AI AGENTS · HUMAN CONTROL

Which tasks can an AI agent handle—and which decisions should remain human?

An AI agent should not receive permission to do everything it can attempt. The useful question is not whether it can do the task, but what happens when it is wrong and who reviews the outcome.

Three levels, not two

Automation versus no automation is too simple. Separate automatic execution, assistance with review, and human decision.

One agent may operate at different levels depending on data, customer and risk.

Tasks suitable for automation with limits

Classify requests

Suggest category, initial priority or department, escalating uncertainty.

Extract document data

Read references, dates or amounts and prepare them for validation.

Prepare recurring drafts

Create responses, summaries or documents from authorized sources.

Update low-risk states

Record a received file or a verifiable completed stage.

Send reminders

Notify people about deadlines or missing information without imposing sensitive consequences.

Tasks where AI should assist, not decide

Answer complex questions

Find precedent and draft; a person retains judgement and tone.

Analyze incidents

Group signals and suggest causes without closing critical issues.

Prepare quotes

Collect data and draft while price and commitment require approval.

Prioritize work

Recommend transparently and allow correction.

Review documents

Flag gaps or contradictions while responsibility remains human.

Decisions I would keep human

Hiring and firing, medical or legal decisions, credit access, sanctions, sensitive exceptions, major financial commitments and hard-to-reverse actions.

AI can prepare information, but should not hide who made the decision.

Safety questions before connecting it

Which sources can it read?

Output depends on source quality.

Which actions can it take?

Reading, drafting and sending are different permissions.

What happens when it is uncertain?

It must stop and request review.

Is activity logged?

The source and action should remain visible.

Can it be reversed?

Begin with recoverable actions.

Where I would start

Choose a frequent, bounded task that is easy to review. Measure time saved, escalations and corrections.

A boring pilot that works is more valuable than a spectacular agent with access to everything.

Good agent design reserves human attention for decisions that genuinely require judgement and accountability.