Moushumi at Ethical CFO takes on rescue jobs that other accountants quietly back away from. Here's how we used Claude to chew through five years of bank statements, and the second step we built for everything the model can't see.

I jumped on a call with Moushumi Sikand a couple of weeks ago. She runs Ethical CFO out of Melbourne, Xero Sole Practitioner of the Year, and she's mostly known for rescue work. Clients 1 to 5 years behind on reconciliation, no Xero set up, no GST lodgements, no books.

We spent a Friday building two things. After we did, she left this on my LinkedIn:

Here's how it works.

The rescue-job version of bookkeeping is the same job repeated thousands of times. Click into a transaction. Read the bank description. Decide who the merchant is. Decide what account code applies. Write a short Why. Move on.

Now do that across 5 years of bank statements.

What I built for Moushumi is a Claude Chrome extension that does the clicking for her.

It reads each transaction in the Xero reconciliation screen. Identifies the merchant. Picks the most appropriate Xero account code. Writes a plain-English Why. Fills the Who / What / Why fields straight down the page.

The whole thing runs from one prompt. Claude is told to behave like an experienced Australian bookkeeper. Don't overwrite anything that's already filled. If the description doesn't tell you enough, leave the field blank. And the one rule that matters most. Never click OK. Never click Submit. The human approves every page.

Watching it run is the bit that sells it. Entertainment, hotels, software subscriptions, NAB bank fees. Straight down the screen. The pace of a fast junior bookkeeper, but it doesn't stop.

Moushumi's first question wasn't "does it work."

It was "does it do GST?"

Which is the right question. GST assessment wasn't in the prompt yet. It's the next thing I'm adding. Then she watched the Loom of it processing transactions one after another and gave me the line I keep coming back to:

"It's still going to speed up the process."

That's the rollout pattern in finance right now. Not "AI replaces the bookkeeper." More like "AI does the first 70% so the bookkeeper can spend their brain on the 30% that matters."

There's a limitation worth flagging. When the bank description is blank, Claude leaves the field empty and a human has to think. On rescue jobs, a lot of descriptions are blank. The Chrome extension hits a wall at that point. It can't see what you can't see.

That's where part two comes in.

Once you've reconciled what you can reconcile, you're left with the awkward bits. Transactions with no description. Mystery transfers. Cash purchases. Receipts the client hasn't filed.

For Moushumi's current rescue client, those are sitting in an inbox somewhere. Years of supplier emails. Receipts forwarded. PDFs as attachments. No folder structure, no system, no way to find anything fast.

The play here is a separate AI step. Not the same Chrome extension. A different tool.

You categorise the inbox first, then search across it intelligently.

Claude tags every email by type. Supplier invoice, receipt, customer payment, internal admin. ABNs and amounts get extracted. Once that pass is done, the email database becomes searchable in a way it never was inside Gmail or Outlook on its own. RAG-style retrieval over the cleaned-up inbox.

So when the reconciler hits a blank transaction for $432.16 on July 14, 2023, the bookkeeper can pull up every email that mentions an invoice on or near that date and that amount. The right supplier invoice surfaces in seconds instead of an afternoon.

It's two pipelines, not one. But they work in sequence and they cover each other's blind spots. The Chrome extension does the obvious stuff fast. The email search picks up the obvious's leftovers.

I'd done something similar before with Zoho Mail for a legal-email-logging case. Different industry, same shape. Point AI at the messy data the client already has rather than asking them to organise it first. The organising never happens. The data is already there.

Quick honest beat before we wrap up.

I also built a ClickUp dashboard for Moushumi. Week view across all her clients, capacity planner, holiday-conflict view that found 47 tasks landing during a vacation week she'd been planning. Pretty awesome, I thought.

She said no.

Not because the dashboard didn't work. Because she diagnosed the real problem in about 90 seconds, her ClickUp tasks were structured at too micro a level. Every sub-step was its own task. The dashboard wasn't going to fix that. The data layer underneath was wrong.

A good operator can tell the difference between an AI problem and a data problem. That's worth more than the dashboard.

Operators already using AI for the soft work, like commentary, drafting, and summarising, were the easy first wave. They got comfortable trusting the model with the writing. That trust matters.

What's happening now is different. AI isn't just drafting copy. It's doing the work. Categorising transactions. Tagging emails. Filling Xero fields. The human's job has shifted from doing the work to checking it.

The frame that fits is AI as a super-junior bookkeeper. It does the click-by-click at speed. The senior CFO reviews. Same trust pattern that already exists between any senior and their junior. The senior knows where the junior makes mistakes. The senior knows which judgement calls to override.

That's the real rollout. Not "AI replaces the bookkeeper." More like AI is the bookkeeper, and the senior is the one who checks.

Moushumi's sending a referral my way. A client who needs a custom inventory and delivery app. That's how this stuff actually spreads.