10 articles
Digits gave AI agents write access to its ledger on October 13, through a connection that already reaches every client you can access: the two things a practice needs before an agent can do real work. Xero delivered client switching, but only in Gemini so far. And JPMorgan drew a line at the customer's password that turns out to sit in roughly the right place.
Bill Gates says AI will cross the reliability threshold for accounting "in the next few years," and the same week the Journal of Accountancy printed an Excel agent reconciling 59,157 general ledger rows in about 10 minutes while Accounting Today tallied Big Four reports with invented citations. Meta and OpenAI both shipped personal agents that can send a client's invoice, and Microsoft moved more of Copilot onto a meter. The work is getting closer to the threshold, and the sign-off you can audit hasn't caught up.
Intuit's agent will run the month-end close from your firm's own SOPs, at no charge until January. Mercury put a general ledger inside the bank account, free through December, with unlimited free seats for the accountant. Claude's close stopped requiring QuickBooks, and a new model prices a yes-or-no decision at a fraction of a cent. The close was offered three times this week with no price of its own, and what the platforms want in exchange is your playbook.
Puzzle, the $60 AI-native ledger we priced against QuickBooks two weeks ago, is selling its firm-facing business to Accrual and will stop marketing software directly to CPA firms. Also this week: OpenAI launched GPT-6 with a financial-statement tie-out as its lead customer story, Docusign is opening its connector to every AI agent on September 30, and a services firm priced an agent-led assurance engagement 30 to 50% below its own rate card.
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.
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.
Last week the joke was that your client's agent could talk to your firm's agent and clear the busywork overnight. That's right — for the friction. But today your advisory relationship is largely a byproduct of all that transactional contact, and automating it blindly severs the thing advisory grows from. Here's how to tell the contact you should automate from the contact that is the whole business.
Every system in your client's business is becoming something their AI can reach and act on — their bank, their payroll, their card processor. Every one except the firm that's supposed to be the most intelligent of the lot. Here's what it looks like when an accounting firm designs for the agent, not just the person — and why you start with the contact nobody enjoyed anyway.
A working agent is a one-off; a practice runs on systems. Three things have to happen for a single agent to become practice infrastructure — it persists across time, ports across clients, and composes with other agents into chains. This piece names the operational disciplines, the platform differences, and the manual chain that lets most CAS practices start today.
2026 is the year of the agent — and you're five minutes away from building one you could reuse every day. An agent isn't a new technology. It's the same AI you're already using with a better job description. The gap between a prompt, a workflow, and an agent is smaller than most CAS practitioners think.
