Handing Work to AI Starts with Understanding the Task

“AI handles execution; people define the request and judge the result” sounds like giving everyone a capable assistant. But a capable assistant still cannot read the thoughts we leave out of the brief.

Take “organize the project materials.” For an archive, the first job is to collect the files. For a discussion about whether the project should continue, we need to establish what has been achieved and what problems remain. The instruction is unchanged; the work it calls for is quite different.

Suppose AI is just as intelligent as we are, but on this assignment it can learn about the task only from what it receives. A manager knows why the work matters and what they want from it, while the person doing it may receive just one small part of the assignment. With different background information, they may mean different things by “done.” AI does not make that problem disappear.

In practice, AI can consult documents and use tools to learn more. Still, the person handing over the work needs to think through what the result will be used for. Being there does not guarantee that we have.

Who are these minutes for?

If you missed a meeting, you may want to know why the issue came up, where people disagreed, and what they decided. If you need to move the project forward, you may be more eager to find what has been agreed and what still needs discussion.

Imagine a meeting that spends most of its time discussing alternatives and settles the next step at the end. A summary that allocates space in proportion to speaking time could cover the discussion thoroughly while tucking the action into a corner. That may work as an account of the meeting. A project lead trying to act on it still has to find the actions.

Looking at that draft, it is easy to feel that AI missed the point. But if we meant “help the lead act” and asked only for a “complete record of the meeting,” the purpose stayed in our heads. The model may have carefully organized what everyone said and still failed to do what we expected.

I use 5W1H in my courses and when organizing information. WHY, WHAT, and HOW also remind me what to explain when handing over work. They help me notice what I have not thought through. For example, start with “Prepare this for the project lead to act on; separate agreed actions from unresolved questions,” then decide which material to provide and what to emphasize. That tells the model more about the job than another set of adjectives such as “complete, professional, and clear.”

If the purpose was explicit and the result still misses it, check whether the model overlooked or misunderstood the instruction. People can leave a brief unclear; models can fail to follow a clear one. We need to find out what happened this time.

Was it left out, or never decided?

A clear purpose can still leave gaps in the source. Suppose the meeting record says only, “Finance needs another look at the proposal; we'll return to it next week.” That supports an unresolved review item. It does not establish that “The finance lead will deliver the final proposal by Friday.” The original wording gives no owner, deadline, or requirement for a final version.

If the meeting did settle the owner and the transcript missed it, a recording or another record may recover the decision. If nobody decided during the meeting, listening again will not produce an owner.

The same cell is blank in both cases, but the next steps differ. One needs more information; the other needs someone to decide. Asking AI to “complete” both can quietly turn organizing a record into assigning work on the organization's behalf.

Once the table is full, the person taking over may not ask where every entry came from. They may assume it was agreed at the meeting and plan around it. An unresolved question now has an apparently definite answer, making it harder to notice.

I keep what the material supports and distinguish missing evidence from matters still undecided. If an old proposal conflicts with the new record, I retain the statements and their sources. A more complete old proposal does not become the current decision by default.

The settled material can still be organized. For the remaining item, the next person needs to know whether to recover an existing fact or ask the relevant people to decide.

Before asking for another version, say what is wrong

“This isn't what I wanted; organize it again” leaves AI guessing. Is the emphasis wrong? Was material missing? Did an action get included even though it was never agreed? The next version may change the headings and layout while leaving the problem untouched.

Stay with these minutes. If they accurately record the discussion but bury the actions, check whether the handoff explained their intended use. If it did and the model missed the instruction, correct that part. If an owner appears without evidence, return to the source. However suitable the person seems, that alone does not justify leaving the assignment in an official record.

In Guidelines for Human-AI Interaction (CHI 2019), Amershi and colleagues recommend making errors easy to correct and clarifying goals or limiting services when the system is uncertain.^[1] For these minutes, that could mean keeping what is right and showing the questionable action beside the relevant original wording, so a person can continue from there.

The finished draft still needs to be checked against its source and purpose. After writing to an online document, check that the important content actually made it in. But each check answers a particular question. If we misunderstood the assignment at the outset, good formatting and complete fields may simply carry out that misunderstanding very conscientiously.

In future reviews, I want to look at the original request alongside the reason for rework. Was the purpose unstated, material not provided, a decision still unmade, or an instruction not followed? Handoff practices that repeatedly help can go into a procedure. Then, next time, we know what to add or correct, instead of starting again with “Please make sure you understand what I mean.”