The shapes of
an AI tool
A capability and the shape it is delivered in are two different things. The capability is intentional on what is needed. The shape is how it is used.
Scroll, click, or use the arrow keys to move through.Scroll, or use the arrows and dots to move through.
The same question, asked twice, does not give the same answer.
A model produces the most probable continuation. The most probable continuation is not the only one available. Run the same prompt on the same input tomorrow and the wording moves. Sometimes the substance moves with it.
This is not a fault to be tuned out. It is how the machine works.
Nobody can say a summary came out wrong because it came out different.
Summarize what changed in the pipeline this month. Two runs give two summaries. One leads with the lost deal, the other with the two that moved to proposal. Both are accurate.
Drafting, summarizing, comparing, explaining, taking a first pass at something a person will read and adjust. Variation inside that range costs nothing.
Nothing says which one is right.
An answer that has to be exact cannot come from a step that varies.
One of those is wrong. Possibly both. Nothing in either answer indicates which. The person reading the report has no way to tell.
A tool moves the exact parts off the model and into code.
The tool does not replace the model. It takes the arithmetic away from it.
Code queries the CRM and sums the column, so the total is the same every run and it is the same total the CRM would give. The model still writes the commentary, because commentary is the part where a range of correct answers is acceptable.
The exact parts become code steps. The judgment stays a model step.
Work that pulls data in, saves it and analyzes it is code work.
Reaching into a system for records, holding them somewhere, and running the same analysis over them every month are not things a prompt does. They are things software does.
A person can paste one export into a window. They cannot paste a query, they cannot save the result anywhere that lasts, and they cannot reliably do it the same way in March as they did in January.
Repetition does not make something a tool.
Summarize the overnight support tickets and send them to me each morning. That runs every day, forever, on a schedule. It is still a prompt. The output is qualitative and a different summary tomorrow is still a good summary.
Doing something forty times does not make it need software. Needing the same answer twice does.
The narrative section
What changed, what stalled, what is worth attention. A different phrasing next month is not a defect. This is a prompt.
The numbers section
The closed-won total, the stage counts, the average days in each stage. Those have to reconcile with the CRM, and again next month against the same rules. This is a tool.
Most real work is both. The question is which part you are looking at.
A prompt varies by design. Build a tool where that is not acceptable.
Qualitative output is fine from a prompt, on a schedule, forever. Quantitative output is a tool, and so is anything that pulls data in and saves it. Those are the parts that have to be exact.