Anthropic published data on its own workforce this month. By August, its AI agents (software that works through the steps of a task on its own, the way a staff accountant works a checklist, rather than answering one question and stopping) were logging six hours of runtime for every hour of staff time, and juniors had fallen from 27% to 9% of its technical hires. The following week, Thomson Reuters and Wolters Kluwer launched agents that take tax work from source documents to finished workpapers, in Wolters Kluwer's case through to a prepared 1040. And while Mercury added payroll to the bank account and ledger its own agent already works in, Wells Fargo wrote into its customer agreement who answers when an agent does something nobody intended.
This week: what Anthropic's figures do and don't say about the entry rung, why the volume work your first-years learned on is moving into the tax software, and why one banking platform is building the agent into the back office while the big banks decide whether to let one in.
1. What Anthropic's numbers do and don't say about the entry rung
On October 16, Anthropic's economics team published "How employment is changing at Anthropic," a study of its own staff. Its headline ratio is that "By August interactive AI agent hours outnumbered employees' six to one," counting agent runtime against headcount at 40 hours a week.
The junior figures come next. The share of junior employees in Anthropic's technical departments "slipped from 21% to 14% since early 2025," and "the proportion of junior technical hires fell even faster, from 27% to 9%." The report says writing code and finding bugs could increasingly be handed to Claude, and an executive assistant told the authors that Claude had already taken over much of the entry-level admin work.
Those are shares, not headcounts. Anthropic tripled in size over the past year, and the report says technical hiring kept growing in absolute terms, so it may be hiring more juniors than before while hiring far more seniors. The authors don't claim the agents caused the shift: they warn that "tool adoption is not random" and that much of Anthropic's practice "cannot be replicated across the economy." And this is the company that sells the agents, describing itself.
What the report does measure is where people's time went. The median employee spends 20% more time in meetings than a year ago, and "meta-work" (building visibility tools and monitoring agent pipelines) rose from 10% to 22% of employees' Claude sessions. The authors write that "human attention is a constant bottleneck," and hiring managers now look for "autonomy and judgment, rather than execution and narrow expertise."
What follows is my argument, not Anthropic's finding. A CAS firm has always trained its people with volume: the 500 returns a tax first-year prepares, or the year of bank recs and month-end closes a new bookkeeper works through, each one reviewed by someone more senior. Jody Padar described that model for the Thomson Reuters Institute on October 16: "For decades, accounting firms developed talent through repetition," and "AI is breaking that progression."
Padar's point is that the volume was never the whole lesson. "Preparing 500 tax returns can teach someone how to prepare a tax return," but judgment comes from experience: real problems, client conversations, and decisions, with room to get things wrong while someone catches it, because "It's this failure that creates judgment." The returns gave juniors hundreds of chances to make a mistake under review. As the software takes the returns, it takes those chances too, so the training has to be designed on purpose rather than inherited from the workload.
My bet is that firms hiring over the next two years will lean the way Anthropic did, toward fewer juniors per senior and a hiring screen built around judgment. Agents didn't create the profession's pipeline problem, since firms were struggling to hire accounting graduates long before this, but they remove the work that turned the juniors you did hire into seniors. Our June piece on AI training argued that training has to happen inside real deliverables with a coach reviewing, and those deliverables now include the agent's.
In practice, have your next hire prepare a small set of returns or closes by hand, then compare the agent's version of the same clients against their own, with a senior reviewing both. They learn the mechanics from the first and judgment from the gaps between the two. Decide now what a first-year in your firm does in 2028, when the software has prepared the return before they open it.
2. The tax software is taking the volume work your first-years learned on
Thomson Reuters launched the next generation of CoCounsel Tax, Audit & Accounting on October 22, "available now for all customers in the United States." CoCounsel is a separate product that connects to UltraTax, and Thomson Reuters describes it as working "from source documents and client data to finished work product." It writes and runs code to check depreciation, amortization, required minimum distributions, qualified business income, threshold tests, and multi-scenario comparisons, then builds multi-tab Excel workpapers with live formulas and drafts memos to your firm's templates. It also supports MCP, the open standard that lets an AI assistant read from and write to another piece of software, much as a bank feed lets your ledger read the bank. Thomson Reuters gave no price and didn't say which AI models it runs on.
Thomson Reuters CTO Joel Hron: "We didn't just build a tool that does math. We built one that knows why the math is right." He also said, "The code execution piece is table stakes now; every company in this space has it." Hron is treating the calculation as solved and claiming the review step for the software as well.
