Insights, guides, and updates
An AI agent was asked to attach a spreadsheet to an email. It couldn't reach the folder, so it found a file with the same name in an old email, attached that, and reported the job done. In a CAS practice the same failure looks like a flawless margin analysis built on the pre-adjustment trial balance, and none of the checks your team was trained to run will catch it.
BILL is repricing AI-agent work from seats to consumption, Xero says one in five ecosystem connections is now a custom-built app, and frontier model prices dropped again. Part 4 of Buy vs. build re-runs the series' five-layer framework against a month of on-the-record evidence, and the question has moved: not whether firms will build, but what's bounded enough to own.
Intuit told shareholders it will deliberately earn less per DIY tax customer next year rather than keep losing them to cheaper software. That decision, plus a free QuickBooks tier with no accountant seat, is the clearest signal yet of what happens to the price-sensitive end of a professional services market. Also this week: Singapore started training half its accountancy profession, agents learned to log into websites, and OpenAI published a postmortem showing its monitoring system was switched off while 1,200 agents broke into someone else's servers.
I spent last week on the floor at Xerocon Denver, and I watched two movements happen at once: the platforms assembling and repricing the whole stack from above, and the firms quietly building their own software from below. Add BILL's pivot from seats to consumption pricing and OpenAI's answer to the client-data question, and the week's message is clear: the gap between falling AI prices and rising platform prices is where your next margin decision lives.
In four weeks, AI disclosure went from ethics debate to infrastructure: Substack scans it, LinkedIn flags it, the EU legislates it, and Anthropic watermarks everything Claude writes. Your firm almost certainly faces no legal risk, but your work product now carries flags you can't remove. The only question left is whether the story that arrives with the flag is one you wrote.
Xero's AI has coded 100 million transactions, and practitioners trust exactly one of its four methods: the one built from written rules. Every coding engine on the market infers your client's intent from history. Not one of them will read it, even though you could write it down today.
