When Work Looks Flat, Ask AI for the Missing Axis── Ask for a way of seeing rather than an answer, and you get candidate columns to add to your records, one of which is worth trying

01The operator asked for a way of seeing, not a fix

One request in the September 2026 record has a different shape from the rest. At the time the operator was running three systems side by side: a public site, a store for daily records, and a database holding both numbers and written notes.

All three worked. Even so, the operator felt they were arranged on a single plane. What went to the AI was not a bug to fix or a feature to add. It was a request for reasoning, ideas and angles that would widen the whole thing. A request for a way of seeing.

This is a different shape of request from the earlier episodes. It is not an idea passed on as a question, nor a name to be defined. It leaves the systems alone and asks for more ways to look at them. What is worth checking is what such a request brings back, and what to do with it.

02Asking for a viewpoint returns candidate axes, not steps

The same situation yields different answers depending on what you ask for. Ask for a fix and the AI returns steps: which file to change, which feature to add. Those steps are built inside the way of seeing you already use.

Ask for a way of seeing and what comes back is not steps. It is candidate axes, ways of slicing your work that your records do not currently hold. Most of them are unusable. Seven in ten cannot be filled from the data you have. But two or three can sometimes be pulled out of records you already keep.

03With too few axes, effort grows but understanding does not

An operation stuck on one plane has a recognisable symptom. Records accumulate. Updates continue. And yet the questions you can answer are the same ones you could answer last year.

The reason is plain. What you can ask depends on how many ways your records can be sliced. Records that carry only a date and an owner produce only counts by date and counts by owner. Multiply the rows tenfold and the kinds of question stay the same.

Adding hands at this point hides the symptom. More summary tables, more screens, more reports. What grew was work, not material for a decision. When an operation feels flat, what is needed is not a faster tool but one more way of slicing.

04Adding an axis means adding a column that widens filtering and grouping

"Widening" is a vague word, so this episode fixes its meaning. Adding an axis means adding a column to your records so that filtering and grouping gain a new option.

There is a basis for defining it that way. A long-established approach to shaping data for analysis splits tables in two: tables that store events, and tables that describe the things those events involve. The second kind covers products, people, places and dates, and it is what enables filtering and grouping. The first stores the events themselves and their numeric measures. How many ways you can cut the data is set by how many of the second kind you hold.

The same approach has a case where one table plays several parts. A single date table is enough; whether you cut by order date, ship date or delivery date depends on which event column it is joined to. So a new view sometimes comes from a different join rather than a new column.

The boundary matters too. Adding tools is not adding an axis. Adding kinds of summary is not adding an axis. You have added one only when a question you could not answer before now has an answer.

05In material review, the record of review comments usually stops at two or three axes

Pharmaceutical material review shows the same symptom. Records of review comments can usually be sliced by case, date, reviewer and comment category. Counts per month and counts per category come out. Then it stops.

Candidate columns are often sitting in the records already. Did the comment rest on the source document or on an internal rule? Did it appear on the first pass or on a later round? How many exchanges did the correction take? Was the material sent back over a matter of fact or over a choice of wording?

Turn one of these into a column and the questions change. Instead of counts per category, you can ask at which stage noticing something would cut the number of rounds. The point is not to add every candidate. A column that a person must type in each time will be full of blanks within three months. Add one column that can be derived from records you already keep.

06AI lists axes well, but its judgement of feasibility is unreliable

Language models can produce candidate axes because they have read a great many ways of slicing records from other fields: logistics, clinical studies, accounting. Listing ways of cutting that sit outside your own work is faster for a model than for a person.

The weakness lies in judging whether a candidate can be built. In one study, more than a hundred researchers compared research ideas written by a language model with ideas written by researchers, without knowing which was which. The model's ideas scored significantly higher on novelty. On feasibility they scored slightly lower. Producing a new angle and having it work are separate things.

The second weakness is agreement. A 2023 paper by Anthropic researchers reports that AI assistants tend to give answers matching the user's view, and that human raters' preference for answers matching their own views is one likely cause. Ask "what about this axis?" and the model will readily list reasons in favour.

So treat what comes back as proposals. Which record supplies the column, who carries the typing if nobody does, and what becomes answerable once it exists: those three are checked by the person.

07From tomorrow: list your axes, ask for ten candidates, add one

There are five steps.

1. List your current axes. Write down the columns your records can genuinely be sliced by. Usually there are three or four.
2. Write three questions you cannot answer. The ones you dread being asked, or that yield the same answer every year.
3. Ask for ten candidates. Hand over the current axes and the three questions, and ask for ten ways of slicing you lack. Do not ask for answers or steps.
4. Ask where each would come from. Make the model name the existing record each candidate could be derived from. Drop the ones with no source.
5. Add one and live with it for two months. Pick a single survivor and add the column. If no new question becomes answerable, remove it.

The fifth step is the one that decides the outcome. When three candidates survive, the urge is to add all three. Add three and you cannot tell which one helped, and all that remains is the typing. Added one at a time, a column that did not help can be identified and removed.

Guidance on working with AI points the same way. Anthropic's documentation for Claude Code recommends that, while the approach is still uncertain, you have the model explore and plan before it writes any code. The same document describes having the model interview you and write a specification before a larger feature begins. Both settle the way of seeing before the building starts.