Solution guide

Document workflows where AI prepares and people decide.

Invoices, forms, contracts and statements still get re-typed by hand in many offices. AI can read them and propose the key details in seconds — but a wrong figure in an invoice is a real problem. The reliable pattern is simple: let AI prepare, let a person check, and keep a record of what was decided.

Request a project quote Try the document review demo

The pattern in four steps

Read. The system takes in a document and proposes values for the fields you care about — supplier, date, amount, reference.
Check automatically. Simple rules catch the obvious: totals that do not add up, dates in the wrong order, a missing reference.
Review. A person sees the document next to the proposed values, corrects anything and accepts. Uncertain items are highlighted, not hidden.
Hand over. Accepted data flows into your accounting, spreadsheet or system — and the decision is logged.

The review step is not a weakness of the design; it is the design. It is what lets you use AI on work where an error would matter.

First ask: is AI even needed?

If your documents come from a handful of suppliers with a fixed layout, fixed rules or a template can read them more cheaply and more predictably than AI. AI earns its place when:

layouts vary a lot from one sender to the next;
the information is in free text rather than in boxes;
documents mix languages or formats;
people spend hours on reading and re-typing.

Sometimes the answer is a mix: rules for the regular senders, AI for the long tail. We would rather say so than reach for AI by default.

What to decide before processing a real document

How accurate is accurate enough? This depends on the field. A wrong supplier name is an annoyance; a wrong amount is a financial error. Set targets per field.
What happens when the system is unsure? It should say so and ask a person — never guess silently.
Who reviews, and how fast? The workflow is only as quick as the reviewing step. Plan capacity for it.
Where do the documents go? Some AI services send content to external providers. What that means for confidential or personal data needs deciding — and checking against the provider's terms — before any real document is processed.
How long are documents and extracted data kept? Agree retention and deletion up front.

Hong Kong documents: two languages, several currencies

Bilingual and mixed-language documents. Invoices that put English and Traditional Chinese on the same page are common. Test on your real documents; do not assume a tool that works well in English performs the same in Chinese.
Multi-currency amounts. HKD, RMB and USD often appear together, so currency recognition and formatting need explicit checks.
Personal data inside documents. Staff records, ID numbers and contact details may appear. Think about the Personal Data (Privacy) Ordinance before documents leave your own systems, and take advice where the content is sensitive.
Scans, photos and stamps. Many documents arrive as phone photos or scans with stamps over the text. Quality varies, so measure rather than assume.

Try the demo — and know what it is

The document review demo shows a sample invoice next to editable extracted fields, so you can see what the review step feels like. It does not use a real AI service: the results are predefined to illustrate the interaction, using fictional data.

Illustrative demo · Fictional data · Simulated AI extraction

Open the document review demo (English)

How we would start

Collect real examples — a spread of typical and awkward documents, with sensitive details handled as agreed.
Define the fields and the rules that make a result acceptable.
Measure first. Try simple rules and AI on your examples and compare accuracy before committing to a design.
Build the review screen early, because it is where the value and the risk both sit.
Start with one document type and widen only when it works.

What drives the effort: how many document types, how varied they are, how many fields, the integrations at the end and the review workflow. See the cost estimator for a first sense of range.

Frequently asked questions

How accurate will it be on our documents?

We cannot promise a number before testing on your documents. Accuracy depends on layout, language and scan quality, and it differs by field. That is why the first step is to measure on real examples, and why a review step stays in the design.

Does our data leave our systems?

It depends on the approach. Some AI services process content on an external provider's servers. We would explain the options, check the provider's data terms and agree what is acceptable before any real document is used.

Can it read Chinese and English documents?

Often, yes — but performance varies, especially with mixed-language layouts and low-quality scans. We would test on your own documents rather than assume.

Could a simpler automation do the job?

Sometimes. For fixed layouts, rules or templates can be cheaper and more predictable. We would compare the options with you and recommend the simplest one that works.

What does it cost?

It depends on the number and variety of document types, the integrations and the review workflow. There are also running costs if an AI service is used. Scoping makes both visible before you commit.

Drowning in paperwork?

Tell us which documents take the most time and what ends up in your systems. We will help you work out where AI helps, and where it does not.

Request a project quote