Xero announced a team of engineers this week whose job is to go into accounting firms and build their workflows on Xero’s platform. It’s in beta with five firms in Australia, New Zealand, and the UK, and it has no price yet. In the same week, Anthropic and Google moved more of a firm’s daily work inside their AI apps, Microsoft was due to widen access to its accounting skills in Excel, and Harvard Business Review published the first field study I’ve seen from inside tax firms using AI.
Below: those stories and what to do about each, plus ChatGPT’s new business listings, then four quick hits, including agentic tax prep coming to CCH iFirm for Canadian firms.
1. Xero will build your workflows for you, and nobody has said who owns them
On November 12 Xero announced Xforce, which it describes as “a team of forward-deployed engineers (FDEs) and partner consultants” who work with accounting firms “to design their workflows for the AI era, on a trusted platform they control.” A forward-deployed engineer is a vendor’s engineer who works inside your business and builds the system around how you actually work. Xforce is in beta with “a handful of design partners in Australia, New Zealand and the UK”: Liston Newton Advisory, Nexia Christchurch, Shepherd Hensman, Azets, and Sumer. There’s no US or Canadian partner, no price, and no date, just a form to register your interest.
In June I wrote that you can’t build Thrive’s tax agent, because it came out of six months of OpenAI’s forward-deployed engineers working inside a roll-up of more than 30 CPA firms, and no firm can buy that at any price. Now a ledger vendor is offering firms a version of that model, on its own platform.
Xforce sits beside XeroForce, Xero’s do-it-yourself tool for building agents (AI that carries out a sequence of work on its own) by describing what you want in plain language, which Xero’s results announcement describes as “early in the alpha,” meaning early testing with a small group. I read Xforce as Xero concluding that its AI products won’t land in a firm without Xero’s people setting them up; Xero hasn’t said so. That’s the same setup work many CAS firms already sell to their own clients, now offered by the platform.
No Xero page says who owns the workflows its engineers design on Xero’s platform, or whether they leave with you. The release says “firms decide which clients and data the workflows touch,” which covers access, not ownership. In July I asked whether you could take your instructions with you, and pointed out that agents built inside a client’s Xero organization leave with the client; Xforce adds a third party to that question, because the procedures are yours, the engineer is Xero’s, and so is the platform.
The money side is still a sketch. Xero says about a million organizations used its generative AI features over the past year, and its investor presentation lists three ways to make money from AI: bundled features and add-ons, Xforce itself, and consumption, described as “Credits, API usage, workflows.” The API is the connection your own AI tools use to read and write Xero data, and Xero says it’s “establishing the foundations for future usage-based pricing,” but no rate or date appears anywhere in its results.
If your clients are mostly on QuickBooks, this still applies to you. In September Intuit showed investors an agent that “runs the close end to end following the firm’s own SOPs and policies,” so both ledgers now want your procedures inside their platform. If you serve clients on both, you face a harder choice: one set of procedures held by Xero and another held by Intuit, maintained twice and drifting apart. I think many mixed-ledger firms will prefer to keep their procedures somewhere neither ledger controls, whether that’s a separate agent-management platform or general AI tools such as Claude, ChatGPT, or Gemini.
Before any vendor builds workflows for you, get three answers in writing: who owns what gets built, whether you can export it in a form another tool can read, and what happens to it if you leave the platform or the client leaves you. And keep the master copy of every procedure in your own files, whoever builds the automation on top of it.
2. The AI apps are becoming the office, and they’ll hold your procedures too
The general AI tools want your procedures too. Anthropic and Google moved more of a firm’s daily work into their AI apps this week, and Claude, Google’s Gemini, and Microsoft’s Copilot can each now follow procedures your firm writes down as skills, or soon will. A skill is a written procedure the AI reads and follows whenever that task comes up, much like the desk instructions you’d leave a new hire for the month-end close.
Start with the meeting. Claude Meetings, in beta on Claude’s Team and Enterprise plans, joins Google Meet calls as a visible participant and “saves a transcript, summary, and screen descriptions, not audio.” By default it stays out of any call that includes someone from outside your organization, and leaves if one joins partway through; an admin can change that, and then “what guests say is included in the notes and transcript.” Anthropic tells you to “make sure everyone on your calls is OK with being transcribed,” but neither of its pages says how long transcripts are kept or whether they’re used to train models.
Claude Sheets also arrived in beta on every paid plan; it exports to .xlsx when you need Excel and doesn’t connect to live data. On Team plans, Meetings and Sheets are on by default. And on December 1, Claude’s chat and Cowork (the mode where Claude works through files and tasks for you) merge into one conversation for Team plans unless an owner keeps it off, and once your account switches over you can’t switch back.
