AI Excel tools write formulas well but stumble on complex tasks, sspai review finds

The review tested OpenAI, Anthropic and Google Gemini models alongside DeepSeek, Kimi-K3, GLM-5.2 and other systems. Some ran through native Office add-ins such as ChatGPT for Excel and Claude for Excel; others used the open-source Pi for Excel sidebar add-in. Standalone workbenches including Codex, Claude Cowork and Workbuddy submitted files and prompts end to end. Reasoning was set to high or deep thinking where supported. Tests covered data cleaning, repetitive formula writing and final delivery, with cases labeled A01 through A11 and Q11.

The review’s summary said AI formula writing is now mature, but delivery after complex workflows remains a problem. It also said the harness is as important as the model, that Office sidebar add-ins were the best harness in the tests, that users should beware black-box decisions made by AI, that prompts must be sufficiently rigorous, and that after receiving an AI-processed file users should modify the original data once to check whether calculated cells update automatically.

In data cleaning, the review found that models can identify fuzzy data where different expressions carry the same meaning, and can use regular expressions or Python to process text in batches. But behavior diverged. In test A09, which matched employee information by name, there were duplicate names and two names that could not be found. Claude Cowork stopped and asked how duplicate employees should be handled, offering options to add a note, match only the first, or write both. ChatGPT for Excel, Claude for Excel and DeepSeek in Pi for Excel mentioned the duplicates but used XLOOKUP’s default behavior and matched the first value. Kimi-K3 in Pi for Excel added a cell comment. Claude and Gemini in Pi for Excel did not mention the duplicates and used XLOOKUP to match the first value. Workbuddy went further and fabricated the two closest-looking names for the missing entries.

The review said most products made decisions on their own when prompts did not specify cleaning or completion rules for conflicts, duplicates or missing matches. It argued that asking the user becomes more important as tasks grow more complex, and that the best current remedy is to tell the AI to ask when there is ambiguity. Even then, not all models followed the instruction. Few models completed data cleaning with zero intervention. Common failures included failing to unify synonyms or heterogeneous fields and failing to understand text-formatted numbers. The review attributed some of the latter problem to CSV conversion and pandas’ automatic type inference, which can silently convert text-formatted numbers while Excel still treats the original cells as text and excludes them from calculations. That creates mismatches between Python results and Excel’s native summaries.

Pivot tables were harder. In tests A10 and A11, which explicitly required creating and manipulating pivot tables, 16 of 22 output files, or 73%, created pivot tables. Fourteen referenced the correct data source range, but only six, or 27%, produced correct calculations. Sorting pivot tables also defeated several models. Kimi K3 in Pi for Excel repeatedly hit the harness’s maximum limit while trying to sort a pivot table and ultimately failed to sort by sales from high to low.

Deep thinking helped reasoning and reduced hallucinations, but the review found longer reasoning did not always produce better results. In 45 tests on long, complex tasks including A08, A10, A11 and Q11, 51% required human intervention. For short tasks A01 through A07, the intervention rate was 12%. Models sometimes overthought small issues. Kimi K3 in Pi for Excel spent more than 25 minutes on A10 trying to use the Office.js API and gave up on sorting the pivot table. After a human pointed out a classification error, DeepSeek v4 Flash took 10 minutes to recognize a data type mismatch. Claude Cowork was stuck for 10 minutes on A08 because of a leftover LibreOffice file. Workbuddy with the Hy3 model ran for 40 minutes on A10 without solving the problem or producing further output.

Harness limits also mattered. Pi for Excel caps a single reasoning run at 4,096 tokens, so models that thought too long were cut off and could continue only after user prompts. The review also observed models wasting tokens by copying data into their reasoning. Kimi K3, working on A11, flip-flopped six or seven times over whether a missing payment method should be filled as “unknown” or “unknown payment method,” with reasoning logs showing phrases such as “I keep flip-flopping” and “Ugh.” When told to continue, it restarted from the beginning rather than resuming, and did not finish the task. The review said Kimi and GLM-5.2 took significantly longer on average than GPT, Claude and Gemini and needed more rounds to reach a result.

The review concluded that stronger reasoning alone does not yield better results unless the model works well with its harness. It recommended rigorous prompts, explicit instructions for AI to ask about ambiguity, caution over black-box decisions, and post-delivery checks by changing source data to see whether formulas recalculate.