What is LLM spreadsheet processing?

LLM spreadsheet processing means applying a language-model prompt to individual rows in a CSV, Excel file, or CRM export and saving the answers in output columns.

Use it for text that needs interpretation: classifying a product, extracting details from a description, summarizing a review, or translating a catalog. A row-based prompt reads the columns you reference. It does not automatically inspect the whole spreadsheet or discover patterns across every record.

Choose the right processing method

For a fixed transformation such as removing a known prefix, use a formula or an explicit rule. For a task such as deciding whether a description refers to software or a service, an LLM can interpret the text.

Datablist's AI Editing follows a different workflow: it generates JavaScript from your prompt and a sample of items, then executes the code across selected rows. It is useful for repeatable transformations, but the generated code needs review and representative exceptions in the instructions. It is not an independent LLM judgment on every row.

If the answer requires current web evidence, use a research workflow such as AI Agent. A prompt asking a general language model to “research this company” does not by itself establish that it searched the web.

A practical spreadsheet prompt

Suppose your file contains Product Name and Description. You want a Product Type column containing Software, Physical product, Service, or Unclear.

Insert the matching properties using the prompt editor. The literal separators keep the input sections visible when a description contains several lines. They help organize the prompt; they do not guarantee that a model will ignore instructions embedded in the source text.

An illustrative result is Software for “A subscription application for managing invoices”, and Unclear for an empty description. Keep those outputs alongside the original values and a stable source ID. These examples define the intended rule, not a measured accuracy result.

Run the workflow in Datablist

Import the CSV or Excel file, choose the rows you need, and open Ask ChatGPT/OpenAI from the enrichment catalog. Configure the prompt with the relevant properties and map the answer to a new output property.

For several output fields, align the names and types in the structured-output settings with your prompt. A configured field type helps keep the result usable, but it does not verify the factual answer. See structured LLM output.

Choose Continue with Instant Run and test a representative sample. Include empty text, ambiguous records, multiline descriptions, and any cases where the answer should remain Unclear. Review the output before processing the remaining intended rows.

Check a bulk run before export

Keep source IDs and original text unchanged. Check the run status separately from whether an answer is correct: a completed request can still return an unsuitable category. Retry failed requests deliberately and inspect existing-data settings before replacing accepted results.

Estimate the credit budget from the selected enrichment and sample usage. Do not treat the row count alone as a fixed price; text length, model, and processing choices can affect usage.

For the full workflow, read run ChatGPT prompts on Excel and CSV rows. Use LLM data cleaning for classification and cleanup decisions, and batch LLM processing when planning larger runs.