Microsoft is putting accounting procedures straight into Excel. Its plan, posted November 3, was to open its accounting skills to Microsoft 365 Copilot users from November 13 without the early-access sign-up, in Copilot in Excel and in Microsoft’s own Cowork tool (no relation to Claude’s). Its support page lists nine skills, covering reconciliation, journal entries, close management, financial statements, variance analysis, budgets and forecasts, tax preparation, and audit and compliance support, and the plan says “Nothing is posted automatically to an ERP or system of record without approval,” with audit trails “currently available in Excel” and “planned for Cowork.” As of this weekend, that support page still says the skills are “rolling out to the Microsoft Frontier program,” its early-access group, and no Microsoft page confirms the wider release has started.
Microsoft’s skill-builder creates custom skills “from your instructions, previous prompts, or examples of completed work” and saves them to OneDrive, though Microsoft says it’s still rolling out and hasn’t said what file format it uses. Google added Skills to Workspace on November 10, “reusable, custom instructions that guide Gemini to apply your team’s rules, templates, and reference files,” in Gmail, Docs, Slides, Drive, and Chat, switched on by default and written in the same SKILL.md file format Claude uses.
Claude, Microsoft, and Google are third-party vendors too, so moving your procedures into one of their apps doesn’t settle the ownership question from story one. The difference is narrower: with Claude and Google, a skill can be a plain text file your firm writes and keeps, in a format both now read. The same file won’t behave identically in each, because each AI has different tools and reaches different data, but the procedure stays a document you hold.
Before December 1, find out which defaults on your firm’s Claude and Google plans someone chose and which simply arrived: meeting transcription, outside guests on calls, Sheets, the merged Claude mode, and Google’s Skills. Then pick one recurring procedure, such as your bank reconciliation review, write it as a skill file, and try the same file in more than one tool.
3. At one tax firm, the billable-hour target worked against the AI tool
On November 11 Harvard Business Review published a study from inside four accounting firms using AI through Current, a platform that provides technology and shared services to partner firms. Across roughly 7,000 moderate-complexity individual and trust returns in the 2025–26 season, AI tools “saved tax practitioners an average of a third of their time” on preparation and lifted throughput by about 50%, and one firm took on about $300,000 of extra work late in the season.
At one firm, “billable hours worked had been an important measure of performance, so a tool that reduced the amount of time to prepare a return was seen as working against the interests of the people expected to use it.” If your staff are judged on hours, a tool that removes hours threatens their performance review, whatever it does for the firm.
The study’s other finding is one I’ve been arguing for weeks: review doesn’t shrink the way preparation does. “AI could produce a tax return in less time than it took for a reviewer to verify it,” so review time fell only modestly on some returns and not at all on others, and errors “tended to recur around particular document types or edge cases.” In the authors’ words, “the productivity gain moved the bottleneck rather than eliminating it.” One firm also redesigned its intern program so juniors prepared practice returns without AI.
Read the numbers with the authors’ own caveats. The gains are company-reported, it’s four firms on one platform in one season, and lead author Nitin Nohria chairs Thrive Capital, whose subsidiary Thrive Holdings is an owner of Current. That’s the same Thrive Holdings behind the Crete tax agent I wrote about in June, so treat this as evidence from an interested party; the authors disclose all of it.
For this tax season, change what you measure to returns completed and accepted without rework, so the tool’s savings count in the preparer’s favor, and log review corrections by document type, such as K-1s and rental schedules, so the review time the AI didn’t remove goes where its mistakes keep recurring.
4. ChatGPT now lets you claim your firm’s listing, but check your legal name first
On November 10 OpenAI opened Businesses in ChatGPT in beta, which means you can now claim what ChatGPT shows about your firm. A business owner or someone authorized to represent the business can claim a listing, edit its details, and see “aggregate activity over the past 28 days, including mentions and website clicks.” You don’t see individual prompts or conversations.
Verification is the step to prepare for. After claiming by phone, business email, website domain, or Shopify, you verify through Persona, an identity-checking service, with “a government-issued ID, business registration information, and proof that you’re authorized,” and OpenAI warns that your business name, address, and “any registration or tax number” must “match your official business records.” Jason Staats found his legal entity name didn’t match his trading name and got stuck; his workaround is to claim the existing listing, then add a second business under the legal name.
Business Agents, which answer customers inside ChatGPT, are limited to “eligible ecommerce and online service businesses in the U.S.,” so there’s no agent for an accounting firm yet. OpenAI says “Claiming does not guarantee higher placement or sales,” and it doesn’t state a price. Claiming isn’t limited to the US: OpenAI’s help article says eligible businesses can claim an existing listing “or add one where ChatGPT is available and the required verification is supported,” which includes Canada. The US-only limit is on adding a new listing for local stores and restaurants that can’t find themselves, and the article doesn’t say whether an accounting firm counts as one.
