Imagine your manager asks an AI assistant to “help us research options” and leaves you to figure out whether the answer is finished.
A comparison arrives. It has links, a confident recommendation, and a closing offer to contact the winner. You were expecting a summary of the notes your team already collected. Your manager was expecting a purchase decision.
The assignment sounded clear until somebody had to accept the work.
If you run operations, that last part is familiar. An exciting idea arrives from above. A pile of output arrives from the tool. You inherit the gap between them.
Here’s the thing. “Help us research options” describes an activity. It leaves the job open at both ends. Where does the information come from? What decision is this preparing? And at what point does the assistant stop?
You can spend the afternoon improving that comparison. Or you can settle the assignment that keeps creating the afternoon.
Three jobs are hiding inside that sentence
Stay with this hypothetical research assistant. No particular company or product here. Just a common kind of internal task.
Gathering information is one job. Comparing approved options is another. Deciding which option the company will buy is another. The work can connect, but each part needs a different boundary.
An assistant asked to gather information may need to search widely. An assistant comparing supplied notes needs to stay inside those notes. The person buying needs to consider commitments and information that may never belong in the assistant’s assignment.
“I thought it was obvious we only wanted a comparison.”
Obvious to whom? You can literally point to the missing words. The assignment never said “from these notes.” It never said “for my decision.” It never said “do not contact anyone.”
That does not excuse a bad answer. It tells you what your side of the work still needs.
In the Professional Recipe, Context means the details of this particular request. Format means the finished thing that comes back. Guardrails are the hard stops. A vague research request asks one sentence to cover all three without actually defining any of them.
The job is what the station owns and what it produces. If you cannot write it, you cannot hire it.
Pick the piece of work you can actually hand over
The AI Employee Handbook starts with a page for each role. Here, assume you have picked the role and the person who checks it. The work now is deciding what belongs on that page.
Try this assignment:
“Compare the options named in these approved notes and recommend one against the stated requirements.”
Better. We have a starting point, a limited set of choices, and a result somebody can use. We are preparing a human decision.
For the rest of this example, call the choices Option A and Option B. The manager needs a weekly list export and separate access for people who only view the work and people who can change it. Those are the two requirements in the request.
The approved notes say both options support the weekly export. They describe separate view and edit access for B. They say nothing about that requirement for A.
The assistant can compare what the notes support. It can say B is the better-supported choice against these two requirements. It cannot decide that A lacks a feature just because the notes do not mention it.
Right? “We do not know” and “it cannot do that” would lead the manager to different decisions.
And if the notes do not support a recommendation at all, the answer needs to say that. Recommending one is the goal when the evidence allows it. Picking a winner at any cost is a different assignment.
Make the starting material part of the job
A request with missing inputs can still produce a full page. That is what makes the problem easy to miss.
Write down the required inputs: the named options, the approved source notes, the requirements, and the manager who will use the comparison. Each request needs that set.
For our example, the weekly export requirement belongs in the request. The assistant should not infer it because the company usually asks for exports. A different comparison next month may have a different purpose.
This is the distinction between the house and the current order. Training, in the Professional Recipe, describes who the company is and what it stands for. Context carries the requirements for this particular comparison. The job description tells the assistant which current details it needs before it can do useful work.
“Use your best judgment about what matters.”
Sometimes that is permission you want to give. Here it would quietly turn a comparison assistant into the person setting your buying requirements.
Say what is missing instead. A visible question is something the manager can answer. An invented requirement can steer the whole comparison without anyone noticing.
Describe the thing you want back
Now decide what will land on the manager’s desk.
For this job, the output has four parts: a recommendation, the reason for it, references to the supplied evidence, and unanswered questions. If the evidence is insufficient, the recommendation says so and explains what prevents a choice.
That is specific without dictating every step the assistant takes. You do not need to require three rounds of thinking, five headings, and a particular order of keystrokes. You need a comparison the manager can inspect.
In our example, a useful answer might say:
“Option B is the better-supported choice against the two stated requirements. Both options have a weekly export in the notes. Only B’s notes confirm separate view and edit access. A’s access options remain unanswered.”
Then the references point to the relevant sections of the supplied notes.
The manager can see the reason, check it, and decide whether to gather more information. The uncertainty has survived the trip from source notes to recommendation.
Does that make sense? The assistant has done a useful job without pretending it has finished every job around it.
The end of the job belongs on the page
Here’s the thing. A narrow task can still drift if you only describe what to do.
Write the forbidden work just as plainly. Do not purchase. Do not contact a vendor. Do not invent a missing requirement. Do not turn an unanswered question into a fact.
“But contacting the vendor would help finish the comparison.”
It might. The assistant can identify that need. The manager decides what happens next.
You are making a choice about responsibility here. Calling a vendor creates a conversation somebody has to own. Purchasing creates a commitment. Both might eventually be appropriate work for another role, with its own rules. They are outside this one.
Yeah. Helpful can still be outside the job.
That boundary also protects the operations person who would otherwise discover the extra work after it happened. The assistant’s offer to do more should arrive as a proposal for the manager, not as news that a new task has already begun.
”Done” needs something you can point to
There is one more loose phrase to remove.
“Send me a useful comparison.”
Useful matters. On its own, it gives the checker very little to work with. They might mean short. They might mean comprehensive. They might mean “choose the option I was already leaning toward.”
Try a completion rule you can inspect: every named requirement is addressed for each option; claims point to supplied evidence; unanswered questions remain visible. The manager who requested the comparison checks those things and either accepts it or returns it with the missing part named.
That is not the same as proving the eventual purchase was successful. No. Narrower than that. It establishes whether this comparison did the job it was given. A sound comparison cannot promise that a tool will solve every problem the company later brings to it.
The choice of wording and the examples of an acceptable comparison deserve their own attention. For now, get the boundary down. What is this station responsible for delivering, and what can its checker refuse?
Your Monday Move
Pick one AI-assisted task you already use and take its normal checker through one page. Keep these six fields:
- Station name: Options Comparison.
- Work owned: Compare named options from approved notes and recommend against stated requirements when the evidence supports a choice.
- Required inputs: Named options, approved source notes, stated requirements, and the manager using the comparison.
- Output: Recommendation or reason a choice cannot yet be made, supporting reason, source references, and unanswered questions.
- Forbidden work: No purchases, vendor contact, invented requirements, or unknowns presented as facts.
- Checker and completion rule: The requesting manager checks that every requirement is addressed, claims have supplied evidence, and unknowns remain visible.
Replace the example with your actual task. Do not leave any field blank. Then read it to the checker and ask them to point to where the job ends and name one reason they would return the work.
If you disagree, you have found the management decision to make before another round of output. Settle that decision on the page. No new tool or permission is needed for this exercise.
You should finish with an assignment another person can recognize as complete. That is the work you were trying to hand over in the first place.
The job is what the station owns and what it produces. If you cannot write it, you cannot hire it.
Original framework. Distilled from client work.
