Case study: how an Ipoh accounting firm cut month-end from 12 days to 4 with document-AI

A 12-person accounting and tax-advisory practice in Ipoh old town was bleeding 12 working days every month to a manual bookkeeping close that should have taken four. The two partners were spending their best evening hours rekeying Maybank and CIMB statements, sorting WhatsApp receipt dumps, and chasing missing invoices. After eight weeks with our document-AI build, the month-end close was down to four days, the partners were back doing tax advisory at 7pm instead of typing, and the junior staff had time to learn IFRS for SMEs. Here is what we built, what it cost, and what your firm would pay.

The business

A 12-person, second-generation accounting and tax-advisory practice operating out of two adjoining shophouses along Jalan Sultan Yusuf Shah, within walking distance of the white-coffee stalls on Jalan Kek Lok. Founded in 2003 by a father who is now semi-retired, the firm is run by his two children — both qualified chartered accountants — and ten staff: four semi-seniors, three juniors, a payroll specialist, an admin, and a part-time IT support. They serve around 180 active SME clients across Perak — manufacturing in the Kinta Valley, retail in Greentown and Canning Garden, F&B, and a small slice of professional services. About 60% of the book is on full bookkeeping, payroll, and tax-submission; the rest is tax-only or year-end audit. The firm uses SQL Account, exports to Excel heavily, and was running on a paper-receipt culture nobody had been asked to break.

The problem

The month-end close was the firm's slow bleed. Every cycle: paper and PDF receipts arrived in a stack (or, increasingly, a 30-to-60-photo WhatsApp dump); a junior rekeyed them into SQL Account; the semi-senior reviewed the trial balance; one of the two partners rekeyed bank-statement lines by hand. By the time all 180 clients had closed, the calendar said the 25th of the following month. The partners called it "the 12-day close" and they hated it.

Three things were eating them alive:

• Bank reconciliation alone took six days a month. One of the partners spent two to three hours every evening for the first two weeks of each month matching bank-statement lines to journal entries. On the first call, almost offhand, he said: "I trained as a chartered accountant, and I am spending my evenings playing spot-the-difference with a Maybank PDF and a stack of petrol-station receipts. That is not the job I signed up for."
• Receipts arrived in seven different shapes. Paper, a 30-photo WhatsApp dump, PDFs, a 200-line spreadsheet from the biggest client — a 60-employee hardware retailer in Menglembu, Kinta Valley — that always needed a second pass.
• The five-day review window was eating the partners' nights. Even after a junior rekeyed everything, the senior review took another four to five days. Clients were calling, the partners were apologising on WhatsApp at 9pm, and the firm had lost two juniors to Big Four poaching calls in the previous 12 months.

The two partners said it plainly in week zero: "We do not want to replace ourselves. We want to do our actual job — the part where a qualified human looks at the numbers and tells a client their margins are slipping — and we want to stop being the bottleneck on data entry."

What we built — and why this approach

A focused document-AI layer on top of the firm's existing SQL Account + Excel stack — not a replacement, not a new ERP, not a six-month migration. The principle: change the data-entry work, change nothing about how the partners run a job. The build was in four parts.

1. A unified intake — a dedicated firm WhatsApp Business line, a dedicated firm email intake, and a small in-tray folder for paper scans. Any document that lands in any of those three channels goes into the same queue, classified by client and document type, and read.
2. A document-AI reader that handles the seven input shapes — paper scan, photo dump, PDF, CSV, Maybank/CIMB/Public Bank PDF statements, the Menglembu spreadsheet. It extracts vendor, date, amount, SST code, and project tag, with a confidence score per field. Anything under 90% confidence goes to a junior review queue with the source document side-by-side.
3. A bank-reconciliation engine that does the matching, not the partner. Statement lines load automatically and are matched to journal entries by amount, date window, vendor, and reference-string similarity. The partner sees only the unmatched lines — usually five to twelve per client per month, almost always legitimate items the AI could not confidently match. This is the part the partner was doing by hand, and it now takes 15 minutes per client instead of two hours.
4. A small review console for the semi-seniors — a daily list of "needs your eyes" items with the source document and the extracted fields side-by-side. The semi-senior confirms or fixes; a fix feeds back into the model.

