Most people are having the wrong argument with AI. They tune prompts, chase the newest model, and compare token prices — and still feel slower than they should, or quietly nervous about what they’re shipping. The bottleneck almost never turns out to be the prompt. It’s the decision that comes before the prompt: should this task go to the machine at all, and how much of it?
Get that decision wrong in one direction and you under-delegate — grinding through email, first drafts, and data cleanup by hand while a tool that costs a few dollars a month sits idle. Get it wrong in the other direction and you over-delegate — pasting an AI answer into a client deck, a legal reply, or a financial model without the faintest idea whether it’s true. Both mistakes are expensive. This playbook is a single, repeatable filter that keeps you out of both ditches.
The one question that routes everything
Every task you could hand to AI has two properties that matter more than any other. Score those two and the right move is nearly automatic.
- Stakes — what happens if the output is wrong and you don’t catch it? A typo in a personal note is a shrug. A wrong number in a board deck, a fabricated case in a legal filing, or a bad dosage in health advice is a fire. Rate it Low or High.
- Verifiability — how fast and cheaply can you tell whether it’s right? Code either runs the tests or it doesn’t — high verifiability. A confident claim about a niche regulation, buried in fluent prose, might take an hour to check — low verifiability. Rate it Low or High.
Notice what’s not on the list: how impressive the output looks. Modern models are fluent by default, and fluency is exactly what disguises a wrong answer. Stakes and verifiability are the axes because together they tell you how much of your own attention the task still requires.
The delegation matrix
Cross the two axes and you get four quadrants. Each one has a default move.
- Low stakes + high verifiability → Automate. Let AI run and spot-check occasionally. Reformatting notes, drafting throwaway replies, converting a table to JSON, renaming files, summarizing a transcript you also have. If it’s wrong, you’ll see it instantly and nothing breaks.
- High stakes + high verifiability → Draft. Let AI produce the first version, then verify every checkable element before your name goes near it. Code, spreadsheets, contracts with defined terms, anything with numbers you can recompute. The machine does the typing; you do the checking. This is the workhorse quadrant.
- Low stakes + low verifiability → Assist. Use AI as a thinking partner, not an author. Brainstorming names, pressure-testing an idea, drafting angles you’ll rewrite anyway. You can’t easily “check” a brainstorm, but nothing rides on any single suggestion, so the risk is capped.
- High stakes + low verifiability → Keep. Do it yourself; let AI touch only the scaffolding. A medical decision, a firing, a public apology, an original strategic judgment, a claim about a fact you can’t confirm. AI can format your thinking or catch a typo, but the substance stays with the human who’s accountable for it.
The trap quadrant is the last one, and the danger is that its outputs look identical to the “Draft” quadrant. A fluent, well-structured, completely unverifiable answer to a high-stakes question is the single most dangerous thing a language model produces. When you can’t verify and the stakes are real, polish is not evidence.
The 60-second routing script
Before you hand anything over, run these four questions out loud. They take less time than writing the prompt.
- “If this is wrong and I miss it, who pays?” If the answer is a client, your employer, your reader, or your body — stakes are High.
- “How would I actually check it, and how long would that take?” If you can’t name a concrete check, treat verifiability as Low, not High-because-it-sounds-right.
- “What’s the smallest slice I can delegate?” You rarely hand over a whole task. Give the machine the outline, the boilerplate, or the first pass — keep the judgment.
- “What will I verify before this leaves my hands?” Name the check now, while you’re calm, not after the output has charmed you.
The delegation brief
For anything in the Draft or Automate quadrants, a thirty-second brief beats a clever prompt. Paste this, fill the blanks, and you’ll get usable output on the first try far more often:
- Role & goal: “You’re helping me produce [X] for [audience]. The goal is [outcome].”
- Inputs: the real source material, pasted in. Do not make the model guess facts — give it the facts.
- Constraints: length, format, tone, and what to avoid.
- Uncertainty rule: “If you’re not sure of a fact or number, flag it as [UNVERIFIED] rather than guessing.” This one line turns invisible risk into a visible checklist.
- Definition of done: “A good answer has [criteria]. Before finishing, list anything I still need to verify.”
That last instruction — make the model surface its own weak points — is the cheapest quality upgrade available. Then run those flags through a real verification pass; the AI Trust Ladder lays out exactly how deep to check based on stakes.
A worked example
Say you owe a client a weekly performance report. Break it into slices and route each one:
- Pull and format the raw numbers — high verifiability (you have the source), low-to-mid stakes → Automate/Draft. Let AI assemble the table; you reconcile totals against the source.
- Write the plain-English summary — high stakes (client-facing), high verifiability (you know what happened) → Draft. AI writes it; you correct any claim that overstates results.
- Recommend next month’s strategy — high stakes, low verifiability (it’s a judgment) → Keep. You write this. AI can tidy the wording, but the call is yours; it’s why the client pays you.
One report, three different routes. That’s the whole skill: stop asking “can AI do this?” and start asking “which parts, and how much do I check?” Once a workflow like this repeats every week, it’s a candidate for real automation — that’s where the $2 Test takes over, turning a proven manual routine into a standing agent.
The five failure modes
- Fluency blindness. Confusing a well-written answer for a correct one. The fix: verify against a source, never against your gut reaction to the prose.
- Silent scope creep. A task starts in “Assist” and drifts into “Keep” territory as the stakes rise, but you never re-route. Re-score when the stakes change, not just when the task starts.
- Verifying the wrong thing. Checking grammar and formatting while waving through the one number that matters. Verify the load-bearing claim first; cosmetics last.
- Delegating the accountability. “The AI said so” is not a defense to a client, a boss, or a regulator. Whoever presses send owns every word.
- Over-keeping. The quiet, opposite failure: hoarding low-stakes, checkable work out of habit or pride while your hours evaporate. Audit your “Keep” pile monthly and push everything you can down into Draft.
Your first week
Don’t reorganize your whole life. For five working days, keep a sticky note with the four quadrants next to your screen. Every time you’re about to do something a machine could touch, say the quadrant out loud before you start. By Friday you’ll route most tasks in a couple of seconds, and you’ll notice the two piles that were costing you: the low-stakes chores you should have automated weeks ago, and the high-stakes, low-verifiability calls you’d been lazily letting AI make. Fix those two and you’ve captured most of the upside.
Our take: As models get more capable, this decision gets more important, not less — because the better the output looks, the harder it is to feel the difference between “checked” and “charmed.” The people who win with AI in 2026 won’t be the ones with the cleverest prompts. They’ll be the ones with the cleanest judgment about what to hand over and what to guard, applied so automatically it looks like instinct. Delegation, not prompting, is the durable skill.