A bookkeeper who subscribed last week has 10 clients on 10 separate Microsoft tenants. Which means 10 Outlook accounts. Every morning he used to open all of them.

So he asked Claude to help him build a Power Automate flow. Now it scans the 10 inboxes at 7am and sends him one email. Here are your tasks, here are your meetings.

Took him an afternoon. Costs nothing on top of what he already pays Microsoft.

Westpac did the same thing this week. With 35,000 staff, six partner logos and their own NVIDIA AI factory.

And the number that stuck with me wasn't on the logo slide.

Dan Jermyn, their Chief AI Officer, told investors that an intranet migration which "typically would cost low seven figures to achieve and take about a year" got done for US$12,000 in tokens. About A$16,880.

A bank with a data centre full of GPUs got its headline win from the same tool the bookkeeper used.

I read the whole Westpac deck so you don't have to. Two things worth keeping.

First, they split AI into two buckets and never mix them.

Everyday AI is Copilot in the hands of every person. They don't count licences. They count people using it more than 10 days a month.

Operations at scale is five narrow agents. They read payslips and bank statements. That's it. 32,000 payslips a week, and about 150,000 banker hours back a year.

Five agents. Nothing clever. Each one reads a document and checks it against a rule.

Second, the expensive bit wasn't the AI. It was pulling 285 source systems into one place so the agents had clean data to read.

That's the part you don't need. A practice runs on Xero, a practice manager, Outlook, a shared drive and maybe a CRM. Five systems, not 285. Same shape, one two-hundredth the size.

Takeaway: the bank's whole stack is a copilot for people, a few narrow agents for documents, and one place the data lives. The 285 is their problem, not yours.

Here's the recipe, and it's the same one whether you're Westpac or a three-person practice in Penrith.

A document arrives. The agent reads it. Checks it against a rule. Writes the result somewhere. A person decides.

Automate the first four. Keep the fifth.

The bookkeeper's morning digest is that recipe pointed at an inbox.

His next one is the same recipe pointed at a bank feed. Scan transactions for GST exceptions. Same supplier, same transaction type, but this one's missing the 10%. Flag it, don't fix it. That's Westpac's account scrutiny agent, pointed at a BAS.

A finance manager in Perth is doing it for month-end. Prepayment and amortisation journals, reconciliation review. Xero just announced an agent that does exactly that, in early access. Try the built-in before you build.

Takeaway: don't start with a platform. Start with one document that arrives every week and one rule you check by hand. Automate the check. Keep the decision.

The same bookkeeper has a client, a concreting company doing about $25M a year. They spent three to four months building a data lake and teams of AI agents on Airtable.

His words: largely a flop. AI was meant to be the silver bullet.

Westpac built a data platform because they had 285 systems and a reason. A $25M concreter has about five and doesn't. Wrong layer first.

One honest note on Westpac too. Their adoption and productivity numbers are internal, not audited. And their own risk slide flags third-party AI concentration. Six partners is six dependencies.

Takeaway: copy the shape, not the size. If you're building a data lake before you've automated a single check, stop.