Login / Register ID | EN
This page has no official English version. It was translated automatically and may contain errors. Read the original in Indonesian →
Deck dan Dashboard Buatan AI: Rapi di Layar, Belum Tentu Benar di Angkanya
Foto: Pexels
Office AI

AI-Created Decks and Dashboards: Neat on Screen, Not Necessarily Accurate in Numbers

Tim Taalenta · 29 Agustus 2026

Throughout 2026, the functionality of AI tools shifted. Previously, AI output was in the form of text that could be quickly reviewed. Now, the output is in the form of finished documents. Kompas Tekno reported the launch of ChatGPT Work in July 2026, which can create documents, spreadsheets, and presentations from materials provided by users. Anthropic took a similar path with Cowork, an agent in a desktop application that reads files in the work folder and returns the results.

This change is convenient, but it moves one risk to a less visible place. Incorrect paragraphs can still be caught when read. Incorrect graphs look exactly the same as correct graphs.

What a Neat Presentation Hides

A bar chart does not indicate how many rows of data failed to be read. Summary slides do not specify which periods are used when the source contains three years at once. Growth figures do not clarify whether empty rows are counted as zero or skipped. All these decisions are made in the middle of the process, then disappear behind the final output that looks presentation-ready.

A frequently cited case comes from Australia. Tirto reported that a report prepared by Deloitte for the Australian Department of Employment contained academic references that never existed and fabricated court ruling citations. The issue only emerged after a researcher from the University of Sydney examined its bibliography. The company then published a revised version and returned part of the contract funds.

What failed there was not the model's capability. What failed was the lack of checks before the document was used for official purposes.

Three Inexpensive Checks

Rechecking the entire work would eliminate the benefits of automation. What makes more sense is targeted checks at the points that most often go awry.

First, reconcile the numbers. Compare the number of rows in the source data with the number of rows that actually made it into the graph. An unexplained discrepancy almost always means there is silent filtering, usually due to inconsistent date or number formats.

Second, manually test one point. Take one bar or one cell on the dashboard, then recalculate from the raw data. If one point matches, the calculation method is likely correct. If it does not match, the entire visual needs to be reviewed.

Third, request the assumptions to be documented separately. The periods used, filters applied, treatment of empty cells, and exchange rates if there are currency values. Unwritten assumptions cannot be debated in the meeting room.

Responsibility Remains with Humans

The Circular Letter of the Minister of Communication and Informatics Number 9 of 2023 on AI Ethics places credibility and accountability as values that must be upheld, with oversight involving both providers and users. In everyday practice, that principle means simply: the name on the deck is the name that will be questioned when the numbers are wrong, not the name of the model.

There is one rarely discussed side effect. When preparing materials becomes quick, the biggest temptation is not to create more decks, but to stop reviewing because the results already look good. In fact, the workload does not disappear; it just shifts from preparation to checking. The skill that is increasingly valuable in today’s workplace is the ability to read a graph and ask where that number came from.

Sources

Read also:

Ditulis oleh Tim Taalenta