The request here is straightforward, and the answer is yes: this can be done at scale, but not by asking ChatGPT to read 30,000 company names and make judgment calls. That is a misuse of the tool, and anyone who has tried it knows the output drifts into inconsistency around row ten thousand. The practical move is to stop treating this as a research problem and start treating it as a classification problem. You do not need perfect accuracy for a first-pass filter. You need a defensible, repeatable method that gets you to 80 or 90 percent correct, then you manually review the edge cases.
What works is a two-step approach. First, build a keyword-based classifier that scans each vendor name and flags obvious matches. Software, hardware, cloud, consulting, infrastructure, data, systems, analytics, security, and related terms go in the Yes bucket. Energy, hotel, law, logistics, real estate, and similar go in the No bucket. This alone will handle a meaningful share of your list with zero web research. Second, for the remaining vendors that do not match either bucket, export that subset and run a lightweight lookup against a public company description source, such as a business registry or an API that returns industry tags. You are not asking an AI to reason about each company. You are asking it to retrieve a structured field, which is a task it handles reliably.
The user who posted this is right to note that ChatGPT cannot handle that much data in one pass. But the deeper issue is that they are framing the problem as a conversation when it is actually a pipeline. A spreadsheet formula can handle the keyword pass. A simple script or even a mail-merge style lookup can handle the enrichment pass. The web research part is only necessary for a small fraction of vendors, and you can sample that fraction to validate the accuracy of your broader method. You are not building a perfect system. You are building a filter that reduces 30,000 uncertain rows into a few hundred that need human eyes.
The practical takeaway is this: do not try to make the tool do everything. Break the job into passes, use the cheapest and fastest method first, and only escalate to slower methods for the residual. Your goal is not to be right on every row. It is to get to a clean enough dataset that your sourcing analysis is not distorted by a column of blank or guessed values. That is achievable in an afternoon with a spreadsheet, a keyword list, and a free or low-cost company data source. The question is not whether it can be done. It is whether you will accept a method that is good enough to move forward.