A general-purpose AI tool will draft your resignation letter, summarize a friend’s difficult message, generate a eulogy, and score a list of job candidates, all in the same afternoon, all with the same fluent confidence. It will not tell you which of those four requests was a reasonable use of a drafting tool and which one quietly handed away something that was yours to carry. That judgment call is the entire subject of this article.
The old question — “can AI help with this?” — is almost always answerable with “yes, it can generate something.” That is precisely why it is the wrong question. The useful question is what you give up by letting it, and whether you are willing to pay that price. This article gives you a six-dimension audit to answer that question deliberately, before you paste the task in, rather than noticing the cost afterward.
Why capability is not the test
Generative models are trained to be broadly helpful and rarely refuse a reasonable-sounding request. That combination — high capability, low friction, near-zero refusal — means the tool itself will not stop you from delegating something you shouldn’t. The stopping has to come from you, and it works better as a habit than as a one-off feeling of unease.
There is a well-studied failure mode that explains why this matters more than it might seem. Psychologist Linda Skitka’s research on automation bias — the tendency to use an automated recommendation as a substitute for actually checking the situation — found that people who trust an automated aid start reducing how carefully they verify it, even when contradicting evidence is sitting right in front of them (Mosier, Skitka, Heers & Burdick, International Journal of Aviation Psychology, 1998). A 2026 review in Philosophy & Technology traces this to what researchers call the “cognitive miser” pattern: humans are disposed to minimize mental effort, and a fluent AI answer is a very cheap way to stop thinking (What is Wrong With Automation Bias?, 2026). None of this means AI assistance is bad. It means the ease of delegating is not evidence that delegating was the right call — the ease is the same regardless of whether the task was low-stakes or one you will regret handing away.
A shorter four-question version of this boundary already exists for personal productivity in what not to delegate to AI. This article is the fuller life-and-work audit: it adds authorship, privacy, and reversibility, and it sorts tasks into keep / assist / delegate / never with worked examples across relationships, grief, and high-stakes judgment — not only office workflows.
The six-dimension audit
Before delegating a task, run it through six questions. None of them alone is decisive; together, they tell you whether a task belongs in your hands, in a supervised assist, in a full handoff, or nowhere near a model at all.
- Consequence. If the output is wrong, who is harmed, and how badly? A wrong grocery list costs a return trip. A wrong medical summary or legal filing costs something you may not be able to undo.
- Authorship. Does this output need to represent your voice, judgment, or credentialed expertise, in a context where someone is relying on that being genuinely yours? See keeping your name on AI-assisted work for the disclosure line this raises once you decide to use assistance.
- Relationship. Is a specific person on the other end who is owed your direct attention — a partner, a child, a grieving friend — rather than a polished proxy for it?
- Skill. Is this a task you are actively trying to get better at, where skipping the struggle skips the learning? See what deteriorates when you outsource thinking for how that trade plays out over months, not just once.
- Privacy. Does completing this task require typing in information — someone else’s health details, a child’s data, confidential business information — that should not leave your control?
- Reversibility. If this goes wrong, can you catch it and fix it before it matters, or is the action final the moment it is sent, filed, or spoken?
A task that scores low-stakes, non-authorial, no-specific-relationship, not-a-skill-you’re-building, no-sensitive-data, and easily reversible is a fine candidate for full delegation. A task that scores high on even two or three of these dimensions needs to stay substantially yours.
High-stakes, relational, identity-forming, consent-sensitive, or irreversible actions require stronger human ownership than a quick draft-and-send. When several of the six dimensions score high at once, treat that as a hard stop, not a judgment call to make under deadline pressure.
Keep / assist / delegate / never: ten worked examples
| Task | Consequence | Relationship / authorship | Verdict |
|---|---|---|---|
| Drafting a routine work email | Low | Low | Delegate — draft freely, skim before sending |
| Summarizing a long PDF report for your own use | Low | Low | Delegate — verify any number you plan to quote |
| Writing a performance review of someone you manage | Medium-high | High (your judgment, their career) | Assist — AI can help you organize evidence; the assessment must be yours |
| Telling a friend their plan is a bad idea | Low financial stakes, high relational stakes | High | Keep — see preparing for a hard conversation instead of outsourcing the message |
| Comforting someone who is grieving | Low financial stakes, very high relational stakes | High | Never delegate the act of presence itself — see what AI can delegate in family life, and what it cannot touch |
| Drafting a eulogy from your own notes and memories | Medium | High authorship, but source material is yours | Assist — organize your words; do not let it invent details or sentiment you did not supply |
| Scoring or ranking job candidates | High (someone’s livelihood) | High accountability | Never without an accountable human process — see refusing to rank people with AI |
| Practicing a new language or skill | Low immediate stakes, high long-term stakes | Skill-building | Assist carefully — use it to check your attempt, not replace the attempt; see deliberate practice with AI |
| Reconciling a shared household budget spreadsheet | Medium | Low | Delegate the mechanics; keep the decisions about tradeoffs |
| Apologizing to someone you hurt | Low financial stakes, very high relational stakes | High | Never — see why repair and apology stay human |
Notice the pattern: tasks that are mechanical, low-stakes, and yours to redo if wrong lean toward delegate. Tasks where a specific relationship, a credential, or an irreversible consequence is on the line lean toward keep or never, regardless of how good the AI-generated draft would look.
A common misconception
The most common mistake is treating “AI produced something good” as proof the delegation was fine. Quality of output and appropriateness of delegation are different questions. A beautifully worded, emotionally resonant message to a grieving friend can be a worse choice than a clumsier one you wrote yourself, because the value of that message was never really about word choice — it was about the fact that you sat down and wrote it. The six-dimension audit exists precisely because “it read well” is not one of the six questions.
A second misconception runs the other way: treating every use of AI assistance as some kind of compromise. That overcorrection is also wrong. Delegating a first draft, a summary, or a spreadsheet formula is not a moral failing — it is the entire reasonable use case for the tool. The audit is there to catch the minority of tasks where delegation quietly costs something real, not to make you second-guess routine assistance. For structured, higher-stakes decisions specifically, pair this audit with using AI for better decisions, which covers how to use a model as a sparring partner rather than an oracle once you have decided a task belongs in the “assist” category.
The privacy dimension of the audit deserves its own habit, separate from the other five: before pasting anything into a general-purpose AI tool, ask whether it contains another person’s private information, a child’s data, or anything confidential to your employer. See what ChatGPT remembers, sees, and shares for what actually happens to that data once you send it.
Teams building AI-assisted workflows professionally face a structurally similar problem at scale — deciding which steps in a process keep a human checkpoint and which do not. If that is your context rather than a personal one, human-in-the-loop design patterns covers the same six-dimension logic applied to production systems.
Try it today
Pick three tasks you delegated to an AI tool this week without thinking twice, and three you are about to delegate this week. Run all six through the audit. For each one, write down which dimension — if any — scored high enough that it should have stayed with you, or at least stayed in “assist” rather than “delegate.” The human delegation audit gives you a printable version of this same table, with space to log your own worked examples as you build the habit.
The goal is not a permanent list you memorize once. It is a fast, repeatable check that takes fifteen seconds once it is a habit — enough time to catch the handful of tasks per month where the easy path and the right path are not the same one.
