26 articles
Basis says it takes responsibility for the first pass of your clients' tax returns. OpenAI, four days earlier, refused to claim authorship of work its own model did and took responsibility only for its correctness. Both sentences are about who is answerable, and only one of them means it.
Karbon asked clients what should happen to your fee when AI makes the work faster, and 37% said it should come down, which is close to what the IRS told tax practitioners in June. Deliver something they have never had instead and 54% will pay more. Three of the four walls of the pricing vise moved in a single week.
Anthropic shipped two things this week that answer the same question in opposite ways: a feature that turns one screen recording into an editable procedure file, and a model that checks its own work brilliantly and remembers none of it. Meanwhile OpenAI launched a program teaching small-business owners to do their own accounting, and a top-15 firm committed its whole client accounting practice to a firm-trained AI model. The question running under all of it is which of these leaves anything behind.
Canopy, Drake, and Firm360 shipped AI tax-prep automation on the same day, the largest open-weight model ever released is days from being free to download, and OpenAI's CFO published an AI ROI scorecard with no line for verification. The pattern across the week: everything below the review layer is being absorbed or given away — and the review layer is what's left to charge for.
OpenAI and Anthropic just shipped the same product — an agentic workspace built to hold your firm's entire workflow — while Xero staged a keynote around an ecosystem you never leave. Meanwhile, Grok 4.5 beat Opus 4.8 on the vendor's own headline coding benchmark at $2 per million tokens: a price signal you can only capture if you're still free to switch. This week's roundup is about that tension.
Xero built AI document capture directly into the platform — and said nothing about HubDoc. A Top-30 firm handed its CAS delivery to AI agents most practices can't buy. And the most powerful model on the market came back from a 19-day ban with a meter on it. One question runs through all of it: how much of your stack do you actually control?
This week the IRS and Intuit both made a move on the same thing — the savings AI is creating inside your firm — and pulled in opposite directions. Plus Claude goes multiplayer in Slack, AI-faked receipts force a hard question about whose job it is to verify, and Washington quietly starts gating which models you can use.
This week in AI didn't bring a new model — it brought plumbing. Agentic tools that do the work (Suralink, Melio) and the governance layer that decides who connects them (Anthropic's managed MCP auth) all shipped, but none of them solved the problem that matters most to a multi-client firm: keeping one client's data out of another client's answer. The real divide forming in AI tools isn't who has the most connectors — it's who scopes them.
On Tuesday, Anthropic released the most capable AI model ever made public. By Friday, the US government had switched it off worldwide. Add OpenAI retiring models on its own schedule, Google force-defaulting Gemini, and QuickBooks killing classic reports — and the week's lesson is hard to miss: you don't own your AI supply chain. Here's how to build for it.
Ramp shipped an "AI operating system" built for accounting firms, Karbon and Firm360 opened MCP sockets to whatever AI you bring, Thrive put $1 billion on the table to buy firms like yours — and 69% of tax pros told a surveyor the hourly billing model is ending. Not one major model release all week. The fight has moved up the stack, to who holds your firm's encoded knowledge and who controls the door.
OpenAI shipped a working tax agent at 97% draft accuracy — but it took a custom-built product to get there, not a ChatGPT login. Kick opened its ledger to whichever AI you bring, and made it work across all your clients at once. Microsoft canceled Claude Code over runaway costs, a CFO ate a half-billion-dollar surprise bill, and Opus 4.8 got noticeably more honest. AI implementation is hard, can get expensive, and produces powerful results when done well — the firms thinking 12 to 18 months ahead are the ones who'll come out on top.
A quiet news week with a loud through-line. Intuit cut its coordination layer, KPMG put Claude in front of 276,000 people, and the ICAEW finally put numbers on the profession's AI blind spot — and all three say the same thing: the tool was never the hard part. The system around it is, and building it is leadership work.
