AI in Malaysia · Professional services

AI for accounting and corporate services firms: the documents are the bottleneck

E-invoicing has changed what clients need from their accountants. For a small practice, AI's first job is the one nobody enjoys: getting client documents into the system.

Pikabyte Engineering 3 min read

A small accounting practice in Malaysia sells judgment and spends its days on data entry. Clients deliver a carrier bag of receipts, a folder of bank statements as PDFs, and photos of invoices on WhatsApp. Before any accounting happens, someone has to turn all of that into entries.

E-invoicing is shifting the picture. As the Inland Revenue Board’s mandate has rolled out by revenue band, more transactions arrive as structured data. That does not remove the work. It changes it. Clients now need help getting their own systems ready, and the documents that are still on paper are the awkward ones.

For a practice with ten or twenty staff, this is where AI earns its keep.

The work AI can take

Reading source documents. Receipts, supplier invoices, bank statements, payslips. A language model reads the image or PDF and returns date, supplier, amount, tax and a suggested account code. Staff confirm or correct.

Scans and photographs are not a barrier. In a system we built for a law firm, a model reads figures from bank statements of account that arrive as photocopies with no selectable text. The method is the same for an accountant’s inbox.

Bank statement lines. Converting a PDF statement into rows and suggesting a match against ledger entries. The matching is rule-based, the reading is AI.

First-pass coding. Suggesting an account code from the supplier and description, learning from how the firm coded the same supplier before. A suggestion, shown to a person.

Chasing clients. Drafting the monthly reminder listing exactly which documents are missing.

Drafting routine correspondence. Engagement letters, management letter points, replies to common questions, for a partner to review.

The work AI should not take

The judgment. Whether an expense is deductible, how a transaction should be treated, what to tell the client. That is the profession.

The arithmetic. Language models are unreliable at sums. Totals, tax computations and reconciliations should be calculated by ordinary software. Use the model to read, and code to add up.

The sign-off. Nothing goes to a client or an authority without a qualified person’s review.

Four design rules

1. Extract into a fixed structure. The model fills defined fields. Anything else is rejected.

2. Let it say “I cannot read this”. An empty field sent to a person is safe. A guessed amount is not.

3. Show the source beside the result. A reviewer should see the receipt and the extracted entry together. Checking then takes a few seconds per item.

4. Record who approved. For every entry: which document, what was extracted, who confirmed it.

Confidentiality

An accountant holds the most sensitive data a business has. Before using any AI tool:

  • Do not use consumer chatbots for client data. Use a business service where the terms say the data is not used for training.
  • Know the region. Where is it processed?
  • Tell clients. Add it to the engagement terms.
  • Limit what is sent. The model needs the invoice, not the whole client file.

These are requirements under the amended Personal Data Protection Act, and they are what a careful client will ask about.

A new service line

There is an opportunity as well as a saving. Your clients are small businesses facing the same e-invoicing and digitalisation pressure. Many have operations systems, point-of-sale and spreadsheets that do not talk to their accounting software.

A practice that understands how to connect those is offering something more valuable than bookkeeping. In one project, we connected a law firm’s case system to its accounting package so that invoices are created directly from the matter. No re-keying, no missed bills. Accountants are well placed to spot where a client needs that.

Where to begin

  1. Count the hours your team spends on data entry in a typical month.
  2. Choose one document type with high volume, usually supplier invoices or bank statements.
  3. Trial extraction on one willing client’s documents.
  4. Measure the correction rate honestly.
  5. Extend to the next document type.

A supplier should be able to show extraction working on your own sample documents before you commit. We build that demo first for every client, before any quotation.

A practice that gets documents in automatically has more hours for the work clients actually value.

Sources

This article is general information and is not accounting, tax or legal advice.

Written by

Pikabyte Engineering

Pikabyte builds custom business systems, applied AI and secure cloud infrastructure from Malaysia. Talk to us.

Start a conversation

Want this applied to your business?

Tell us what is slowing the business down. You will talk to an engineer, not a salesperson, and get a straight answer on whether software is the right fix.