# How to Know if AI Can Do Your Task: A Simple Test

*How-to — 2026-10-02 — by Mahmoud Zalt*

Four questions that tell you whether AI can handle a specific task, how much oversight it needs, and when to fix the task before automating it.

**Short answer.** **Ask four questions about the task: can you describe it in five lines, is everything it needs written down somewhere reachable, can you spot a bad result in under two minutes, and what does undoing a mistake cost.** Four yes answers means hand it over now and let it run. Three means hand it over behind an approval step. Two or fewer means the task is not ready, and the fix is usually writing down a rule that currently lives in someone's head.

Every article about AI tells you what it can do in general. That is not the question you actually have. Your question is about the specific thing on your list, the one with the awkward exception and the client who always wants it differently.

General capability lists cannot answer that, because whether AI can do your task depends less on the AI than on the task. The good news is that you can work it out yourself in about three minutes, without trying it first.

The test below is the one we use internally before handing anything over. It has no scoring spreadsheet and no maturity model. It has four questions, and the reason it works is that each one maps to a real property of how this technology succeeds or fails.

It also assumes the right setup. A chat window fails question two on almost every task, because it has no access to anything. An **AI Employee** on **Sistava** connects to the mailbox, files, calendar, and records the task actually lives in, holds a standing brief, remembers your corrections, and runs on a schedule. That changes which tasks pass.

## At a Glance

- **4** Questions that decide whether a task is ready
- **3 min** How long the test takes per task
- **5 lines** The length of a brief a good task needs
- **2 min** How fast you should be able to spot a bad output

## How do I know if AI can do my task?

Run these four questions in order and count the yes answers. Each one tests a different failure mode, and a task that fails any of them fails for a reason you can usually fix. Do not skip ahead to the answer, because the value is in noticing which question you hesitated on.

The questions are deliberately about your task and your business, not about the model. Nothing here requires you to understand how AI works, and that is intentional.

- **Can you describe it in five lines? **What starts it, what information it needs, what a good result looks like, what tone, and where it stops and asks you. If you cannot write those, the rule is in your head and the AI will invent a different one.
- **Is everything it needs written down and reachable? **The rate card, the past emails, the customer record, the policy. If a key input exists only in your memory or on a whiteboard, the output will be plausible and wrong.
- **Can you spot a bad result in under two minutes? **You have to be able to check the work quickly. If verifying takes as long as doing it yourself, you have moved the work rather than removed it, which is the most common quiet failure.
- **What does undoing a mistake cost? **A bad draft costs nothing. A wrong refund costs money and trust. This question does not decide whether to hand the task over, it decides whether a human stays on the trigger permanently.

## What do the scores actually mean?

Four yes answers with a cheap undo cost means hand it over and let it run unattended after a week of drafts. Four yes answers with an expensive undo cost means hand it over but keep the approval step forever. Three yes answers means fix the failing one first, and it is almost always question one or two.

Two or fewer means the task itself is the problem. That is genuinely useful information, because a task nobody can describe and nobody can check is also a task no human employee could take over from you. Automation just made the gap visible.

| Score | What it means | What to do |
|---|---|---|
| 4 yes, cheap to undo | Ideal candidate. Repeatable, checkable, forgiving | Hand over now, draft mode for a week, then let it run |
| 4 yes, expensive to undo | Capable but consequential. Money, law, or a customer | Hand over the preparation, keep human approval permanently |
| Fails question 1 | You cannot describe the rule you are applying | Write the five lines first. That takes an hour and helps regardless |
| Fails question 2 | A key input lives in your head or on paper | Get it into a document or a record, then retest |
| Fails question 3 | Checking the work costs as much as doing it | Split the task so the output is small and skimmable |
| 2 or fewer | The task is undefined, not just unautomated | Do not automate. Define it, or accept it stays yours |

## What does the test look like on real tasks?

It looks like a five item list where two pass, two need a fix, and one should never be handed over. Ken runs a print shop with four staff doing signage, business cards, and short run booklets. He ran the test on the five jobs he most wanted off his plate.

Quoting standard print jobs scored four yes with a cheap undo. The rules are on a price sheet, the spec arrives in the enquiry, and a wrong quote is caught in ten seconds. It went first, and now every standard enquiry gets a same day quote instead of a next day one.

Chasing artwork from customers also scored four. Reordering paper stock failed question two, because the reorder points lived in his head, not in the system. He spent forty minutes writing them into the stock sheet and it passed on the retest.

Quoting complex jobs failed question one. His pricing on unusual work depends on whether the customer is likely to change their mind twice, which he cannot express as a rule because it is a judgment about people. And handling complaints scored three yes with an expensive undo, so the AI drafts the reply and Ken always sends it himself.

**A failed question is usually an hour of work, not a dead end.** Most tasks that fail this test fail on question two, because an input lives somewhere the AI cannot reach. Writing it down takes under an hour and improves the task even if you never automate it.

## What if my task is mostly, but not entirely, describable?

Split it. Almost every task that feels borderline is actually two tasks stapled together: a mechanical part with clear rules and a judgment part that needs you. Handing over the mechanical half is where most of the time saving lives anyway.

