March 19, 2026
Clippy’s back—with receipts
A survey on LLMs for spreadsheet intelligence
Spreadsheets get ChatGPT brains — half cheering, half panicking
TLDR: A new survey maps how text-based AI could let people talk to spreadsheets instead of writing formulas, while warning about trust and privacy. The comments split between joy at fewer formula headaches and fear of un-auditable errors—high stakes when spreadsheets power budgets, payrolls, and big decisions.
A new academic survey says the future of spreadsheets is talking, not typing: Large Language Models (LLMs)—the chatty AIs behind tools like ChatGPT—could turn “write a scary formula” into “just say what you want.” The paper maps a full workflow for AI-in-spreadsheets, catalogs tasks like writing formulas and cleaning messy tables, and flags the big headaches: trust, accuracy, and privacy. Translation: less VLOOKUP pain, more “can we believe this thing?”
The comments went nuclear. One side is popping confetti: “No more formula spaghetti! I’ll tell Excel what I want!” They see AI as a friendly sidekick that can explain steps, fix typos, and clean data in seconds. The other side? Red alert. Power users and finance folks warn that AI “hallucinates” (makes stuff up), doesn’t show its work, and can’t be audited. “If it’s wrong, who’s on the hook?” Privacy hawks shouted at anyone pasting payroll into a chatbot. Engineers begged for version control, explainable steps, and locked critical cells.
Memes flew: “VLOOKUP is dead; long live vibe-lookup,” “Clippy just got a PhD,” and “SUM-thing feels off.” A fragile truce emerged: let AI draft, humans approve. Even the survey leans cautious, calling for trustworthy systems that show formulas, cite steps, and keep data safe. Democracy vs gatekeeping? The spreadsheet wars just went prime time.
Key Points
- •The survey reviews LLM applications for spreadsheet tasks, covering methodologies, capabilities, benchmarks, and challenges.
- •It defines spreadsheet intelligence as a workflow of independent stages.
- •Existing research is categorized according to these stages and organized into a taxonomy of tasks.
- •Downstream tasks and their corresponding benchmarks are listed to form an end-to-end evaluation pipeline.
- •Open challenges and future research directions are discussed to guide trustworthy LLM systems in spreadsheets.