Sorting whole tasks into AI or human gets the answer wrong, because most tasks are several different steps. Sort the steps instead, with a sheet and prompts to paste.
Split each task into steps. Hand AI the steps that end in material, keep the judgments, and keep every commitment under a person’s name. Give each AI step a named checker and a set moment, and review the line quarterly.
Most advice about what to give AI sorts whole jobs and whole tasks: summarising is fine, hiring is not, drafting is fine, strategy is not. It sounds sensible and it falls apart on an ordinary Tuesday, because few tasks are one thing. A weekly update is part gathering, part deciding what matters and part sending something with your name on it. AI is good at the first part, uneven at the second, and the third is a promise only you can make.
A team lead asks AI to draft the Friday update for leadership. It pulls the project notes into a clean summary in a minute, puts the items in a sensible order and writes in a calm, confident tone. What it can’t know is that one project is quietly at risk, because the evidence is a hallway conversation that never made it into the notes, so the draft says the project is on track. The lead reads it as a summary, sends it as a status, and the writing was never the problem. A judgment about what was true got handed over without anyone deciding to hand it over.
The claim this guide rests on is that you should give AI the steps that end in material and keep the steps that end in a judgment or a commitment. Material is anything you’ll read, check and edit before it matters: a summary, a list of options, a first draft. A judgment is deciding what the material means. A commitment is the point where something leaves your hands and someone else starts relying on it. Below you’ll find a way to sort any task like this, a filled-in sheet for a manager’s week, and prompts you can paste into your own work.
Take any task and write its steps in the order they happen. Then ask one question of each step: what does this step end in? Nearly every step ends in one of three things.
| A step that ends in | What it looks like | Who takes it | Examples |
|---|---|---|---|
| Material | Something you will read, check and edit before it matters | AI does the work, you review it | Summarising notes, listing options, a first draft, pulling figures into a table, reformatting a document |
| A judgment | A call about what the material means or what is good enough | You. AI can argue the other side, but the call is yours | Which risk matters, whether a number is believable, who to interview, what to leave out |
| A commitment | Something leaves your hands and someone else relies on it | You, by name | Sending the update, approving the spend, telling someone their role is changing, signing off the figure |
A Future Factors working method, based on our own judgment and not a research-validated model.
The sort works because it follows where the damage happens. A mistake in material is cheap while it’s still material, since you’re about to read it anyway. The same mistake inside a judgment gets quiet, because a judgment looks like a conclusion. Inside a commitment it becomes someone else’s problem. So the useful question isn’t how hard the step is, and it isn’t whether AI is able to do it. AI can send an email. The question is who answers for it afterwards.
AI can prepare anything. A person makes every commitment.
Two things keep this from becoming a slogan. First, steps move. A judgment you’ve made the same way many times, against a clear written standard, can become material that you review. Second, this sits one level below the question of whether a task suits AI at all. For that, our fit test for any task scores the whole task. Use it to decide whether to bring AI in, and use the sort here to split the task you’ve decided to use it on.
The quickest way to learn the sort is to run it on last week’s calendar. A first-line manager’s week tends to hold the same handful of recurring jobs, and most of them split cleanly once you break them into steps. Here is a filled-in sheet for five of them, in the three-column shape we use whenever AI takes on real work.
| The job | What AI does | What the human still owns | How it gets checked |
|---|---|---|---|
| Weekly update to leadership | Pulls notes, tickets and messages into a draft in a fixed format | Which items are really at risk, what to leave out, and sending it | Read against your own list of open risks before it goes. Any “on track” you can’t back up gets changed |
| Preparing for a one-to-one | Summarises the person’s recent work and open actions from notes you supply, and suggests questions | What the conversation is for, and anything about performance, pay or wellbeing | You read the source notes behind anything the summary says about the person |
| Hiring shortlist | Puts CVs into one layout and lines up stated requirements against evidence | Who gets an interview and why, and the ranking itself | Two people shortlist separately from the same CVs and compare. Local rules on automated screening may apply, so check them |
| Month-end variance commentary | Drafts commentary from the figures you paste in | Whether the figures are right, which explanation is true, and the sign-off | Every number traced back to the source sheet, and the finance lead signs |
| Reply to an unhappy customer | Drafts two versions at different tones from the facts you give it | What you’re willing to offer, and pressing send | A second person reads anything that makes an offer or admits fault |
A worked illustration of the sheet, not measured data. Swap in your own jobs, owners and checkers.
Look at what the sheet does and doesn’t do. AI takes real work in all five rows, usually the part that eats the most time. What it doesn’t do is send, approve, rank, promise or decide. That split is what makes it safe to hand over the time-consuming part.
