Use this level for orientation or when you do not yet know enough to define a deliverable. Ask for source links and separate verified facts from suggestions.

Level 2: Transform

At this level, the AI works on material you choose. It can extract dates from a document, compare two policies, classify support tickets, or turn notes into a consistent format. OpenAI’s file documentation describes synthesis, transformation, and extraction as supported uses for uploaded documents.

The source boundary makes review easier. Compare the output with selected passages or rows, and tell the AI what it must preserve. A summary that changes a number or omits a condition has failed even when it reads well.

Level 3: Produce

The AI now owns more of the workflow and returns something closer to finished work. OpenAI describes ChatGPT Work as an agent for longer, multi-step tasks and finished deliverables such as documents, spreadsheets, presentations, reports, and Sites.

The prompt now needs an outcome and acceptance test, not just a topic. "Summarize this meeting" is a Level 2 request. "Create minutes, a decision record, an action list, and a follow-up draft, each traced to the transcript" is Level 3.

Level 4: Act

The AI crosses from drafting into changing a system. A connected app may send or edit a message, move a file, create an appointment, or change a record when the action is available and permitted. The same polished text can have a different risk once it leaves the draft.

OpenAI classifies actions such as sending communications, deleting content, moving files, and changing permissions as important actions. Its app settings can require approval at different thresholds. Use the strictest practical setting, and make the prompt stop before any action you want to review yourself.

Level 5: Repeat

A scheduled or triggered task can produce a brief, monitor a source, or rerun maintenance without a fresh prompt each time. Repetition saves attention, but it also carries stale assumptions into the next run. A reliable one-time prompt is not automatically a reliable recurring process.

Define what happens when a source is unavailable, a permission changes, or the output conflicts with previous data. Schedule reporting before scheduling external changes. Review the task’s scope and access periodically, and turn it off when the project ends.

Choose by your situation

Use Level 1 when the task is still vague. The purpose is to learn enough to write a better assignment.

Use Level 2 when you have authoritative input and need a bounded conversion. Examples include extracting requirements, normalizing labels, or comparing versions.

Use Level 3 when the result can be inspected before use. Meeting deliverables, a morning work brief, and a folder-renaming plan all fit here when they remain drafts or reports.

Use Level 4 only when the agent needs to change a destination. Start with reversible actions and a preview. Sending, deleting, purchasing, publishing, changing access, and modifying security settings deserve an explicit human approval immediately before execution.

Use Level 5 after the same workflow has succeeded manually and the inputs are stable. Keep alerts and reports separate from write actions. A monitor that reports a change is safer than one that reacts by changing the system on its own.

Before any level above a basic answer, write a task contract:

Outcome: The exact artifact or state to deliver.

Inputs: The files, sources, systems, and time window the AI may use.
Constraints: Facts, format, scope, privacy rules, and forbidden actions.
Approval boundary: The point where the AI must stop and ask before acting.
Acceptance checks: Observable tests that decide whether the result passes.

For example, "organize my files" gives the agent no safe finish line. A contract can instead request an inventory and proposed rename table for one copied folder, forbid file changes and deletion, and require unique destinations and matching counts.

What could change the choice

Data sensitivity may lower the appropriate level even when the task is routine. An approved business workspace, retention policy, and narrow connector may support work that should not enter a personal account. Follow the rules of the organization that owns the data.

Reversibility also matters. A saved private draft is easier to undo than a sent email. Renaming files with a rollback map is safer than permanent deletion. Financial, legal, health, employment, security, and public-facing decisions need stronger review because a plausible mistake can carry real consequences.

Product capabilities are not uniform. Local file access, connected apps, schedules, action permissions, and administrator controls vary by provider, plan, device, workspace, and region. A prompt cannot grant access that the product or administrator has not allowed.

Finally, untrusted content can contain instructions aimed at the agent. Keep web pages, emails, and documents within a narrow task scope. Stop if the agent tries to access an unrelated system, expand its permissions, or follow instructions that conflict with your contract.

Limits and evidence

The five levels and task-contract fields are an editorial decision framework. They have not been validated as a universal standard, productivity benchmark, or measure of model intelligence. The right level depends on the task and the consequences of error, not on a product label.

OpenAI’s current documentation supports the distinction between conversational Chat, multi-step Work, and connected app actions. Anthropic’s current Cowork safety guidance supports the distinction between read and write tools and recommends selective access, dedicated folders, backups, and human oversight for higher-risk work. Both providers can change product availability and controls.

Sources

• ChatGPT Work and Codex | OpenAI Help Center
• Apps in ChatGPT | OpenAI Help Center
• Use Claude Cowork safely | Claude Help Center
• File Uploads FAQ | OpenAI Help Center

Related articles

• GPT-5.6 model selection guide: Sol, Terra, Luna, and reasoning effort
• Qwen 3.8 27B thinks by default, and the default effort is xhigh
• Who spends the time your AI saves?