What to Hand to AI
A trust test you can run in three seconds.
Faced with "can I hand this to AI," the two most common answers are both wrong: hand over everything, or hand over nothing. The right answer is not about your attitude toward AI; it is a judgment about the task — three seconds, three questions.
Symptoms you'll recognize:
- You delegate everything, then one incident sends you back to doing it all yourself.
- Afraid of errors, you delegate nothing, and AI becomes a decoration.
- "It's just a suggestion, nothing major" — until it is.
The three-question test
Can the error be reversed? Can the sent email be recalled, the deleted file recovered, the paid money refunded? If not, no autonomous execution. If it is reversible — drafts, first drafts, temporary content — hand it over; wrong answers cost a redo.
How wide is the impact? Does an error touch one person and one record, or every user and the whole production environment? Wide: keep a human checkpoint. Single and locally fixable: AI can do it.
Can you verify it? When it is done, can you tell whether it is right? With a source, a standard answer, or a way to check, fine. If it can fabricate and you would never know — and models are extremely good at this — be careful.
One "no" among the three and the verdict is: downgrade, either add a human confirmation or shrink the scope. Look through the AI Incident Database and most entries trace to these three: an irreversible action executed automatically, an impact underestimated, a verification skipped. incident-db
Four classes of tasks
All green (reversible, narrow, verifiable) → full automation. First drafts, format conversion, headlines, summaries. This is where AI belongs most: high volume, editable, cheap to be wrong.
Irreversible or wide → AI does it, a human approves. Sending email, submitting orders, publishing content. AI does the heavy lifting; the person presses the last button. This is the classic spot for a human checkpoint in permission tiers.
Unreliable but checkable → AI does it, a human verifies. Summarizing external material, translating contracts, extracting data. AI produces a draft; a person checks the key points, using the three verification moves to filter out most fabrication.
Too many "no"s → do it yourself. Key decisions, life and property, irreversible operations you cannot verify. Some things should never be handed over. The EU's AI Act gates high-risk uses separately, for the same reason: when consequences get heavy enough, you cannot rely on the model's good behavior. eu-ai-act
Note: the categories are not fixed. The same task can move from "do it yourself" to "AI plus human check" by adding a verification step, or to "AI plus human approval" by adding a confirmation gate. What improved was not the AI; it was your process design.
The boundaries people slip past
A few kinds of tasks look delegable and are not.
Reversible-looking but not. "Just delete this section" — but it is the only copy. Reversibility is about having a backup and a recovery path, not about the operation itself.
Underestimated impact. "Just an internal email" — that gets forwarded to a client. Estimate impact by worst-case propagation, not by the direct audience.
Skipped verification. "It's only summarizing, no need to check" — the model's confidence in its fabrication far exceeds human expectation. The more a task looks like it needs no verification, the more you should ask: if something goes wrong, would I even know? Security threat models for LLM applications put the same items at the top: permission abuse and over-delegation. owasp-llm
Finally, apply the test in product design: do not build a "fully automatic" AI feature. Decide up front which class every step belongs to, and build the human checkpoints and verification points into the flow. Trust is not achieved by praying; it is achieved by design.