AI Revolution AtlasAsk Dr. Mira
Menu

Task decisions

Should AI do it, help with it, or stay out?

Use a practical framework to screen tasks by stakes, reversibility, sensitive data, error detection, and required judgment.

12 minute readLast reviewed 2026-06-20

Plain-language summary

What this guide covers

Automation means the system performs a task with little or no human action. Augmentation means the system helps a person perform a task. Many real workflows are mixed: AI drafts, sorts, or suggests while a person reviews, decides, and owns the outcome.

Why it matters

Most bad AI decisions start with choosing the wrong level of control. A low-risk draft may be a good AI-assisted task. A sensitive decision about a person may need strict human review or no AI at all.

What you will learn

  • Define automation, augmentation, and mixed workflows.
  • Use a task-screening framework for AI decisions.
  • Apply the framework to five sectors.
  • Identify common mistakes in AI task design.
  • Design human checkpoints for review and accountability.

Guide section

Three ways AI enters work

The right question is not only whether AI can do something. It is what role AI should play.

Automation means AI or software performs a task with little human action, such as routing a low-risk message to a folder. Augmentation means AI supports a person, such as drafting a first version of a reply that a worker edits. A mixed workflow combines both: software may sort, summarize, and suggest, while a person handles exceptions, approval, and responsibility.

Guide section

The five-part task screen

Use these questions before deciding whether to automate, augment, or avoid a task.

Screening factorLow-risk signalHigh-risk signalLikely AI role
StakesSmall inconvenience if wrong.Could affect rights, money, health, safety, grades, employment, or trust.Low stakes may allow automation; high stakes need human-led review or no AI.
ReversibilityMistake can be quickly fixed.Mistake is hard to undo or leaves a lasting record.Irreversible tasks need stronger human checkpoints.
Sensitive dataUses public or non-sensitive information.Uses private, confidential, personal, student, employee, medical, legal, or financial data.Sensitive data requires approved tools, minimization, and policy controls.
Error detectabilityA person can easily check the output.Errors are subtle, hidden, technical, or outside the reviewer’s expertise.Hard-to-detect errors require expert review or avoidance.
Judgment requiredTask has clear rules and few exceptions.Task requires values, context, empathy, tradeoffs, or accountability.Judgment-heavy tasks should stay human-led with AI as limited support if allowed.

Guide section

A practical decision sequence

This sequence works for a school, office, team, nonprofit, or small business.

  1. Write the task in one sentence.
  2. Name the person affected by the output.
  3. Score stakes, reversibility, sensitive data, error detectability, and judgment as low, medium, or high.
  4. Choose a role: automate, augment, augment with expert review, or do not use AI.
  5. Define the human checkpoint before the output is used.
  6. Test with safe examples before using real data.
  7. Measure quality, errors, time saved, and user trust.
  8. Set a review date because tools and rules change.

Guide section

Five sector examples

The same framework can produce different answers in different settings.

Example

Office work

AI may summarize a non-confidential meeting transcript and suggest action items. A person should check names, deadlines, decisions, and missing context before sending. Automation may be acceptable for formatting notes; approval stays human.

Example

Education

AI may create practice questions from a teacher-approved reading. It should not secretly complete a student’s final work or grade sensitive student performance without school policy, validation, and human oversight.

Example

Small business

A shop owner may use AI to draft product descriptions from approved product facts. AI should not invent claims, warranties, safety instructions, or customer promises. The owner checks accuracy and tone.

Example

Software

AI coding help may speed up boilerplate, tests, or explanations. Developers still review for security, licensing, maintainability, and correctness. Code that handles payments, health, or personal data needs stronger checks.

Example

Public-facing services

A public agency may use AI to draft plain-language explanations of forms. It should avoid making eligibility decisions without validated systems, transparency, appeals, and human accountability.

Guide section

Human checkpoints make mixed workflows safer

A checkpoint is a clear place where a person reviews, approves, edits, rejects, or escalates.

Good human checkpoint

  • The reviewer knows the goal of the task.
  • The reviewer has enough expertise to spot likely errors.
  • The output is labeled as AI-assisted when disclosure is required.
  • The reviewer checks facts against trusted sources.
  • The reviewer checks privacy, bias, and tone.
  • There is a clear path for unusual or high-stakes cases.
  • The final owner is named.

Try it

Exercise: screen one task

Choose one task from your work, school, or personal learning. Score the five factors. Then write the safest AI role in one sentence.

  1. Task:
  2. Stakes: low, medium, or high.
  3. Reversibility: low, medium, or high.
  4. Sensitive data: low, medium, or high.
  5. Error detectability: low, medium, or high.
  6. Judgment required: low, medium, or high.
  7. AI role: automate, augment, expert review, or avoid.

Avoidable errors

Common mistakes and better approaches

Automating because a demo looked good.

Better approach: Screen the real task and test with real review standards.

Putting a human in the loop without giving them time or expertise.

Better approach: Design checkpoints that reviewers can actually perform.

Ignoring data sensitivity.

Better approach: Use approved tools and minimize sensitive data before any AI step.

Remember this

Key takeaways

  • Automation lets software perform a task; augmentation helps a person perform it.
  • Many AI workflows should be mixed rather than fully automated.
  • Stakes, reversibility, sensitive data, error detectability, and judgment guide the decision.
  • Human checkpoints must be specific, staffed, and meaningful.
  • Low-risk drafting is different from high-stakes decision-making.
  • Review dates matter because tools and rules change.

Questions readers ask

Frequently asked questions

Is augmentation always safer than automation?

Not always. A weak human review can create false safety. Augmentation is safer only when the reviewer has enough time, skill, information, and authority.

Can a task move from augmentation to automation later?

Possibly, if evidence shows reliable performance, risks are low, errors are detectable, and rules allow it. The decision should be reviewed and documented.

What tasks should usually stay human-led?

Tasks involving high stakes, private data, complex values, unclear facts, or responsibility for people should usually remain human-led with strict limits on AI support.

Why include reversibility?

A mistake that can be corrected before anyone is harmed is different from a mistake that affects a grade, job, benefit, customer, or safety record.

Sources and review notes

Sources were accessed on the dates shown. Links open the original organization’s page.

  1. SRC-01
    Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileNational Institute of Standards and Technology · Published 2024-07-26 · Accessed 2026-06-20
  2. SRC-02
    Generative AI and Jobs: A Refined Global Index of Occupational ExposureInternational Labour Organization · Published 2025-05-20 · Accessed 2026-06-20
  3. SRC-03
    AI PrinciplesOrganisation for Economic Co-operation and Development · Accessed 2026-06-20
  4. SRC-08
    Guidance for Generative AI in Education and ResearchUNESCO · Published 2023-09-07 · Accessed 2026-06-20
  5. SRC-09
    AI Act | Shaping Europe’s digital futureEuropean Commission · Accessed 2026-06-20

Your next step

Run the task screen

Use the five factors on one real task before adding AI to it.