The AI is being considered for more work. The person fixing its comparisons is still being asked why they take so long.
Imagine that conversation in an internal review. The assistant compares named options from approved notes and recommends against the team’s stated requirements. Its job is clear. A manager checks the result before using it. Completed comparisons keep arriving.
“We could give it more of these.”
Then the operations colleague opens one of the finished comparisons beside the version that first came back.
“I still have to rebuild the references.”
The recommendations may be useful. But the links are too vague for a reader to trace the supporting claims, so someone goes back through the notes and repairs them. In this hypothetical team, that has become a recurring part of getting the work ready.
You have been reviewing what the AI produces. The person beside you has been carrying what it takes to use it.
Before giving the role a bigger job, put those two things in the same conversation.
Review the job it has now
Here’s the thing. An AI role can become permanent without anyone deciding it deserves to be.
The first version worked well enough to keep trying. People learned how to get around its rough edges. The corrections became part of the routine. Now the role is established, and the review starts with what else it could do.
That skips the question the existing work should answer.
“Are the comparisons helping us decide?”
Not just arriving. Helping the manager compare the named options, understand the evidence, and decide what happens next. The job description gives you the purpose to return to before deciding whether the role should grow.
The handoff tells you whether a particular result is ready for its intended use. The performance review asks whether producing results this way continues to be worthwhile.
In The Professional Recipe, Feedback has two parts: look at evidence of what is happening, then revise the relevant part of the recipe when that evidence shows a problem. A recipe is the repeating job written down. Your review needs both the evidence and a decision about what it means for that job.
The role has to keep earning its place.
Start with the work you already have
You do not need a new dashboard for this conversation.
Choose a completed period of ordinary work. Write down the dates, which comparisons you are reviewing, and where your records are incomplete. Use the source notes, corrected versions, review records, and decision notes you already keep.
Then look at three things together.
Did anyone use the result? Find where a comparison helped the manager make or advance the decision it was prepared for. Maybe it supported a choice. Maybe it exposed an unanswered requirement that the manager needed to resolve first. Ask the recipient to explain what it helped them do, and look for that in the decision record.
A file being opened does not establish that. Neither does a thank-you message. You are trying to find the work the comparison contributed to.
What did people put into making it useful? Include preparing the source material, reviewing the result, and correcting it. In our example, rebuilding the references belongs here. So does any work required to prepare the approved notes the assistant receives.
If time is recorded, use it. If someone is estimating, label it an estimate. Do not turn a remembered impression into a measured saving. Without a relevant comparison of the work with and without the assistant, the time benefit remains unknown.
Which material corrections keep coming back? Count documented repair types across the work you examined. A missing source reference matters because it prevents the reader from checking the recommendation. A preferred opening phrase may not change the usefulness at all.
State how much work you actually reviewed. The pattern in those cases is evidence about those cases. A few good examples do not establish a reliable performance rate for every future assignment.
Right? These measures need each other. Little correction work could mean the role is doing well. It could also mean nobody relies on its output enough to check it.
Missing evidence is an answer you can act on
Suppose the team kept the final comparisons but not the corrections. The manager remembers finding some useful, but cannot identify which decisions they informed. Nobody recorded the preparation effort.
You have some material to examine. You do not yet have enough to decide that the role is saving time or deserves more responsibility.
“We do not have enough to tell yet.”
That is an honest finding. Name the missing piece and how you will capture it through the existing process. Perhaps the next review needs the initial and corrected versions together, or a note about how each comparison was used.
Set a point to return to the question. Until then, leave the existing limits and review requirements in place. Do not convert missing evidence into a confident decision to expand, retire, or declare success.
And do not let a tidy average hide a serious miss. If a comparison made an unsupported claim that could change the decision, examine it even when most of the other work looked fine. The consequences matter alongside the count.
The review has to permit more than “improve it”
Once there is enough evidence for the decision you are making, you have options.
Keep the role as it is. The comparisons are helping with the intended decisions. The preparation and checking are worth the benefit. There is no unresolved problem that makes continuing the role unreasonable.
“Keep this job as it is.”
That can be a good performance-review outcome. You do not owe the meeting a new improvement project just because the meeting happened.
Revise or narrow the role. The comparison work is useful, but a recurring problem is adding avoidable effort. Find where it starts before choosing the fix.
In our imagined case, perhaps the approved notes do not include usable source references. The input needs attention. Or perhaps the references are available, but the instruction for the finished comparison is vague about making each claim traceable. That calls for a different change.
Check which situation you actually have. Adding another instruction will not manufacture information that was never supplied. And if the role serves a smaller, clearer class of comparisons well, you can propose keeping it within that class while addressing the wider problem.
Propose retiring the role. Perhaps the evidence shows the team seldom uses the comparisons, or preparing and correcting them takes effort that the benefit does not justify. You are allowed to decide this is no longer a worthwhile way to do the job.
Before acting on that decision, name what happens to the underlying work. Does an existing process take it back? Does the need itself no longer exist? Retiring an AI role should not leave a necessary job waiting for someone to notice it.
Yeah. The role can leave. The team’s need still deserves an answer.
Make the next review able to change your mind
Here’s the thing. “We updated the instructions” is not the result you were trying to get.
For the source-reference problem, the intended improvement is that a manager can trace the comparison’s claims without someone rebuilding the references first. After an accepted change, look for that in subsequent ordinary work. Watch whether it introduced some new preparation burden elsewhere.
You are testing the fix. Let me be more precise. You are checking whether the whole job improved, including the work people do around it.
Choose the next review point to match the role’s actual use and the importance of its mistakes. A role used occasionally may need longer to produce enough evidence than a role used throughout the week. The calendar appointment alone does not make the evidence sufficient.
If the repair did not help, revisit the diagnosis. If the role now earns its place, keeping that scope is still a valid choice. A good review can support discussing one bounded next assignment, but that is a separate decision with its own requirements.
Does that make sense? Improving the current role does not tell you how it will perform on work it has never owned.
Your Monday Move
Pick one existing, low-risk internal AI role. Bring its responsible leader together with the colleague who prepares, checks, or repairs the work.
Use a defined, completed period and the records you are allowed to review. Look at actual use, surrounding human effort, and recurring material corrections. Mark estimates and gaps plainly. You are literally putting the evidence of benefit beside the evidence of burden.
Write one decision: keep the current role, revise or narrow it, or propose retirement. Give the reason, owner, and next review point. If the evidence is insufficient for that decision, write that instead and name the smallest missing input you need.
For a proposed change, say what later work should show if it helped. Any retirement needs a plan for necessary work the role would stop doing. Put accepted changes through the team’s normal process, with existing permissions and review rules intact.
The performance review is where you stop treating the role’s existence as evidence of its value.
Keep the work that helps. Change what the evidence says needs changing. Let a role stop when its job no longer earns the effort around it.
The role has to keep earning its place.
Original framework. Distilled from client work.