Why a focused build on top of the existing stack, instead of a full AI bookkeeper or a software migration? Three reasons. The first is plain cost. The second is the Malaysian accounting reality: every client's "stack" is different, every format is different, and the firm's job is to absorb that mess. The third is the part nobody likes to say out loud — the partners' judgement is the actual product. The build protects that judgement; it does not replace it. The same shape of build, on a different document mix, is what we did in the Alor Setar printing shop document-AI case study.

How the operation changed

The first thing that changed was the WhatsApp dump. The biggest client in Menglembu had been sending a 200-line spreadsheet plus a 40-photo WhatsApp dump every month-end. After week two, the partner told the client: "Just WhatsApp everything to this number. Don't worry about ordering it." The client's admin sent the same dump. The system read the spreadsheet, read the photos, matched the lot, and produced a TB draft in 18 minutes. The junior's job on that client dropped from a full day to a 90-minute review.

The second thing that changed was the partners' evenings. The two-to-three-hour nightly reconciliation block collapsed to a 15-minute review per client, done in the office, in daylight, with coffee. The same partner started running a 7pm client call every Tuesday to talk through the month's management accounts. He said on week six: "I am a chartered accountant again. I have an opinion about my clients' numbers. I forgot what that felt like."

The third thing that changed was the junior-to-senior pipeline. With six days a month of bank-reconciliation time back, the firm put the juniors on a paid six-month IFRS for SMEs study group. The quiet attrition risk went quiet, and the partners hired a third junior because the workload was now genuinely sustainable.

The numbers, 8 weeks in

• Month-end close cut from 12 working days to 4. The partners wanted five; they hit four. The fifth day is now a buffer for late-arriving documents, not a crunch.
• Bank-reconciliation time cut from ~6 days to 11 hours per month. Per the partner's own log; he tracked it manually for the first four weeks, then stopped because it stopped being a problem.
• Junior rekeying hours down 71% on the top 20 clients by volume (week zero vs week eight).
• Zero clients lost in the transition. The most common feedback was "you seem less stressed this month".
• The partners got 14 evening hours a week back. One of them is back at the badminton club on Wednesday nights.
• Total program cost was RM 28,800 (build + first 90 days of tuning). The firm's monthly bookkeeping revenue across the 180 clients is roughly RM 78,000, so the payback was inside 12 working days — well inside the 30-day target the partners had set on the first call.

What it would cost for your business

We size every document-AI build on three things: client-count complexity, document-format mix, and how much of the existing stack you want to keep. Indicative bands, in the same shape we use for the Perak services hub:

• Low band — RM 14,800 to RM 22,000. For a three-to-six person practice, 40 to 90 active clients, mostly PDF and WhatsApp dumps, existing SQL Account or Xero. Five to six weeks.
• Mid band — RM 26,000 to RM 42,000. The build we did for the Ipoh firm — 180 clients, mixed document formats, full review console, bank-rec engine that runs end-of-day. Seven to nine weeks.
• High band — RM 55,000 to RM 95,000+. Multi-branch practices, firms with audit and corporate-secretarial arms, or client-facing dashboards and a real mobile app. 10 to 14 weeks.

All three bands are flat-fee, fixed-scope, and include the first 90 days of tuning. We do not bill by the hour. We do not mark up human-coding as if it were AI work. The build model is the same as for our AI-agency track work — the engine and the architecture are the same regardless of region.

What this looks like for your business

If you are running an accounting or tax-advisory practice in Ipoh, Taiping, Sitiawan, Teluk Intan, or anywhere in Perak, the build we did for the 12-person firm is the same shape most firms need. The intake layer changes a little — some firms prefer email-first, some WhatsApp-first. The reader and the reconciliation engine are the same.

We have done the same shape of build for an Ipoh dental group, an Alor Setar printing shop, and a Taiping auto-parts distributor. You can read those: the Ipoh dental clinic patient-communication build and the Taiping auto-parts order-automation build. Each one is a real anonymised firm; each one had a different pain, the same kind of fix.

If you want to talk about whether your firm is a fit, drop us a line and tell us your client count, your existing stack, and the part of the month that hurts the most. We will tell you on the first call whether the build makes sense, what it would cost, and how fast it would pay back. We do not pitch you either way — about a third of the firms we talk to, we end up recommending they wait, because the numbers do not yet justify the build.