When the Team Formats Excel and ChatGPT Sets the Priorities

I was in a plant recently where a team was working with an Excel file and ChatGPT.

The team adjusted column widths, added borders, aligned entries, revised headings, and applied color standards, while checking the file for consistency and continuing to make formatting changes. At the same time, ChatGPT reviewed their conversations and data so it could propose what the team should do next.

The division of labor assigned the mechanical task to the team and the judgment task to the machine.

Nobody had formally decided that spreadsheet formatting required human attention or that ChatGPT should determine the next actions, but that was how the work had been allocated.

Although the technology made this allocation possible, the behavior behind it has been present in improvement work for a long time.

Visible work attracts attention

Teams tend to favor work that provides evidence of completion because progress is easy to observe.

A spreadsheet can be cleaned up, a presentation can be revised, a board can be relabeled, and a tracker can be reorganized. These tasks have defined outputs, so progress can be observed while the work is being performed. A team can adjust column widths, colors, borders, labels, and alignment, then inspect the result, and the task is complete when the file meets the selected format.

Operational judgment has different characteristics: a team must identify the relevant facts, separate symptoms from causes, evaluate tradeoffs, and choose an action. Although the available information may be incomplete and causes may remain uncertain, the team must select among several reasonable actions, each of which has operational consequences.

Completion is established through a decision under uncertainty, and because the effect may remain unclear until the process responds, the team must decide when the analysis is sufficient and accept responsibility for proceeding.

This difference affects how a team allocates its time because a bounded task provides frequent confirmation that work is advancing, while an open-ended task produces additional questions. When both are available, the bounded task can consume the meeting because formatting provides a defined end state, while judgment continues to require discussion, investigation, and commitment.

ChatGPT adds another option because the team can assign the open-ended work to a system that will produce a complete response. When it returns proposed priorities, actions, or conclusions in a structured form, the judgment task gains an apparent endpoint.

Although a generated recommendation may provide relevant input, generating an answer and making a decision are separate activities. The response does not establish that the team examined the process, selected the correct priority, understood the operational constraints, or accepted responsibility for the result.

The risk begins when the existence of a complete response is treated as evidence that the judgment work is complete.

The same behavior appears in MDI

Anyone who has deployed managing for daily improvement has seen a related pattern.

A team can spend a morning working on the layout of a board: the boxes need consistent sizes, the headings need clearer wording, the colors need standard definitions, and the charts need uniform formatting. While the board changes throughout the session, each change can be reviewed as soon as it is made.

The process problem displayed on the board may receive much less attention.

The board exists to support a management process by making abnormal conditions visible, establishing accountability, and supporting timely action. Its layout matters only to the extent that it helps the team perform those functions.

Although this behavior often receives labels such as poor discipline, weak ownership, or lack of engagement, those labels provide limited help. Leaders contribute to the result when their reviews devote more time to presentation details than to the team’s analysis, decisions, and follow-up.

ChatGPT changes the available response because a team can submit notes, conversations, and data and receive a structured proposal. While the tool organizes information and presents a sequence of actions, the team can continue working with a formatted file and a proposed plan without completing its own evaluation.

The old version produced a polished board beside an unresolved problem. The current version can produce a polished spreadsheet and a machine-generated action plan that the team has not fully evaluated.

The division of labor requires a decision

The division of labor between a team and a machine should be established deliberately.

Mechanical work is usually the clearer assignment for the machine because formatting, transcription, consolidation, classification, calculation, and the application of defined rules are repeatable. Their outputs can be checked against explicit requirements, and errors can be identified without reopening the operating decision.

Judgment remains the team’s responsibility: the team determines which condition matters, which evidence is relevant, what action is appropriate, what risk is acceptable, and what work should stop. These decisions depend on process knowledge and operating context, while the team remains accountable for implementation.

ChatGPT can support that work when it summarizes discussions, identifies missing information, organizes possible causes, compares options, and drafts proposed actions. Because each output remains an input to the team’s decision process, the team must test the proposal against conditions in the operation.

That review requires more than approving a response that sounds reasonable. The team should be able to state the problem in operational terms, identify the evidence supporting the selected action, explain the relevant tradeoffs, name the person responsible, and define how the result will be checked.

If the team cannot do that, the judgment work remains incomplete.

This distinction also defines the role of leaders. A leader reviewing the work should examine how the decision was made. Questions about evidence, assumptions, constraints, expected results, and follow-up establish whether the team performed the required analysis, while reviewing the appearance of the file confirms only that it was formatted.

Accountability stays with the team

Although a generated recommendation can be detailed, organized, and plausible, none of those characteristics transfers accountability.

The team operating and managing the process will implement the action: they will allocate labor, interrupt production, adjust standards, accept cost, and respond to the result. They need enough understanding to make those commitments deliberately.

Responsibility remains with the team even when ChatGPT proposes the priority or drafts the action plan, because the team approves the action, assigns the work, monitors the process, and evaluates the result. If conditions change or the action creates an unexpected effect, the team must decide how to respond.

The practical control is straightforward: define the division of labor before beginning the work. Specify which tasks the system will perform, which decisions the team will make, what evidence the team will review, and how the final action will be approved and checked.

That definition gives the system a defined role in the work and leaves authority with the people responsible for the operation. While mechanical activities can be automated, priorities, actions, approval, and follow-up remain assigned to the team.