A finance lead pastes a variance table into a chat and asks for commentary. The draft says marketing overspent because of the timing of campaign launches. It’s plausible and well written, and it’s a guess, because the table shows the overspend and says nothing about the reason. The sort catches it at the right spot: the numbers are material, the explanation is a judgment, and she’s the one who knows whether it’s true.
To run this on your own week, paste the prompt below, then work through the short routine after it.
| Here is a list of the recurring tasks in my week: [paste five to eight]. For each task, break it into the steps in the order they happen. Tag every step as Material (something I will read, check and edit), Judgment (a call about what the material means or what is good enough) or Commitment (something that leaves my hands and someone else relies on it). Then give me a table with three columns: what AI could do, what I should still own, and how I would check the AI’s part. Flag any step where I might be tempted to skip the check because the output looks polished. If you can’t tell from my description which tag a step deserves, ask me instead of guessing. |
Treat the tags as a first pass. You know which judgments are routine and which are not, and the model doesn’t.
If you’d rather build one of these rows into a full repeatable process, our guide to choosing your first AI-enabled workflow walks through one end to end. For writing the delegation brief itself, see using AI to delegate work without losing control.
Managers tend to assume that people either want AI to take as much as possible or want it kept away from their work. The people doing the work are more particular than that. A Stanford team asked 1,500 workers across 104 occupations about 844 tasks, and the workers had a positive attitude toward AI agents automating 46.1% of them, even after being asked to weigh concerns such as job loss and enjoyment.[1] The reason chosen most often, in about 69% of the responses where people wanted automation, was that it would free up their time for higher-value work.[1]
I’d take the shape from this more than the percentages. The same study asked how much human involvement people wanted, and for the largest group of occupations, 47 of 104, the preferred answer was an equal partnership, where person and agent together outperform either alone.[1] That is worker-reported preference set against expert ratings of capability, so it tells you where people want the line, not where it should sit. It does line up with the sort above: hand over the steps that eat time, and stay in for the steps that need you. The authors also report early signs that valued skills may shift from processing information toward interpersonal and organisational ones, which they present as a signal and not a settled trend. Our guide to human skills in the AI era goes into which of those skills are worth building.
The practical use is a ten-minute exercise before you finalise anyone’s sheet. Ask each person on your team to list the two steps they’d most like off their plate and the one step they’d want to stay in on. Then compare their answers with yours. They’ll differ by role, which is a good reason not to write one AI policy for a whole department.
The step that gets skipped first is the review. Output that reads well invites a quick skim, and a quick skim is how material quietly turns into a judgment nobody made.
A Microsoft and Carnegie Mellon survey of 319 knowledge workers, who shared 936 first-hand examples of using generative AI at work, found that higher confidence in the AI went with less critical thinking, while higher confidence in their own ability went with more.[2] It’s self-reported, so it describes how people say they work and not a measured drop in quality. The workers also described their role shifting toward verifying, integrating and stewarding what the AI produced.[2] That fits what teams tend to see: the more a step goes right, the less checking it gets.
So the check has to be designed into the step. Waiting for a feeling that something’s off doesn’t work, because a polished draft doesn’t produce that feeling.
Every step you hand to AI gets a named checker and a set moment.
Three things make a check real:
You can also use AI to point at what to check, as long as it isn’t the one deciding whether the evidence holds.
| Here is the material I prepared with AI and the call I’m about to make: [paste both]. Don’t tell me whether my decision is right. List (1) the three claims in the material I would most want to verify before relying on it, and where I would check each; (2) what would have to be true for my call to be wrong; (3) one person who would object and what they would say; (4) anything that sounds certain but has no source in what I gave you. If a point depends on information you don’t have, say so. |
The model can find what to check. Whether the evidence holds is still your call.
A commitment is the step where being wrong gets passed along, which is why it stays with a person whatever helped to draft it. A 2024 tribunal decision in British Columbia shows how that looks from the outside. An airline’s chatbot told a customer he could claim a bereavement fare after he had travelled, while another page on the airline’s own site said otherwise. The airline argued the chatbot was effectively a separate entity responsible for its own actions. The tribunal rejected that, said it should be obvious the airline is responsible for all the information on its website, and found it liable for negligent misrepresentation.[3] It’s one tribunal’s ruling in one jurisdiction, and I wouldn’t treat it as a global rule. The reasoning is the part worth carrying: the people on the receiving end of a commitment won’t distinguish between what you wrote and what your tool wrote.