Ken's complex quoting is the example. The measuring, the material calculation, and the drafting all pass the test easily. Only the final number needs his read on the customer. So the AI prepares the whole quote and leaves the price blank.

## Comparison

| Dimension | Traditional | With Sista |
|---|---|---|
| Quoting | Reads the spec, calculates materials, drafts the document | Sets the final number based on the customer |
| Support | Sorts, tags, answers the documented questions | Handles anything angry, unusual, or from a key account |
| Hiring | Screens against stated requirements, schedules, takes notes | Decides who to interview and who to hire |
| Invoicing | Prepares, reminds on schedule, flags what is overdue | Decides when to stop chasing and pick up the phone |
| Content | Drafts, formats, schedules, keeps the cadence alive | Decides what the business is actually saying |

If a task fails on the judgment side for reasons that feel permanent, that is worth understanding properly rather than working around. The structural limits are laid out in [what AI still cannot do](/insights/what-ai-still-cannot-do), and they explain most of the hard no answers.

## How do I run the test properly?

Take your three most annoying recurring tasks, score each one on paper, and act on the highest scorer. Doing it on paper matters, because the point is to notice which question made you pause. That hesitation is the actual finding.

### Running the test on your own work

1. **List your three worst recurring tasks** — Recurring is the filter. A one off task is never worth the setup, however painful it was. Write down how often each one happens.
2. **Try to write five lines for each** — Trigger, inputs, what good looks like, tone, and the stop line. The one you cannot finish is telling you something important about your own process.
3. **Check where each input actually lives** — For every piece of information the task needs, name the file, mailbox, or record it sits in. Anything you cannot name is in your head and needs writing down.
4. **Time yourself checking one past output** — Take a result you produced last week and see how long it takes to judge whether it was good. Over two minutes means the task needs splitting.
5. **Write the undo cost next to each** — In money or in trust. This decides whether the approval gate is temporary or permanent, and it is the one answer you should never soften.
6. **Hand over the winner and retest the rest** — Act on the highest scorer this week. Fix the failing question on the others and they will usually pass a fortnight later.

People are often surprised that the test rules out fewer tasks than they expected and rules out different ones than they expected. The task you assumed was too complicated usually passes. The simple one you assumed was easy often fails question two.

That is because complexity is not the constraint. Describability, reachable inputs, and checkable output are the constraints, and those are properties of how organized your business is rather than how hard the work is.

## What if I get it wrong?

You will, on one of the first few, and it costs almost nothing if you kept the approval step on. That is the entire purpose of the draft week. A task that turns out to be a bad fit shows up as drafts you keep rewriting, and you will know by day three.

Plans start at 50 a month, so the cost of a wrong guess is a week of reviewing drafts rather than a wasted project. That is a much cheaper way to find out than a planning exercise.

Once you have scored a few tasks you will start seeing which role should own them, and that is the next decision rather than this one. The framework below covers choosing the first employee rather than the first task.

## FAQ

### How do I know if AI can do a specific task?

Ask four questions. Can you describe the task in five lines covering trigger, inputs, good output, tone, and stop line. Is every input written down somewhere reachable rather than in your head. Can you tell a bad result from a good one in under two minutes. And what does undoing a mistake cost. Four yes answers means hand it over. Three means fix the failing one first. Two or fewer means the task itself is undefined.

### Which question do most tasks fail on?

Question two, by a wide margin. A key input exists only in someone's memory, on a whiteboard, or in a filing cabinet, so the AI produces something plausible and wrong. The fix usually takes under an hour of writing the rule into a document or a record, and it improves the task even if you never automate it, because a new human hire would have hit the same wall.

### Does a hard or complex task automatically fail?

No, and this surprises people. Complexity is not the constraint. Describability, reachable inputs, and checkable output are the constraints. Plenty of tasks that feel complicated pass easily because they follow a clear rule, while simple tasks fail because one input lives in your head. Score the task rather than trusting your intuition about how hard it feels.

### What if a task only partly qualifies?

Split it. Almost every borderline task is really two tasks stapled together: a mechanical part with clear rules and a judgment part that needs you. Hand over the mechanical half and keep the judgment. Quoting is the classic case, where reading the spec, calculating materials, and drafting the document all pass, and only the final number needs your read on the customer.

### How does the undo cost change what I do?

It does not decide whether to hand the task over, it decides whether the human approval step is temporary or permanent. A cheap, reversible mistake such as a bad draft or a mislabeled email means you can remove the gate after a good week. An expensive or irreversible mistake such as money leaving, a contract signed, or a record deleted means the gate stays forever, regardless of track record.

### What if I score a task wrong?

It costs you a week of reviewing drafts, which is why the approval step during the first week is not optional. A badly fitting task announces itself quickly: you find yourself rewriting every draft, and you will know by about day three. That is a far cheaper and more reliable way to find out than a long planning exercise, because the test only really settles once real work runs through it.

Whether AI can do your task depends on the task, not the technology. Describable, reachable, checkable, and reversible are the four properties that matter. Score your three worst jobs on paper this week, and act on whichever one scores highest rather than whichever one annoys you most.

**Tags:** can-ai-do-my-task, ai-decision-test, ai-delegation, ai-employee, task-automation