Generative AI for back-office work: invoices, quotes, Emirates ID and Arabic documents

Generative AI can now read supplier invoices, quotes, Emirates IDs, trade licences and Arabic letters, pull out the fields your systems need, check them against your rules and draft the reply, leaving a person to approve anything it is unsure about. It works best as one controlled step inside a workflow. This guide covers the document workflows that matter in UAE finance, operations and HR, how document AI works, which major services read Arabic, what they cost per page, how e-invoicing changes invoice capture and what the PDPL expects for ID documents.

Sources: KPMG Lower Gulf, Is your finance function AI ready? (26 August 2025) · Ministry of Finance, UAE Electronic Invoice mandatory fields v1.0 (23 February 2026) · Azure Document Intelligence pricing. Checked 14 October 2026.

Generative AI in the back office: the document workflows in the UAE

Back-office paperwork in the UAE has a recognisable shape: a handful of document types, in two languages, arriving by email, WhatsApp and scanner. KPMG found that 54 per cent of UAE finance leaders plan to prioritise generative AI over traditional AI in the coming year.

Source: KPMG Lower Gulf press release (26 August 2025)

Supplier invoices, purchase orders and quotes

Payables teams retype supplier, TRN, number, dates, net, VAT and total, then match each invoice to an order. Extraction removes the typing; rules catch the errors that cost money, such as VAT that is not 5 per cent of the standard-rated net, a duplicated invoice number or a TRN that does not match the supplier record. The UAE introduced VAT on 1 January 2018 at that standard rate. Incoming quotes become comparable line items, and generative AI can draft your own bilingual quotes from the approved price list.

Source: UAE Ministry of Finance, VAT

Emirates ID, passports, visas and trade licences

Onboarding an employee, tenant or corporate customer means keying the same identity and licence details into several systems. The Emirates ID number has 15 digits and stays with the holder for life, which makes format checks simple. The government portal also says companies may not hold on to a person’s card, although they may collect the personal data printed on it.

Source: u.ae, Emirates ID (updated 17 February 2026)

Contracts, tenancy documents and bilingual letters

AI can summarise parties, dates, rent, renewal and notice periods into a register, and compare a draft against your template. In Dubai, tenancy contracts are registered through Ejari using the Unified Tenancy Contract and the applicant’s Emirates ID, so a property team’s document set is predictable. For letters and replies in Arabic and English, one rule keeps generative drafting safe: names, numbers and dates are copied from the record, never written by the model, and a person signs off.

Source: Dubai Land Department, Register or renew a tenancy contract

Our guide to what AI automation is introduces document extraction. This article covers building it properly.

How document AI works: OCR, layout, extraction, rules and review

A dependable document workflow has five layers. Disappointing projects usually skip the last two.

• OCR. Optical character recognition turns the pixels of a scan or photo into text, with a confidence score for each word.
• Layout. A layout model finds tables, key-value pairs and reading order, which matters where right-to-left Arabic sits beside left-to-right numbers.
• Extraction. A language model maps the text to your schema: supplier, TRN, total, expiry date. Generative AI helps most here, coping with hundreds of supplier layouts without a template for each.
• Validation rules. Ordinary code checks formats, totals, dates and lookups against your master data. Rules never guess.
• Human review. Anything below the threshold, or failing a rule, goes to a person with the page image and suggested values side by side. Each correction becomes test data.

This is generative AI for workflow automation in a back office: one generative step, boxed in by deterministic checks. The model reads and the rules decide, so a person only sees the cases where the two disagree or confidence is low.

The model reads and the rules decide, so a person only sees the cases where the two disagree or confidence is low.

Which services read Arabic: Azure, Google and AWS compared

“Supports Arabic” means different things on different products, and the gaps sit where UAE documents live: invoices, ID cards and handwriting. This is what each provider’s documentation listed on 14 October 2026.

Sources: Microsoft Learn, Read and Layout language support · Microsoft Learn, prebuilt model language support · Google Cloud, Document AI language support (updated 13 October 2026) · Amazon Textract FAQs

Azure is the only one of the three whose ready-made invoice model lists Arabic, and the only one listing Arabic handwriting. Google’s general processors read Arabic, but its Custom Extractor officially supports only English when extracting with generative AI. Textract does not list Arabic at all.

Source: Google Cloud, Document AI language support

None of the ready-made identity models covers the Emirates ID. That means OCR plus custom extraction inside strict format and expiry rules, with a person checking every result until your error rate is known.

What document AI costs per page

Page prices are published and low. These are list prices on 14 October 2026, converted at the dirham’s fixed peg of AED 3.6725 to the US dollar.

Sources: Azure Document Intelligence pricing, UAE North region selected · Google Cloud, Document AI pricing · u.ae fact sheet, AED 3.6725 to the US dollar

A worked example: 2,000 one-page supplier invoices a month on Azure’s prebuilt invoice model cost USD 20, about AED 73. Build effort, integration and the minutes people spend on exceptions decide a project, not the page fee. Our breakdown of what AI automation costs in the UAE prices the model tokens, platforms and maintenance around it.

E-invoicing changes invoice capture from 2027

The Ministry of Finance is explicit that PDFs, Word documents, images, scanned copies and emails are not e-invoices. Under the e-invoicing system, invoices travel as structured data between accredited service providers and are reported to the Federal Tax Authority. Businesses with revenue of AED 50 million or more had to appoint a provider by 30 October 2026 and must go live by 1 January 2027. Businesses below that threshold appoint one by 31 March 2027 and go live on 1 July 2027, and in-scope government entities follow on 1 October 2027.

