Why plain-text CSV prompts can produce made-up numbers, and how Loam grounds your AI
A CSV gives your AI useful context. If it reads the file only as text, though, it may produce a plausible number instead of calculating from every row. This comparison shows that plain-text case beside a grounded Loam answer with a query you can check.
The workaround everyone reaches for first
When you have a spreadsheet and a question, a common first step is to paste the CSV into ChatGPT and ask. It is a useful workaround for small files and simple questions. We tested that plain-text workflow.
The dataset: a 500-row quarterly sales report with order IDs, dates, regions, products, quantities, unit prices, totals, and sales rep names. Real-world structure, nothing exotic. We pasted the full CSV into the context window and asked 10 straightforward questions, each in its own message in the same conversation. ChatGPT had all the data.
Then we gave Loam the same CSV and the same 10 questions. Loam calculated each result from the rows and showed the query. Same data and same questions, with two different ways of answering.
The results: 10 questions, side by side
Click any question to see both answers in full, including Loam's query and the rows it cites.
Raw ChatGPT
0/10
correct answers
Loam (grounded)
10/10
correct answers
Why this happens
ChatGPT generates text that looks right, which is exactly what you want when you're drafting or reasoning. But when you paste a CSV and ask "what was total revenue," plain context doesn't add up the "total" column. It produces a plausible-sounding number from the patterns it sees. Sometimes that's close. On a 500-row file with questions that span many rows, it often isn't.
The gaps can be subtle. "$1.24M" instead of "$1.04M" doesn't set off alarms. Neither does naming the wrong leading region. These are the kind of answers that land in a slide deck or a planning doc and get acted on before anyone re-checks them.
This is the AI grounding problem. Ungrounded, an answer is generated from patterns. Grounded, the same question runs against your actual data and the answer carries its source. The difference shows up the moment a question needs real computation.
What grounding changes
Each Loam answer in this comparison includes the query that produced it and the rows it drew from. You can verify any number by reading the query and seeing which rows contributed to the result.
None of this is a knock on ChatGPT. It's a strong tool, and the paste-a-CSV workaround is a reasonable first move. The point is narrower: when a decision depends on real numbers from your data, you want your AI to query the data, not read it out of text. Grounding is what closes that gap. You can also bring your own AI to it.
You can try this on your own files. Open your CSV in a sortable table to explore it first, then ask questions about your data after adding a CSV or TSV. The tool sends up to the first 200 lines after the header to an AI service and asks it to cite supporting lines.
Frequently asked questions
Was this useful?
Open your CSV in the browser
View and sort your CSV locally. To ask questions, open the separate chat tool and add a CSV or TSV. It sends up to the first 200 lines after the header to an AI service and asks it to cite supporting lines.