1,500+ invoices a month into AutoCount, with 80%+ less time per invoice.
A Malaysian accounting firm now moves supplier invoices into AutoCount with 80%+ less time each, and an accountant approves every draft.
The 30-second version
A Malaysian accounting firm was keying 1,500+ supplier invoices a month into its clients’ AutoCount books by hand, at about 5 minutes each, with more arriving as it took on clients. We built an AI automation that reads each invoice, codes it from that client’s own booking history, checks it, and creates a draft in AutoCount once an accountant approves it. Each invoice now takes under a minute including review, 80%+ less time, and the team recovers 100+ hours a month. The firm runs it every month and keeps connecting more client books.
Manual key-in
Every invoice opened, read, coded and typed into the right client’s book by hand: 125+ hours a month.
AI invoice automation
One workflow that reads, codes and checks each invoice, then creates an approved draft in AutoCount.
80%+ less time
About 5 minutes per invoice down to under 1, including the accountant’s review.
100+ hours back
More than half of one full-time staff member’s month, returned every month. Live within a month of kickoff, and in use every month since.
Who we built this for
The firm’s clients are not named. The profile below reflects the real engagement.
- Industry
- Accounting & bookkeeping
- Business type
- Malaysian accounting firm
- Team involved
- Accounts & bookkeeping
- Accounting system
- AutoCount Cloud, one book per client
- Invoice volume
- 1,500+ supplier invoices a month, and growing
- Client books
- A growing number of companies
- Project scope
- Custom AI invoice-to-AutoCount system
- Timeline
- Live in one month, running monthly since
AutoCount wasn’t slow. Getting invoices into it was.
An accounting firm doesn’t keep one set of books. It keeps one for every client, and every invoice had to land in the right one, coded the right way.
Manual key-in at volume
1,500+ supplier invoices a month, about 5 minutes each: 125+ hours of typing every month, and more with each new client.
Every client codes differently
The same kind of expense goes to a different GL account in each client’s book. Staff had to remember which one, client by client.
The wrong-client risk
With many clients flowing through one team, the easiest mistake is filing an invoice in the wrong company’s book.
Duplicates slip through
The same invoice arrives twice, emailed and then uploaded, and the only check was someone’s memory.
Documents everywhere
PDFs in email, scans in shared folders, phone photos from clients. Nothing had one place to live.
Off-the-shelf OCR fell short
Generic OCR reads the text, but it doesn’t know each client’s chart of accounts, so the coding, checking and entry were still manual.
Six manual steps for every invoice
Around 5 minutes each, repeated for 1,500+ invoices a month, with every check done by eye.
- Step 1
Collect
Gather invoices from email, shared folders and client messages.
Scattered - Step 2
Read
Open each invoice and read the supplier, dates, amounts and lines.
Slow - Step 3
Look up
Find the supplier’s creditor code in that client’s book.
Per client - Step 4
Code
Pick the GL account from memory, different for every client.
Inconsistent - Step 5
Key in
Type it all into AutoCount.
~5 min each - Step 6
Check
Spot duplicates and wrong-client invoices by eye, if at all.
Error-prone
What success needed to look like
We set the targets up front, so the build was measured against outcomes, not features.
Cut key-in time
Replace 5 minutes of typing per invoice with a quick review of fields the AI has already filled in.
Code like the accountant
Suggest the GL account each client’s books would actually use, not a generic category.
Catch the wrong book
Catch an invoice addressed to one client before it becomes a draft in another client’s book.
No duplicates
Stop the same supplier invoice from being entered twice, however many times it arrives.
Accountants stay in control
Automate the typing, but keep every decision about the books with the people responsible for them.
Grow with the client list
Taking on a new client shouldn’t mean rebuilding the system or calling us in.
One automation between the invoice and AutoCount
We built a custom AI automation that sits in front of the firm’s AutoCount books. It reads each invoice, matches the supplier, suggests the GL account from that client’s own booking history, runs the safety checks, and creates a draft purchase invoice once an accountant approves it.
The goal was not just to extract invoice data. Generic OCR could already do that. The firm needed the whole accounting step handled: each client’s coding rules, the checks an accountant does by eye, and an approval before anything reaches the books. AutoCount stays the system of record, nothing is written until someone approves it, and everything it writes is a draft.
An AI automation layer, connected to AutoCount.
Invoices come in from wherever the firm’s clients send them. The automation layer reads, codes and checks them, and an accountant approves. AutoCount stays the system of record; it just stops needing a typist.
The same invoice, in under a minute.
The system does the reading, coding and checking. A person steps in to review and approve.
- Step 1
Capture
Uploaded, emailed in or dropped into a shared Drive folder.
- Step 2
Read
AI pulls the fields from a typical invoice in about 10 seconds.
- Step 3
Code
Supplier and GL account suggested from the client’s history.
- Step 4
Check
Duplicate, currency, document-type and wrong-client guards.
- Step 5
Approve
An accountant reviews and approves on the dashboard.
- Step 6
Draft
A draft purchase invoice lands in the right AutoCount book.