If an agent or an automation does the sending, the commitment was made when someone approved the automation, and that approval deserves the same care as the message. Before a step runs without review, ask three questions:
Two or three honest noes mean a person goes back into that step. Where the line falls also depends on the role, because the same step carries different weight in different seats.
| Role | Steps that usually move to AI | Steps that tend to stay with the person | Why |
|---|---|---|---|
| Team lead or people manager | Gathering status, drafting updates, summarising one-to-one notes | Feedback, pay and promotion conversations, who gets which assignment | Each one changes someone’s working life |
| Finance lead | Variance tables, first-pass commentary, reconciling formats | Which explanation is true, forecast assumptions, sign-off on figures | The signature is what carries accountability |
| HR partner | Policy FAQ drafts, grouping survey comments, formatting job descriptions | Case outcomes, anything touching employment status, sensitive disclosures | The consequences are hard to reverse and often legally regulated |
| Operations manager | Process documentation, incident summaries, schedule options | Which risk to accept, safety and compliance sign-off, vendor commitments | Others act on these without re-checking them |
A starting point based on our judgment. Your regulator, contract or policy may move a line in either direction.
AI can still help on the right-hand side as a preparation partner. It can organise your notes before a hard conversation or play the other person in a rehearsal, as in our guide to preparing for a difficult conversation. The conversation, and the decision behind it, stay with you.
The line shouldn’t stay where you first drew it. A step you reviewed closely in October may be reliable enough to spot-check by January, and a step you handed over may need to come back after the tool or the inputs change. So review the sheet on a schedule, and measure it by use, persistence and impact, not by whether the tool got opened. Use: is the sheet being followed for the steps on it? Persistence: is it still being followed a month later without a reminder? Impact: did the work change, with fewer corrections, a faster turnaround, or a mistake caught before it went out?
| Question | Yes | No |
|---|---|---|
| 1. Did the checker catch something in the last month, or confirm on purpose that the output was clean? | 1 | 0 |
| 2. Is the check still happening without a reminder? | 1 | 0 |
| 3. Are the tool, the inputs and the stakes the same as when you set the line? | 1 | 0 |
| 4. Did the step give back time that went to a judgment step? | 1 | 0 |
| 5. Would someone notice and mind if the step went back to being manual? | 1 | 0 |
| Reading the score | 5: keep the line. 3 to 4: keep it and fix the weakest answer. 0 to 2: pull the step back toward you and rerun it for two weeks. | A step can only move further toward AI after a 5 and at least two clean checks. |
A Future Factors working tool. Run it once a quarter, one row of the sheet at a time.
The line between you and AI is a decision you revisit, not a setting you leave.
For a team, the same habit has a longer path. Training builds the capability, so people learn to sort their own work into material, judgment and commitment. Workflow design gives that capability somewhere to be used, for example a sheet attached to every recurring job. Reinforcement turns it into behavior, such as a manager asking to see the sheet in the monthly review. Adoption is what you see when teams keep updating the line without being asked. A team can start on its own with one shared sheet. Working through your team’s real tasks with an experienced trainer shortens the learning curve and gets everyone to the same standard, instead of leaving it with one enthusiastic sorter.
This week, pick five recurring tasks and run the sort prompt on them. Mark every commitment, name a checker for every AI step, and book the review for two weeks from now.
It can be, for steps that are internal, low in stakes and easy to undo, such as a reminder or a routine status note. The commitment moves to whoever approves the automation, so write down who that is and what the automation may not send. Check your organization’s policy first, since some require review for anything external.
Ask them to bring theirs and compare the two. Differences usually show where someone knows something you don’t, such as a step that looks routine and isn’t. Run the comparison openly as a conversation about the work, not as a compliance check, or the sheets will go underground.
Yes, when it’s routine, you have a written standard, and you’ve compared the AI’s version with your own on enough real cases to see where they differ. Even then, keep a small sample check, because the inputs change after the standard was written.
A leader’s week holds more judgments and commitments, so the risk is approving an automation or a draft without really reading it. Someone earlier in a career holds more material, so the risk is forwarding output upward without knowing which parts are checked. The same three-way sort works for both, and the checker is usually the difference.
It can prepare, by organising your notes or helping you rehearse, but the decision and the conversation stay with a person. Rules on automated decision-making in hiring and employment vary by place and may apply to ranking or screening, so check yours before any AI output influences who is selected, promoted or let go.
Sources were read on October 14, 2026: Shao and colleagues on automation and augmentation across the U.S. workforce (arXiv, 2025, version dated February 2026), Lee and colleagues on generative AI and critical thinking (CHI 2025, via the Microsoft Research page), and the British Columbia Civil Resolution Tribunal decision in Moffatt v. Air Canada (2024 BCCRT 149), which we read through a law firm’s summary because the CanLII page blocks automated access. The material, judgment and commitment sort, the manager’s sheet, the line review and the role table are Future Factors working methods based on our own judgment, not research-validated models. The scenes in this guide are illustrations, not client stories. We left out task-time and productivity figures from workplace AI studies, and reported refund amounts from consulting-report cases, because we couldn’t confirm them against a primary source today.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.