You gave the bargain a smaller job.
Imagine you are comparing two AI options for a piece of internal work. Both are already available to your team. You expect one to cost less, and you would like to use it for the routine work that does not need your most expensive option.
The job is a short how-to sheet, made from an invented process description. A colleague should be able to see what must be ready before starting, follow the actions in order, and know what to do if a required input is missing.
One answer is a tidy summary. The starting conditions are missing, and the colleague would need to work out how to handle the missing input. You look at it and say, “That’s basically there. We can add those bits.”
The other answer gets checked against the whole job.
Before you have compared a single cost, you have given the first option a different finish line.
Cheap compared with what?
There is nothing wrong with looking for a less expensive way to get useful work done. The problem is declaring the answer before the work is finished.
In the public introduction to his cost guide, Nate B Jones points to the cost of an accepted result, including the retries and human cleanup needed to get there. That is the principle worth bringing into the operator’s decision.
Here is how I would use it on this small comparison.
Keep the job the same. Keep the meaning of acceptable the same. Then look at what each option needs to reach it.
Here’s the thing. Your preference for the bargain can change the standard without anybody admitting that the standard changed.
“It’s only an internal document.”
It was an internal document when you wrote the job, too. The person using it still needs to know where to begin.
“They’ll know what we mean.”
Then you are including that person’s knowledge and repair work in the result. Fine. Count that arrangement honestly.
The option may still be the better choice. You have not established that by quietly accepting less from it.
Compare the price of finished work.
Let the checker use the same standard
In our imagined example, the useful result was agreed before either option answered. There is no need to invent a new review system. Bring the person who normally checks this kind of sheet and keep the existing requirements in view.
Can a colleague start from it? Are the actions in a usable order? Does it preserve the missing-input instruction from the supplied process?
Those questions apply to both answers.
A summary that needs somebody to supply the missing parts is still unfinished against that job. A longer answer that includes everything but tangles the order may also be unfinished. Length, polish, and familiarity with the tool do not settle it.
If practical, let the checker read the outputs without the option names attached. The point is to make it harder for “this is the cheaper one” or “this is our best model” to become part of the quality standard.
The Professional Recipe calls its evidence-and-revision rhythm Feedback. Here, the evidence has one job: help you decide which option should handle this kind of work.
You already know what the sheet needs to do. Feedback connects what you observe to the choice you make next. It does not ask you to collect numbers for their own sake.
The help belongs inside the comparison
Suppose you decide to repair the summary yourself.
You add the starting conditions. You put the missing-input instruction back. You read it again. Now it meets the standard.
That is a legitimate way to get the job finished. It is also a different amount of human involvement from accepting the first answer without changes.
“I only spent a minute on it.”
Maybe. Record the time you actually spend. Include checking, deciding what needs repair, and checking the repaired version. The visible edit can be the smallest part of that work.
You need to measure the tool’s cost. Let me tighten that. You need to understand the cost of the arrangement that gets this job finished, with the tool and the person in it.
The same applies if you ask the AI to try again. Another attempt belongs to that job. So does the human work of explaining the miss. An option does not get a clean first-attempt result because the final answer looks clean after several tries.
Agree the retry limit before starting. When an option reaches that limit without acceptable work, leave the job marked unfinished. Do not remove it from the comparison because there is no nice finished sheet to display.
Right? The unsuccessful request is part of what you learned about asking that option to do this work.
Money and minutes tell you different things
Now you can look at cost without pretending every kind of cost is the same number.
If you can identify a charge for the attempts, record it. If you cannot, leave that amount unknown. Having access through a subscription does not make the work free, and the full subscription bill is not automatically the charge for this one request.
Keep active human time visible beside any attributable charges. The person checking and repairing the sheet has other work to do. That matters even when you have not assigned a dollar value to their minutes.
Waiting is another distinction. A response can take longer to arrive while requiring little attention. Another can appear quickly and keep someone occupied repairing it. Write down what you observed without calling both things time saved.
Keep one-time setup separate, too. Preparing an option for the comparison can be worthwhile, but a few requests do not tell you how often that setup will be reused.
You may later put a value on staff time or spread setup cost across an expected workload. Those are assumptions to state when making that decision. They are not money this small comparison has already proved you saved.
Does that make sense? You can know which arrangement asked more of the checker while still not knowing which had the lower monetary cost.
That is useful information. It is also a limit on the conclusion.
Leave room for the bargain to win
Here’s the thing. This is not an argument for always buying the strongest option.
If the less expensive option can do the agreed job with acceptable effort, that is a reason to consider using it there. If another option needs less human help, that difference belongs in the decision. If the evidence is mixed or the money cannot be attributed, you may need another observation before choosing.
The exercise has to allow all of those outcomes.
Otherwise you have built a way to defend the answer you wanted. You can do that with a beautifully organized comparison sheet just as easily as with a feeling.
For this job, you are deciding who or what should produce a usable how-to sheet. You are not declaring a winner for every kind of work in the business. A tool that suits short instructions may tell you very little about a different task.
Yeah. A limited answer can still be an actionable answer.
“Use this option for this job, with this amount of checking.”
Or: “We do not have the cost information needed to call it cheaper yet.”
Both are more useful than a winner whose result was allowed to mean something different.
The Monday Move: compare one finished job
Pick one actual choice between two AI options you already have permission to use. Choose a small internal job and three ordinary requests from that same kind of work, using public or invented material.
Give both options the same source material, the same job, and the same acceptance standard. Keep their working conversations separate. Agree the retry limit with the person who will check the outputs.
For each request, keep a short note of whether it reached acceptable work, the attempts it took, the active human help, and any charges you can actually identify. Preserve unfinished work and the reason it remained unfinished. Keep setup and waiting separate.
Then make the narrow recommendation those observations support. If a needed number is missing, name it. Three requests can show you what to investigate or try next; they do not establish a permanent model ranking or a savings claim.
Return to the line you were tempted to say at the start: “This one is cheaper.”
Make sure both sides of that sentence contain the same finished job.
Source translated from Nate B Jones. Operator framing by AI in Crayon.
