01 / Accounting & Finance Ops
Xero just moved AI agents into your ledger - some of it is live now
Xero used Xerocon Denver this week to lay out where your books are heading: a “month-end agent” that reviews reconciliation status and drafts routine journals before handing you a full account of its actions to approve, cash-flow tools that spot gaps before they arrive, and scenario planning grounded in your actual Xero data. Also announced: Xero data flowing directly into Excel, Word and PowerPoint through Microsoft 365 Copilot, and a Xero connector for ChatGPT arriving in coming weeks.
The honest reading: most of this is badged “coming soon” or early access. What is live and proven is Auto Bank Reconciliation - over 100 million transactions auto-matched since launch, with an explanation attached to every match. Xero's design principle throughout is the right one: the AI acts, you approve. That's the pattern worth copying in any automation, not just accounting.
Owner takeaway: Open Xero this week and check whether Auto Bank Reconciliation is switched on and how many of your transactions it's already matching - then ask your bookkeeper which of the announced agents they plan to use, so you're not paying for manual reconciliation that software now does.
Original source: Xero Blog - Xerocon Denver 2026
02 / Everyday Tools
The AI assistant moved into the browser tab - where your real work lives
Anthropic upgraded its Claude in Chrome extension this month: the side panel now runs a full working session that can see the page you're on and act in it - clicking, typing, filling forms - using your existing logins, with the conversation carrying over to desktop and mobile apps. It's on the higher plans now and rolling out to Pro subscribers through late August.
Why this matters more than another chatbot feature: most small-business “systems” are really browser tabs. Supplier portals, insurer dashboards, government forms - none have integrations, and nobody is building them any. An assistant that works inside those pages addresses the admin no API will ever reach. Credit where due: Anthropic plainly documents the main risk, prompt injection - hidden instructions in a webpage that can redirect the AI - and has built approval checkpoints before consequential actions. The first practical test would be a read-only job: collecting invoice details from a vendor portal into a spreadsheet.
Owner takeaway: If anyone in your business pays for Claude, have them try one low-stakes browser task this month with approvals left on - and keep it away from anything involving payments until the pattern is proven.
Original source: Anthropic - Claude Cowork comes to the Chrome side panel
What I would do first
Start in Xero, because it's already open on your screen: check Auto Bank Reconciliation is on and doing its share. It's live, it's included, and it's the clearest example of the pattern both items here point to - AI doing the routine work, with you approving the result. The browser assistant is the same idea one step further out; give it one boring, read-only task before you trust it with anything that matters. Neither test costs a cent, and both will tell you something true about how much routine work your business is still doing by hand.
Running a business that needs someone accountable for technology decisions, not just automations? See the fractional CIO seat.