Vision AI for Operations: Receipts, Stock Shelves and Site Photos

The camera is now a data-entry device

Multimodal AI models, the kind behind ChatGPT, Claude and Gemini, can look at a photograph and return structured information: the supplier, date, VAT amount and line items from a receipt; the number of facings and gaps on a stock shelf; the state of a boiler installation from a site photo. For a small business, that turns the phone camera every employee already carries into a data-entry device.

This matters because a large slice of operational admin in a small firm is exactly this kind of work: reading something visual and typing it into a system. Receipts into accounting software. Shelf checks into a spreadsheet. Site conditions into a job sheet. Vision AI does not remove the work entirely, but it shifts your team from typing to checking, which is faster and considerably less painful.

Receipts and expenses: the easiest win

Receipt capture is the most mature use of vision AI, and you almost certainly should not build it yourself. Dext, Hubdoc (bundled with Xero subscriptions) and the built-in receipt snap in QuickBooks all extract supplier, date, net, VAT and category from a photo, then push the transaction into your ledger for approval.

HMRC accepts digital copies of receipts for record-keeping under Making Tax Digital, so once a legible image is stored in your accounting system you generally do not need the paper. That alone kills the shoebox problem.

Where receipt capture slips up

• Faded thermal paper and crumpled receipts produce missing or misread digits
• Handwritten amounts, such as tips added to card receipts, are often skipped
• VAT is not always itemised on small receipts, and the software may guess the treatment
• Foreign-currency receipts need a sanity check on the exchange rate applied

The fix is a simple rule: staff photograph receipts on the day they get them, and a named person reviews anything above a set value before it posts to the ledger.

Stock shelves: audits from a photograph

If you run retail, wholesale or field merchandising, shelf photos can now answer questions that used to need a clipboard: which products are out of stock, how many facings each line has, whether price labels match the current promotion, and whether a competitor has quietly taken your space.

Enterprise retailers use dedicated image-recognition platforms such as Trax for this. A small chain does not need that budget. A lighter-weight version is entirely workable: a staff member photographs each bay on a weekly walk, the images go to a multimodal model with a prompt describing what the shelf should contain, and the output lands in a spreadsheet flagging gaps and anomalies.

• Products hidden behind others get missed, so counts are estimates, not stock figures
• Similar packaging within a brand family gets confused, especially at a distance
• Poor lighting and glare on fridge doors degrade accuracy sharply

Treat photo audits as an early-warning system that tells you where to look, not as a replacement for proper stock counts.

Site photos for trades and field teams

Builders, electricians, plumbers and landscapers already photograph everything for their own protection. Vision AI makes that archive useful. A survey photo set can be turned into a pre-filled quote checklist: the model is asked to identify the consumer unit type, note visible pipework, estimate room dimensions from reference objects, and list anything that looks non-standard for the quoter to price properly.

Progress photos can be summarised into plain-English client updates, and before-and-after pairs can be checked for completeness against the job spec. Apps such as CompanyCam handle the organising and timestamping; a multimodal model handles the describing.

One hard boundary: never use AI judgement for compliance-critical calls. Whether an installation is safe or meets Building Regulations is a decision for a qualified person on site. AI can flag something worth a second look; it cannot sign anything off.

Accuracy caveats that apply to all of it

Vision models fail in a particular way: confidently. A model that misreads £83.40 as £88.40 will not tell you it was unsure. Build your process around that behaviour rather than hoping it improves.

• Add arithmetic checks: line items should sum to the total, and net plus VAT should equal gross
• Set value thresholds: anything above, say, £100 gets human eyes before it posts
• Sample-audit a percentage of low-value items each month to catch drift
• Keep the original image attached to every extracted record so checking takes seconds
• Watch for faces and number plates in photos; under UK GDPR these are personal data, so store and share site images with the same care as any other

Build or buy: a short decision guide

The choice comes down to how standard your job is. If thousands of businesses have the same problem, a product already exists and its vendor has spent years on failure cases you have not met yet.

• Receipts and invoices: buy. Dext, Hubdoc or your accounting platform's built-in capture
• Documents at volume (delivery notes, application forms): buy the plumbing. AWS Textract, Google Document AI and Azure AI Document Intelligence are built for exactly this
• Unusual workflows (your shelf layouts, your survey checklist): build a thin layer. Multimodal APIs from OpenAI, Anthropic and Google cost pennies per image, and a no-code platform such as Make, Zapier or Power Automate can wire a photo upload to a model and a spreadsheet

Build only where the process is genuinely yours, and start with one workflow rather than an everything-at-once project. If you want help scoping which is which, our team does exactly that.