Of everything in this series, this is the use case that most reliably makes people say "hang on, it can do that?" Because modern Claude models are multimodal, vision built in, per the What Is AI article, the camera on your phone is now a data entry device. Point it at paper, ask for structure, receive a spreadsheet. The pattern is one sentence long and it dissolves a whole category of admin.
The five everyday versions
Receipts and invoices to line items. Photograph the supplier invoice or the shoebox of receipts and ask for a table: date, supplier, description, amount, GST. Ask for it as CSV and it drops straight into your spreadsheet or accounting import. The end of month shoebox ritual becomes a ten minute job.
Whiteboard to typed plan. The planning session ends, everyone photographs the whiteboard, and the whiteboard's fate is usually to be erased into oblivion. Instead: photo in, and ask for the content typed up, organised, with the actions pulled out. Pairs perfectly with the meetings article, the wall becomes the minutes.
Business cards and details to contacts. The stack of cards from the expo, photographed in batches, becomes a clean contact table, names, companies, roles, numbers, emails, ready for the CRM. Add "draft a short follow up note to each referencing where we met" and the expo actually converts.
Site photos to documented findings. Walk the site, photograph what you see, then ask for a structured condition report: what is shown, apparent issues, locations, suggested follow ups for a human to verify. The judgement about what matters remains the professional's, what disappears is the evening of typing up what the photos already show.
Handwriting to records. The paper forms customers still fill in, the technician's job sheet, grandma's recipe cards, photographed and transcribed into clean text or a table. Handwriting quality varies and so does accuracy, which brings us to the fine print.
The fine print, because numbers
Vision extraction is genuinely strong and improving, but it can misread, a 7 for a 1 on a crumpled receipt, a smudged handwritten quantity, and it will occasionally do so with the same confidence as everything else. So the trust and verify rules apply with teeth:
- Anything financial gets its totals checked, ask for the column sum and verify it matches the source, mismatches catch line errors instantly
- Ask the model to flag low confidence reads, "mark any value you're unsure about", it is good at knowing when the ink was against it
- Human review scales with stakes: skim the whiteboard transcription, actually check the invoice batch headed for the BAS
The workflow is still ten times faster with checking included. The checking is just non negotiable.
Try it in the next ten minutes
No setup required, this one is pure habit. Open Claude on your phone, photograph the nearest receipt, and type: "Extract this into a table: date, supplier, items, amounts, GST, total. Flag anything you're unsure about, then confirm the line items sum to the total." That is the entire skill. From tomorrow, the moment paper appears in your day, some part of your brain will correctly whisper: that is not typing, that is a photo.
A crumpled receipt photographed, and the structured table it became
That wraps the core use cases. Next, into the build in public series, where I point all of this capability at real projects and show you what happens.