Can AI read my suppliers' messy price and stock files for me?

Why supplier files are hard to automate

Old-style import tools expect the same columns every time. Suppliers do not send that. One sends a PDF catalog with a table per page. One sends a spreadsheet with merged cells and a header on row six. One emails "price rise of 4 percent on all items from the 1st". A fixed template breaks the first time any of them changes. AI reads the page the way a person does, so it copes with new layouts.

What AI can do with them

• Read a PDF table. Anthropic's documentation says Claude accepts standard PDFs (not password protected), up to 600 pages per request, or 100 when the request uses a smaller context window, and 128 MB on its own API. Each page is processed as an image, with its text extracted alongside, so tables and charts are read, not just plain text. The same page says a page typically uses 1,500 to 3,000 text tokens depending on density, plus image tokens, with no extra fee for PDFs. The same page warns that dense PDFs with small fonts and complex tables can fill the context window before the page limit, and suggests splitting the document into sections. If you use the Claude chat app instead of the API, its help center lists different limits: PDFs up to 1,000 pages, with tables and images read only in PDFs of 100 pages or fewer.
• Handle spreadsheets. The same documentation says plain text files such as CSV can be sent directly, but Excel files must be converted to text or PDF first. Save as CSV before you upload. In the chat app, Claude's help center says Excel (XLSX) files can be uploaded if code execution and file creation are switched on in your account.
• Map columns. "Item No", "Part #" and "Ref" all become your SKU field.
• Match to your catalog. It can match supplier codes to your SKUs, including near matches, and flag the ones it is unsure about.
• Compare to last time. Give it last month's file and this month's. Ask for every row where the cost changed, with the percent.
• Read the email. A percent rise in a message body can be turned into a calculated price column.

My guide to searching 40 supplier PDFs for an answer covers the question-asking side, and AI document processing explains the wider idea.

What it gets wrong

• Case versus unit price. A column named "Price" may be per case of 12. Tell the AI the unit of measure, or have it report any row where the unit is unclear.
• Look-alike codes. Codes that differ by one character are easy to mix up. Check matches, especially where the match is "close" and not exact.
• Dropped rows. On a long file, a model can skip rows. Always compare the row count in and out.
• Currency and tax. A price that is ex-tax in one file and inc-tax in another needs a clear rule.
• Scanned pages. A blurry scan or a photo of a printed list is harder. Expect more errors and more checking.

A safe way to update prices from a supplier file

1. Save a backup of your current product export. If your store runs on Shopify (a hosted online store platform), note that its CSV import help says imports cannot be canceled once started and that a blank column overwrites the live value with a blank. A backup is not optional.
2. Give the AI three things: the supplier file, your current cost sheet and a short rule list (unit of measure, currency, tax, rounding).
3. Ask for a change table: SKU, old cost, new cost, percent change, and a confidence note. Ask for unmatched rows in a separate list.
4. Check the extremes. Sort by percent change. Anything over, say, 15 percent or any cost drop to zero gets a human look.
5. Approve and apply. Only changed columns go into the import file. Tick "Overwrite products with matching handles", or Shopify ignores rows for products that already exist. If you include a SKU column, include the option name and option value columns too, because Shopify says a SKU column without them deletes the variant option. Try the file on a test store first when you can. Apply costs first. Change selling prices through your own margin rule, not by letting AI guess.
6. Spot-check live. Open five products on each channel after the update.

A worked example (made-up numbers)

A seller of pet supplies gets a 40-page PDF from a supplier with 900 items. The AI returns 900 rows. Comparing to last month, 212 costs changed. Sorted by change, 6 rows show a rise over 30 percent. A person opens the PDF and finds 4 are case prices read as unit prices and 2 are real. Those 4 are fixed, and the other 206 changes go in. The job that took a morning is now done in under an hour, and the review caught what would have been 4 wrong prices. By Anthropic's token figures, a 40-page PDF uses roughly 60,000 to 120,000 text tokens (40 pages at 1,500 to 3,000 each), before the image tokens, which is why the same page suggests splitting dense files into sections. These figures are invented.

Stock feeds and automatic files

Some suppliers offer a stock feed, a file or link that updates on a schedule. If yours does, use it. A plain scheduled import is cheaper and more predictable than asking AI to read it each time. AI earns its place with the files that arrive as PDFs and emails, or change shape. Sellers who need stock counts kept in step across channels should also see multichannel inventory management software.

How I'd approach this

Start by hand. Take your messiest supplier and your most regular one, run both through Claude this week, and count how many rows it got wrong. If the answer is under two percent and you can see which ones, you have a usable process. If you do this every week for five or more suppliers, it is time to turn it into a repeatable flow: the files land in one place, the AI reads them, a change table is built, and you approve it. If you build it through the API, Anthropic's structured outputs feature can force the answer into a fixed set of fields, so every run returns the same columns. That fits what I describe in building an AI workflow. To pick the first job to try, the AI Strategy Sprint is a paid 3-day first step with me, with a free 15-minute AI brainstorm call to start.