Document & Image Intelligence

Three capabilities read content out of documents and images so it can be turned into data you can use. All three run on Amazon Web Services using your installation's existing credentials.

The three capabilities

• Document reading — pulling fields and line items out of invoices, forms and scanned documents.
• Text analysis — detecting what language a piece of text is written in.
• Image analysis — detecting what is in a photograph.

How they are used

These are services ActiveManage and your application call rather than screens you use directly. Where they appear in your system depends on how it has been built — typically behind a document or image upload.

These or the AI Gateway?

There is genuine overlap, since a modern language model can also read a document. As a rule of thumb:

• Document reading for structured extraction at volume. It is consistent and returns properly typed fields and tables.
• The AI Gateway for documents whose shape you cannot predict, or where you want an interpretation rather than a transcription.
• Both together is a common and effective combination — extract the data first, then reason over the result.

Document reading turns an uploaded document into fields and line items. It is built around commercial documents — invoices in particular — and does a good deal of work beyond simply recognising the characters.

What it handles for you

• Number formats. It recognises both UK style (50,000.00) and European style (50.000,00) and converts to one consistent form, working out which is in use from the document itself.
• Units. It strips noise such as "EACH" out of quantity fields and applies conversion factors where a unit such as pounds is detected.
• Field mapping. Your configured label rules are applied first, with the service's own field recognition filling any gaps.
• Awkward line items. It copes with columns all labelled the same, unit price against total price, and quantity columns that were never properly identified.

Every step is logged

Each adjustment is recorded as it is made. When an invoice extracts wrongly, that log is how you find out which rule did it — considerably faster than staring at the original document.

Practical advice

• Expect to tune it. Document extraction is never right first time across a real supplier base. Budget for a period of adjusting label rules as new layouts appear.
• Always review before posting. Put extracted financial data in front of a person before it becomes a transaction. Extraction is very good but not perfect, and the failure mode is a plausible wrong number rather than an obvious error.
• Handle "nothing found". The service can read a document and find no line items at all — a distinct outcome from an empty invoice.
• Watch mixed number formats. A document mixing conventions is most likely to mislead the format detection.

Worked Examples

• Purchase ledger: supplier invoices extracted into a draft, reviewed by accounts, then posted.
• Expenses: receipts photographed on a phone, with totals extracted and the image kept as evidence.
• Onboarding forms: scanned paper forms turned into records rather than being retyped.

Two smaller capabilities sit alongside document reading: working out what language a piece of text is in, and working out what is in a photograph.

Detecting the language of some text

Given a piece of text, ActiveManage can identify which language it is most likely written in, with a confidence score and the alternatives considered. Useful for routing correspondence to somebody who speaks the language, deciding which language to reply in, or tagging free-text records.

Use the confidence score. Short text is genuinely ambiguous — a three-word message may plausibly be several languages — and acting on a low-confidence result produces exactly the sort of error customers notice, such as a reply in the wrong language. Set a threshold and fall back to a default below it.

Only a limited amount of text can be analysed at once; longer text is trimmed automatically, so send a representative sample rather than an entire document.

Analysing images

Given an uploaded image, ActiveManage can detect what is in it — useful for tagging photographs so they can be searched, reading text out of a photograph such as a sign or a serial plate, or screening user-uploaded images before they are published.

Points to consider before switching image analysis on

• Images are sent to a third-party service. Where customers upload photographs, your privacy notice needs to say so.
• Anything face-related carries additional legal obligations in most jurisdictions, including the UK and EU. Take advice before using it, rather than after.
• Confidence scores matter here too. Act on high-confidence results and route the rest to a person.
• Cost scales with volume. Analysing every uploaded image is very different from analysing the ones somebody asks about.

Worked Examples

• Field service: read a serial number from an engineer's photograph instead of typing it.
• Marketplace moderation: screen listing photographs before publication, with anything uncertain queued for review.
• Multilingual support: route inbound e-mail to the right language team automatically.