Sources: Ministry of Finance, eInvoicing programme · Ministry of Finance, targeted amendments (10 May 2026) · Ministry of Finance, Ministerial Decisions on scope and timelines (29 September 2025)

For document AI, this moves the work rather than removing it. Once suppliers are live, their tax invoices reach you as data, so there is nothing to read; the AI work shifts to matching invoices to orders, coding them and explaining exceptions. The Ministry’s mandatory field list also changes your rules: your network identifier is your TIN, the first 10 digits of your 15-digit TRN.

Source: Ministry of Finance, UAE Electronic Invoice mandatory fields v1.0

Reading paper does not disappear. Smaller suppliers can send PDFs until their go-live date, and quotes, delivery notes, receipts and contracts are not tax invoices at all. Design one intake that accepts structured e-invoices and scanned documents and sends both through the same rules. That intake usually lives in your ERP, which is where our CRM and ERP team starts. For phases, penalties and provider choice, read our guide to UAE e-invoicing. This is general guidance, not tax advice.

Design one intake that accepts structured e-invoices and scanned documents and sends both through the same rules.

Emirates ID and onboarding documents under the PDPL

Identity documents are personal data, so the UAE Personal Data Protection Law applies: Federal Decree-Law No. 45 of 2021, in force since 2 January 2022. Article 2 excludes, among others, free-zone companies with their own data protection law.

Article 5 shapes the workflow: data must be collected for a specific purpose, limited to what that purpose needs, and not kept once the purpose is exhausted unless anonymised. So extract only the fields you need, decide whether to keep the image at all, set a retention period per document type and mask ID numbers in logs and prompts.

Consent is not the only legal basis: Article 4 allows processing without it when necessary to perform a contract with the person, to meet obligations in other UAE laws, or for employment obligations. Where you rely on consent, Article 6 requires that you can prove it and that it can be withdrawn easily. Article 18 lets people object to automated decisions that affect them, with listed exceptions, and requires human review on request, so never let a model reject an applicant on its own.

Source: UAE Legislation, Federal Decree-Law No. 45 of 2021, Articles 2, 4, 5, 6 and 18

Location matters too. Google’s Document AI processes in US or EU multi-regions or a few single regions, none in the Middle East, while Azure’s pricing page lists UAE North. Articles 22 and 23 govern sending ID scans abroad; our guides to the PDPL for product teams and cloud data residency go further. This is general guidance, not legal advice.

Sources: Google Cloud, Document AI locations · Azure Document Intelligence pricing, region list

Testing accuracy with a gold set before you trust it

Vendor demos use clean documents; yours include angled phone photos, stamps over totals and Arabic supplier names. The honest test is a gold set: real recent documents covering every supplier layout, scans and photos, Arabic, English and mixed pages, with the correct values keyed once by a person as the answer key.

• Score fields, not documents. Nine correct fields and a wrong TRN still put a wrong TRN in your ledger.
• Run every candidate on the same set, and rerun it after every change.
• Cover each variation. Microsoft advises at least five samples of each variation, such as digital and scanned PDFs, when training a custom model.
• Measure four numbers: field accuracy, the share of documents no person touches, review minutes per exception, and errors reaching the ledger, which should be zero.

Source: Microsoft Learn, accuracy and confidence scores

Then set thresholds. Microsoft explains that a field confidence of 0.95 means the prediction is likely correct 19 times out of 20, and that confidence can decide whether to accept a value or flag it for review. For custom models it recommends an accuracy score close to 100 per cent for financial or medical records. AWS suggests thresholds of 90 per cent or higher where financial decisions are involved. Start strict, review everything for the first weeks, and lower the share of manual checks only as the gold set proves it safe.

Sources: Microsoft Learn, accuracy and confidence scores · Amazon Textract, best practices: confidence scores

Calculator: the hours and dirhams your team gets back

Use a loaded hourly cost, not salary alone. Someone costing AED 10,400 a month all-in, over the 208 hours a 48-hour working week gives, costs AED 50 an hour. The share automated is the proportion that passes every rule above your threshold, known only after the gold-set test. Defaults are illustrative.

Source: u.ae, private sector working hours (8 hours a day or 48 a week)

Estimates only. Hours saved = documents × minutes ÷ 60 × share automated. Full-time equivalents assume 208 working hours a month. The result ignores running and build costs; the payback calculator in our AI automation cost guide adds them.

With the defaults, 1,500 documents at six minutes each is 150 hours a month. Automating 70 per cent returns 105 hours, worth AED 5,250 a month or AED 63,000 a year, and leaves 450 documents for review. Set the share to 40 per cent and the saving falls to 60 hours, which is why clean intake matters more than the choice of model.

Where OXEDGE can help: WhatsApp chatbot service in Dubai.

FAQ: generative AI for back-office documents in the UAE

Can AI read handwritten Arabic?

Some services can. Azure lists Arabic handwriting in version 4.0 of its Read and Layout models, Google’s language table does not confirm it, and Textract reads handwriting in English only. Test on your own forms before relying on it.

Can we keep copies of Emirates IDs?

Companies may not hold on to the physical card, but they may collect the personal data printed on it. Under the PDPL, keep only what the purpose needs, for as long as it needs it, and document your legal basis.

Can AI post invoices without a person?

Only those passing every rule above a strict confidence threshold and within an approval limit you set. Everything else should reach a person.

How much does document AI cost per page?

Published list prices are low. On 14 October 2026, Azure’s prebuilt models were USD 10 per 1,000 pages, about AED 36.73. A worked example of 2,000 one-page invoices a month on Azure’s prebuilt invoice model cost about AED 73. Build effort, integration and exception handling decide a project, not the page fee.

Does generative AI work for Arabic documents?

Support varies by service and by feature, so compare them on your own files. This guide covers which of Azure, Google and AWS read Arabic, and a gold set of your real documents shows the accuracy you can expect before you trust a tool.

If you would rather see your own documents tested before you commit to a tool, that is where we start.