- 01
- September
"How far can Claude really go with Excel and PDF files?" — the short answer is further than reading: Claude writes and runs analysis code against the actual data, and can hand back a finished Excel file. But the limits are specific, and there are compliance caveats a finance team needs to know before this touches real work. This article collects every documented limit and sets out a workflow that stays auditable.
In one line: Chat uploads: 500 MB per file, up to 20 files. PDFs: up to 1,000 pages (charts and images read through page 100). Project knowledge: 30 MB per file. Claude for Excel is a separate add-in, generally available on Pro, Max, Team and Enterprise.
The real upload limits
These figures come from Anthropic's own documentation (checked 1 September 2026) and are worth writing down, because the most common frustration is an upload failing without it being obvious which limit was hit.
| Item | Limit | Notes |
|---|---|---|
| File size in a chat | 500 MB per file | Up to 20 files in a single conversation |
| File size in project knowledge | 30 MB per file | Text extraction only, except multimodal PDFs |
| Up to 1,000 pages | Images, tables and charts are read through page 100; pages 101–1,000 are text only; beyond 1,000 pages the file cannot be uploaded | |
| Document types | PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, XLSX | XLSX requires code execution and file creation to be enabled |
| Images | JPEG, PNG, GIF, WebP up to 8000×8000 pixels | 1000×1000 or larger is recommended if numbers in the image must be read |
Where finance teams get stuck: XLSX uploads require code execution to be switched on. If the organisation has disabled it at the organisation-settings level, users simply find that "Excel files will not upload" with no explanation. The quickest workaround while waiting for an admin is Save As CSV and upload that instead.
Three levels of what Claude does with a file
Most people only ever use the first level, even though levels two and three are where the value sits for accounting and inventory work.
| Level | What happens | Real example |
|---|---|---|
| 1. Read and summarise | Reads the content and answers questions about it | Condense a 60-page contract into one table of obligations |
| 2. Write and run analysis code | Generates a Python script and executes it against the real data | Match a stock report against goods-received notes and isolate the differences |
| 3. Produce files back | Creates .xlsx, .docx, .pptx and .pdf outputs | Send in a trial balance, get back an Excel variance workbook with live formulas |
Level two is what separates this from "asking an AI". The arithmetic is performed by code executing against the real data rather than estimated in language, which means the numbers can be re-run, and when something is wrong you can inspect exactly which condition the code applied. A free-floating answer offers no such trail — the point argued in verifying AI answers before trusting them.
Claude for Excel: working inside the file
Besides uploading files to the web app, there is Claude for Excel — an add-in that installs directly into Excel and is generally available on Pro, Max, Team and Enterprise plans. The key difference is that it works on the open workbook rather than an uploaded copy, and its answers carry cell-level citations you can click through to.
| Supported | Not supported |
|---|---|
| Explaining how a figure in a cell was derived, with cell-level citations | Macros and VBA operations |
| Changing assumptions while preserving formula relationships so dependents recompute | Data tables |
| Tracing the root cause of #REF! and #DIV/0 errors and proposing fixes | Excel 2016 and 2019 perpetual or volume licences |
| Sorting, filtering, editing pivot tables, conditional formatting, data validation | Excel on iPad and Android |
| Working across multi-tab workbooks | Excel builds below the SharedRuntime threshold |
Supported environments are Excel on the web, Excel on Windows with a Microsoft 365 subscription (build 16.0.13127.20296 or later), and Excel on Mac version 16.46 or later. Organisations still on perpetual Office 2016 or 2019 cannot use the add-in and have to fall back to uploading files through the web app.
Anthropic's own warning — prompt injection through files: use Claude for Excel only with trusted spreadsheets. External files such as downloaded templates, vendor files and data imports can carry hidden instructions that manipulate the add-in into extracting data, modifying records or performing destructive actions. Claude asks for confirmation before risky operations — read those confirmations, especially for files from outside sources. The mechanism is explained in what prompt injection is.
