Hand a task to an AI agent when it passes three tests: it repeats, you can undo a mistake, and you can check the result without redoing the work. Score every task you do this week on those three questions, and by the end of this post you will have a sorted list and a first task to start with today.
Three yes or no questions decide almost every task
Published delegation frameworks land on nearly the same tests. Digital Applied's guide to what to delegate first looks for work that is recurring, bounded, reversible and verifiable. Fold "bounded" into "repeats" and you get three questions a solo founder can answer in a minute.
- Does it repeat? It has happened at least twice and will happen again. A one-off is a poor candidate, because you spend longer building the handoff than doing the job.
- Can I undo it? If the agent gets it wrong, nothing has left your hands yet. Drafting an email is undoable. Sending it is not.
- Can I check it fast? If checking takes about as long as doing the task, you have not saved anything. Keep that task.
Score each answer 1 for yes and 0 for no. Then use the score.
| Score | What it means | What to do |
|---|---|---|
| 3 | Repeats, undoable, quick to check | Hand it over. Spot-check the output. |
| 2 | One test fails | Hand over a draft only, with a review step before anything goes out. |
| 0 or 1 | Two or more tests fail | Keep it. Do not delegate. |
The score alone is not enough when the failed test is "undoable". A task that repeats and is easy to check but ends in a send, a payment, a delete or a publish still needs a human to pull the trigger. Researchers writing on AI delegation say irreversible tasks with real-world side effects, such as sending an external email or deleting a database, need stricter safeguards than reversible ones such as drafting an email. So the agent drafts. You send.
Run the test on your own task list in ten minutes
Example: say you run three small apps and write down eight recurring jobs. Here is how a list like that might score. The tasks and scores are illustrative. Use your own.
| Task | Repeats? | Undoable? | Quick to check? | Score | Decision |
|---|---|---|---|---|---|
| Weekly summary of support emails | Yes | Yes | Yes | 3 | Hand over |
| Sort new bug reports into labels | Yes | Yes | Yes | 3 | Hand over |
| Draft replies to user emails | Yes | Yes (draft only) | Yes | 3 | Hand over, you press send |
| Weekly status note to yourself | Yes | Yes | Yes | 3 | Hand over |
| Refund a customer | Yes | No | Yes | 2 | Draft only, you approve |
| Change prices in your billing tool | No | No | Yes | 1 | Keep |
| Delete old user accounts | Yes | No | No | 1 | Keep |
| Decide which app to cut this month | No | No | No | 0 | Keep |
The first four are the kind of work published guides call strong first candidates: recurring tasks with bounded inputs and outputs a human can check, like a weekly status update or a recurring summary that gets a quick human review before it is shared. Classifying email, extracting data from documents and building reports from templates sit in the same group, according to Coderhouse's guide for non-technical teams.
Copy this line for each task in your own list: Task / repeats? / undoable? / how I check it / decision. The "how I check it" column matters most. If you cannot fill it in with a sentence, the task is not checkable yet.
Write the check before the agent starts, not after
Pick one task from your "hand over" pile. Before you run anything, write one sentence that says what "done correctly" looks like. That sentence is your check. Without it you end up reading the whole output and guessing.
A short return checklist keeps your review small. This piece on avoiding endless review suggests three questions: what outcome should this produce, which two or three things must be correct, and what should the AI flag instead of deciding alone. After that you only check those few conditions.
Paste this above any task:
Task: [one sentence]. The result must: [outcome]. These 2 or 3 things must be correct: [1], [2], [3]. If any of these is unclear, flag it and stop. Do not decide it yourself. Do not send, publish, pay or delete anything. Return a draft.
If your review still means reading every line, the checklist is too long. Cut it to the two or three must-be-correct items, or take the task back.
Give the agent more freedom in three stages, not all at once
Passing the test tells you the task is a good candidate. It does not tell you the agent is reliable on it yet. Earn that with a ramp. ATXP's guide to ramping agent autonomy describes starting with constrained scope, such as read-only access and human confirmation, and expanding each stage only after the previous one has been verified. AWS describes the same pattern: agents earn wider permissions through steady reliability and lose them when performance drops.
Here is a version sized for one person:
- Stage 1: draft only. The agent reads and drafts. You approve every output. Stay here for your first several runs.
- Stage 2: small items alone. The agent handles low-risk items by itself. You spot-check a sample and review every high-value item.
- Stage 3: runs on its own. You watch for drift and keep the right to drop back to stage 1.
Move up only after the stage below has held up. Move down the moment your checks catch a miss. A task that worked in week one can drift later, and if nobody is looking you find out from a user.
Sample by risk, not by a flat rate
Do not review every item the same way. Techstrong's piece on evaluating an unpredictable agent says to define "good" with a rubric and sample by risk: higher-value items get more checks, and low-risk repetitive work gets spot checks until the metrics drift. Pick your own numbers and write them down before the first run. For example, you might check every refund draft and one in ten ticket labels. Those numbers are a starting point. Change them after a month.
Skip these four mistakes
| Mistake | What it costs you |
|---|---|
| Handing over a vague task with no stated outcome | You re-read everything to guess if it is right, and the time you saved disappears. |
| Letting the agent send instead of draft | One wrong external message cannot be recalled. |
| Starting with the hardest task | You test your setup on the one job most likely to fail. Start with a recurring, bounded one. |
| Reviewing everything the same way | Review becomes the new full-time job. Check the two or three things that must be correct. |
On first tasks, Boei's list of tasks to delegate gives a useful rule: start with one narrow, repetitive task you do the same way every day, and have the agent draft instead of send. Their list is written for customer service, so treat it as examples, then run your own tasks through the three questions.
Tasks that fail the test have a home
Some jobs should stay with you no matter how good the agent is. Anything that approves budget, sends a customer-facing proposal or publishes outside communication is high-risk, and a human owns it. If a task keeps landing in the "keep" pile, read what you should never automate with an AI agent for the five jobs I would keep a human on.
Once tasks are handed over, you still need to know when one goes wrong. How to monitor AI agents in production when you run several apps alone covers the alerts that tell you a task should drop back a stage. If you are running more than one agent, manage multiple AI agents without losing track covers the list that keeps them straight.
Do this in the next ten minutes
- Open a blank note and list every task you repeated this week across your apps.
- Score each on repeats, undoable and quick to check.
- Pick one task that scored 3. Write the one sentence that says what "done correctly" looks like.
- Paste the return checklist above it and run it at stage 1: draft only.
That is the whole setup. One task, one check sentence, one stage.
Stuck on a task you cannot score?
If you have a manual job that keeps landing at a 2 and you cannot tell if it is worth handing off, email me the task. Tell me what it is, how often it happens and what goes wrong when it fails. Reading the guides here costs nothing, and so does asking.
