Which tasks to hand to AI

Before you delegate a task, ask two questions: how easily can I check the answer, and what happens if it’s wrong?

The idea

Good delegation to AI isn’t about how clever the model is. It’s about your ability to catch a mistake and the cost of missing one. Those two things — verifiability and stakes — place any task on a simple map.

Where a task lands tells you the play: hand it over fully, let AI draft and you verify, use AI only as a sounding board, or keep it in human hands. Same tool, four very different amounts of trust.

How it works

Run every task through two quick checks before you reach for AI:

1. VERIFY  — can I confirm the output is right, cheaply?
             read-it-and-know = easy;  need-to-redo-the-work = hard.

2. STAKES  — if a wrong answer slipped through, what breaks?
             mild annoyance = low;  money, safety, trust, legal = high.

           easy to verify        hard to verify
 high  |  Draft, then verify  |  Keep it manual        |
 stakes|  (AI drafts, you     |  (own it; AI is at     |
       |   check every claim) |   most a second opinion)|
 ------+----------------------+------------------------+
 low   |  Delegate fully      |  Assist only           |
 stakes|  (hand over, glance) |  (AI widens options,   |
       |                      |   you pick)            |

Two more dials sit underneath the map. Off-the-shelf vs custom setup: reach for a plain chat tool for one-off work; only invest in a custom setup (retrieval over your own docs, saved prompts, guardrails) when the same task repeats and correctness matters. Cheap-and-fast vs slow-and-capable: use a small quick model for low-stakes, easy-to-verify jobs, and save the slower, stronger model for the top-right and top-left where a wrong answer is expensive.

When to use it

Play	Fits when…	The trade-off
Delegate fully	Cheap to check, cheap if wrong — summaries, formatting, first-pass tidying.	Still glance at it; “fully” never means “unseen.”
Draft, then verify	You can read and confirm it, but a mistake reaches someone who matters.	Verification is real work — budget the time, don’t auto-send.
Assist only	No single right answer to check, and little breaks if a suggestion is weak.	You still own the judgement; AI just widens the option set.
Keep it manual	Hard to verify and costly if wrong — regulated, safety- or trust-critical calls.	Slower, but a confident wrong answer here is the expensive kind.

Watch out for

• Confusing fluent with correct. A polished paragraph or a tidy-looking number feels trustworthy. Fluency is not verification — that’s exactly why hard-to-check tasks are dangerous.
• Skipping the glance on “delegate fully.” Fully delegated still means a human read it once. The failure mode is auto-pasting output nobody looked at.
• Treating high-stakes math as easy to verify. A budget or dosage can look reasonable and be wrong. If checking means redoing the work, it belongs on the hard-to-verify side.
• Letting a low-stakes tool creep into high-stakes use. The brainstorm helper that’s fine for taglines is not fine for interpreting a contract. Re-place the task when the stakes change.
• Over-building. A custom pipeline for a task you do twice a year is wasted effort. Match the setup to how often the task repeats.

Worked example

An interviewer asks: “Where would you use AI in your day, and where wouldn’t you?” A weak answer is “everywhere” or “nowhere.” A strong one reasons on the map. Summarising a call I was on — easy to check, low cost — I delegate and skim. A client-facing email — easy to read but it reaches a customer — I let AI draft, then I verify facts and tone before it sends. Reconciling the quarter’s budget — a wrong figure is expensive and hard to spot — I keep the math mine and use AI only to sanity-check totals I’ve already computed. That reasoning — verifiability and stakes, not vibes — is what a good answer sounds like.