What makes it different
Plenty of tools can read an invoice. These are the parts that make it safe to let a system write drafts into real clients’ books.
Learns how each client books
Reads up to 8 years of that client’s purchase invoices, payments and journal entries. If the accountant always codes a supplier to a certain account, so does the automation. Each client’s history stays with that client.
Can’t invent account codes
The AI can only choose accounts that exist in that client’s chart of accounts and can be posted to. If a code isn’t in the live book, no draft is created with it.
Catches the wrong-client mistake
Reads the invoice’s Bill-To. If it doesn’t match the client being worked on, the accountant has to confirm twice.
Refuses what shouldn’t go in
Checks AutoCount itself for the same supplier invoice number. Credit notes and statements are refused; quotations, proformas and receipts need a confirmation.
Nothing touches AutoCount before approval
Even a new supplier is only created at approval, following that client’s own creditor-code pattern. Everything goes in as a draft.
A new client in minutes
Add the client’s AutoCount keys in Settings and test the connection. No rebuild, and it starts learning from that book’s history straight away.
What stays with the accountants, and what we chose not to automate.
The system prepares the work. The people responsible for each client’s books still make every decision that changes them.
What remained human-controlled
- Approval of every invoice. Nothing is written to AutoCount until an accountant approves it on the dashboard.
- Wrong-client warnings. When the Bill-To doesn’t match, the accountant must confirm twice before a draft is created.
- New suppliers. A creditor is only created at approval, using the client’s own code pattern.
- Finalising the books. Entries arrive as drafts, so the firm keeps the final say in AutoCount.
What we deliberately did not automate
- Posting final entries. The automation creates drafts only. It does not post to the ledger on its own.
- Credit notes and statements. These are refused rather than guessed at.
- New account codes. The AI can only use accounts that already exist in that client’s chart of accounts.
- Quotations, proformas and receipts. These need a person’s confirmation before they go any further.
From kickoff to live in one month
Built in phases against the firm’s real books, running on them by the end of the month, and extended to more client books since.
Discovery
Mapped how the team handles invoices, and reviewed the existing books (read-only) so drafts match how the firm already books.
Reading & coding
Built the AI extraction, supplier matching and GL learning from each client’s own booking history.
Guards & approvals
Added duplicate, currency, document-type and wrong-client checks, plus review and approval on the dashboard.
Launch
Multi-client setup, error alerts, daily backups and a team guide, then live on the firm’s books.
Each piece chosen for one job
Read accurately, write safely, and keep the firm’s data under its control.
AutoCount Cloud
The system of record. One book per client, connected through its API.
Review dashboard
A custom web dashboard where the team reviews, approves and manages clients.
n8n
The workflow engine connecting intake, AI, storage and AutoCount.
Mistral AI
Reads PDFs, scans and phone photos of invoices.
OpenAI
Structures the extracted fields and suggests GL accounts from booking history.
Self-hosted database
The review queue, client settings and activity log, on a managed server.
Google Drive
Stores every source invoice and the daily backups.
Telegram
Alerts when a run fails, so problems are spotted quickly.
The same team, without the typing.
The volume keeps growing. The time each invoice takes came down by 80%+.
That’s more than half of one full-time staff member’s working month, back every month: time the firm now spends on reviewing, advising and taking on more clients.
- Invoice volume Pilot-phase volume
- 1,500+ supplier invoices a month across the client books connected in the pilot phase. Volume grows as the firm connects more client books.
- Time per invoice Before vs after
- About 5 minutes of manual key-in before; under 1 minute after, including the accountant’s review and approval: 80%+ less time per invoice.
- Hours recovered Calculated
- 1,500 invoices × at least 4 minutes saved ÷ 60 = 100 hours a month at pilot-phase volume. The time saved per invoice and the full project timeline are separate measures.
- Time to read Typical invoice
- About 10 seconds for the AI to read a typical invoice.
Recovered staff time is capacity, not a cash saving. The dashboard at the top of this page is an illustrative interface with sample data.
“This saves our team a lot of time. We’d looked at off-the-shelf OCR software that does something similar, but none of it comes anywhere near the accuracy of a workflow tuned to how we actually work.”
What’s happening now: the firm runs the workflow every month and keeps connecting more client books, and we’re extending it into more parts of the firm’s work.
Why it matters
More of our work.
Three more client projects, with the workflow before and the system after.
How a Capital-Raising Team Cut Investor Research from 15 Minutes to 1
- 70%
- shorter deal pipeline timeline
- 15×
- faster than by hand
How a Private Credit Firm Researches a Target Desk in About 7 Minutes
- ~7 min
- per desk, unattended
- 90%+
- coverage on target desks
How a Recruiting Agency Scores Every Applicant in About 20 Seconds
- ~20 s
- from application to ranked row
- <$2
- AI fees per 500 resumes
How many hours is invoice entry costing your firm?
Start with a free discovery. We’ll walk through how your team handles invoices today, then prepare a tailored initial plan and proposal for your clients’ books. Implementation is quoted after discovery.