Read the terms before you upload anything beyond the basics. They give OpenAI a license to use what you submit for “OpenAI’s and its Affiliates’ development, provision, and improvement of their products or services,” and uploading a product feed means giving up any text-and-data-mining opt-out against OpenAI. Your firm’s description and contact details are one decision; a library of service write-ups is a different one.
Claim your listing this week, with your legal entity name, registration number, and ID ready, and look at what ChatGPT currently says about your firm before you edit anything. If ChatGPT doesn’t show a listing for your firm, look for the option to add one; if it isn’t offered, your firm is probably being treated as a local business and will have to wait.
Quick hits
An IRS official eased off June’s fee line, and the SEC’s chief accountant said accountability can’t be outsourced. Thomas Curtin, director of the IRS Office of Professional Responsibility, told the AICPA National Tax Conference on November 9 that his office is “not suggesting that practitioners must reduce fees simply because AI allows you to complete certain tasks more effectively,” and called June’s AI alert “just guidelines.” He spoke in his personal capacity, the alert itself hasn’t changed, and the Journal of Accountancy’s report is the only record; it also reports him saying AI isn’t a “person” under Circular 230, so you can’t rely on its work the way you’d rely on a colleague’s. The same week, the SEC’s enforcement results announced an Office of Super Intelligence and Analytics (“Super Intelligence” is the term a White House executive order now has federal agencies use for AI), the PCAOB created a Super Intelligence Unit to focus on potential misconduct “associated with the use of Super Intelligence in public company audits without discouraging the appropriate use,” and the SEC’s chief accountant, Kurt Hohl, told CFOs “you can outsource a process, but you really cannot outsource the accountability.”
Agentic tax prep is coming to CCH iFirm for Canadian firms on December 18. Wolters Kluwer announced “agentic tax preparation capabilities” for CCH iFirm on November 5, but its product page shows they aren’t available yet: “Commercial availability for T1 is scheduled for December 18, 2026.” There are two add-ons for eligible T1 customers. Extract & Organize uses “Expert AI agents” to classify uploaded documents, pull out the figures, and set them out for review before they go into the return; Questionnaires, which needs the separate CCH iFirm Portal product, sends clients tailored, pre-filled questionnaires, though the page doesn’t make clear how much of that is AI. Wolters Kluwer claims “up to 20 minutes per return” and calls these “Canada’s first” such tools, with no add-on price published. My own practice runs on CCH iFirm Taxprep and hasn’t had any notice of it, so if you’re a T1 user, ask your rep what “eligible” means before tax season.
Anthropic cut Fable’s price, and a tax benchmark ranked it 28th. Claude Fable 5.2 arrived on November 9 at $8 per million input tokens and $40 per million output tokens (tokens are the units AI usage is billed in, roughly three-quarters of a word each), down from $10 and $50. Two days later Vals’ tax-agent benchmark, a US corporate tax research test, put Google’s Gemini 4 Hydrogen first at 52.30% for $3.87 a test, Fable 5.1 third at 50.75% for $13.18, and Fable 5.2 28th at 32.09%. The same Fable 5.2 ranks first on Vals’ finance-agent and Excel-modeling benchmarks, and Vals gives no reason for the tax result. If you’re choosing a model for tax research, run it on your own files first.
Notion’s AI vendor was breached, and some of the exposed chats held pasted keys. Notion says a breach at Fireworks, a company it used to run some AI models, exposed Notion AI chat transcripts from “less than 1% of Notion workspaces,” “some containing pasted API keys” (the passwords one piece of software uses to log in to another), and it told affected customers to “treat this data as if it were exposed.” Affected customers chose Notion, and Notion chose Fireworks: an AI feature can send your data to a vendor you never evaluated. If anyone on your team has pasted passwords or API keys into an AI chat, rotate them, and add “no credentials in AI chats” to your AI policy.
What this week adds up to
The vendors in this week’s lead stories all want a copy of how your firm works, from Xero’s engineers to the skills in Claude, Copilot, and Gemini. The officials who spoke this week pointed the same way on responsibility: it stays with people, whoever built the workflow. So before anyone else writes your procedures for you, where’s the master copy kept, and could you take it to another tool tomorrow?
If you want to know where your firm stands before any vendor starts writing its procedures for it, take the free AI Readiness Scorecard at theaiaccountant.ai/scorecard. It’s 25 questions and about five minutes, it scores your practice on five areas including data and context readiness, and your report card comes with an action plan for your level.