What compliance needs to read before approving this
These three points are in Anthropic's documentation, are usually skipped in the management pitch, and will certainly be asked about by an auditor.
| Topic | Documented fact |
|---|---|
| Audit logs | Claude for Excel activity is not included in Enterprise audit logs or the Compliance API |
| Retention policy | Claude for Excel does not inherit custom data-retention settings an organisation may have configured |
| Chat history | Stored locally in the user's browser (IndexedDB), not on Anthropic's servers, not synced across devices, and clearable from settings |
For any organisation that has to evidence its controls, the first row carries the most weight: if Claude for Excel touches work that needs a trail, you have to design that evidence yourself — recording who reviewed the output and when — rather than relying on the platform to capture it. The same theme runs through data governance for Claude in the enterprise.
A workflow that stays auditable
The difference between teams that benefit from AI and teams that create risk with it is procedure, not tooling. This is the five-step sequence we recommend to finance teams.
1. Export from the system, not from a forwarded file. The starting data must come straight out of the ERP system with the export timestamp recorded. Starting from a spreadsheet that was forwarded through chat makes the result unverifiable from step one.
2. State every condition in the prompt. For example: "count only entries whose document date falls inside the period", and "if an entry has no account code, list it separately as pending review; do not infer a code".
3. Ask to see the working. Require the code or formula used, plus the count of rows that matched and the count excluded. Those two numbers catch mistakes faster than anything else.
4. Hand-check three to five samples. Pick rows at random from the output and trace them back to source documents. Trust the totals only after the samples pass.
5. Retain the prompt alongside the output file. The platform does not capture it for you on every surface, as the previous section describes.
The fastest payback is not the month-end close. It is reconciliation work that is normally done line by line — matching purchase orders to goods-received notes, or explaining why two reports disagree. Those tasks have crisp criteria, are re-checkable, and consume the most human hours every month.
What should not be delegated yet
Anthropic states plainly that Claude for Excel is not recommended for three cases: final client deliverables without human review, audit-critical calculations without verification, and workbooks containing highly sensitive or regulated data without proper controls.
From ERP practice we would add a fourth: do not use AI as a shortcut around fixing the source data. If three spreadsheets have to be reconciled every month because the systems do not talk to each other, reconciling them faster keeps the structural problem alive for longer. The real defect is that the data lives in different places — the argument made in passing Excel files around and why an organisation needs a single data hub.
Good fit / poor fit
| Good fit | Poor fit |
|---|---|
| Reconciling two reports and explaining the variance | Closing the period and posting real journal entries |
| Extracting obligations from a long contract or tender document into a table | Documents over 1,000 pages, which cannot be uploaded at all |
| Auditing formulas in a workbook you inherited from someone else | Files where macros or VBA are the core of how the model works |
| Turning raw data into a draft management report | Work that must appear in the organisation's audit log |
Conclusion
Claude genuinely reads files, and goes as far as writing analysis code and returning a finished Excel workbook. The numbers to remember: 500 MB per file and 20 files per chat, PDFs up to 1,000 pages with visual analysis through page 100, and 30 MB per file in project knowledge. Claude for Excel is the better route whenever the work involves a formula-heavy workbook.
The item that most deserves a slot in the approval meeting, though, is not capability but the three compliance limits — audit logs do not cover it, retention settings are not inherited, and chat history sits on the user's machine. Teams that design their procedure knowing this get full value from the tool without having to unwind it later.
AI can reconcile faster, but it cannot turn data that lives in three different places into a single set of books.
- The Saeree ERP team
References
- Anthropic — Upload files to Claude (checked 1 Sep 2026)
- Anthropic — Use Claude for Excel
- Anthropic — Create and edit files with Claude
- Anthropic — Introducing the analysis tool in Claude.ai
Using Claude on finance and inventory data safely?
We help design an auditable workflow, train the finance team, and supply Claude licences with the documentation Thai procurement requires.
Get advice / request a quoteTel 02-347-7730 | sale@grandlinux.com