Wolters Kluwer announced the next morning "the launch of agentic tax return preparation in CCH Axcess Tax." Its Expert AI agents identify the documents a return should have, classify and extract them, map the data into an individual 1040, link each amount to its evidence, surface review findings, and generate workpapers, with the preparer approving the return. The release says the capabilities "are built into the platform rather than separate products," but gives no date, no beta or general-availability label, and no price. The day before, Inscope's agentic financial statements went live inside CCH Axcess and CCH ProSystem fx Audit, drafting statements from the trial balance and running footing, cross-footing, and disclosure checks, with Inscope's own claim of a "70% faster" first draft.
In July's roundup I called it the AI tax-prep wars, with newer vendors selling the review layer. Now the incumbents are putting production inside platforms many US firms already run. Intake, extraction, computation, and workpapers are the work a first-year learned the job on.
Neither vendor published an accuracy figure for these releases, and Inscope's numbers are its own. Vals AI's tax agent benchmark, updated October 20, still has no AI model above 50% on its 193 tax tasks. A model is the underlying AI engine, and products like CoCounsel are built on top of one, so the benchmark doesn't score these products directly. It does mean someone at your firm still checks the output.
Software that checks its own math still can't tell you whether it was handed the right inputs. The provenance checks from September cover that gap: the right client, the right year, the right entity, and the right source document behind every number. Those checks, and the signature, stay with your firm whatever the vendor's software claims to verify.
If you're on CCH Axcess or UltraTax, ask your rep when these features reach your account, what they cost, and whether the record shows which numbers the agent produced. CCH Connections runs this week, so Wolters Kluwer firms may get the date and price there. Firms on other packages should ask their own vendor the same questions.
3. Mercury built the agent into the bank account. The big banks are still deciding whether to let one in.
Last week Mercury, a business banking platform that holds customer deposits at partner banks, let ChatGPT and Claude move money from its accounts once the customer approves. Now it's building its own agent into everything behind the account. Mercury Books, launched September 16, is "a full double-entry accounting engine" that Mercury pitches to QuickBooks users as "built to replace it, not sync with it," and on October 20 Mercury added payroll in all 50 states, free for 2026 and then $100 a month plus $50 per person paid.
Its agent, Command, works across both. Mercury says it handles "writing journal entries, organizing your chart of accounts, categorization in bulk" in Books, and for a salary change it "works out the proration, the updated withholdings, and the right pay run" and "then drafts the change for your approval." So one company now holds a client's bank account, ledger, and payroll, along with the agent that drafts the work, and the owner approves it.
Where your firm sits is less clear. Books gives accountants unlimited advisor seats at no extra cost, but the payroll announcement doesn't mention accountants or bookkeepers at all, and it doesn't say whether that seat will see a pay run before the owner approves it.
The large banks are moving the other way. JPMorgan said last week it won't let outside agents sign in with a customer's password, and Citi and Wells Fargo told USA Today on October 23 that they block unauthorized agents. American Express CEO Steve Squeri said on the October 23 earnings call, "we're only going to allow agents access to our systems that are either approved through API or MCP channels," an API being the published route one piece of software uses to talk to another.
Wells Fargo has also written down who answers once it does let an agent in. Its online access agreement, effective November 24, reads: "All transactions that an Authorized Representative or Delegate (including the transactions of their AI Agents) performs on an Eligible Account or Online Financial Service, including those you did not want or intend, are transactions authorized by you." "Unless prohibited by applicable law," the customer is "solely responsible for those transactions," and also "solely responsible for supervising" its delegates and their agents. The language applies only "if and when we permit the use of such agents," and the agreement never defines an AI agent, but it puts the bank's position in writing before a single agent is let in.
Mercury and the big banks disagree about whether an agent gets in. They agree on who answers for it: the owner, whom Mercury builds in as the approver and Wells Fargo writes in as the liable party. Neither puts your firm in the loop by default, so your seat is whatever the client grants you. And catching a wrong pay run or a miscoded entry before the owner clicks approve takes the judgment juniors used to build by doing that work by hand.
For clients who bank with Mercury, find out what your advisor seat can see in payroll. Where your firm has delegated access to a client's bank account, read that bank's agreement before you put an agent on their payables. Either way, the engagement letter is where you and the client settle whose agent it is, who reviews its work before the owner approves, and who answers for it.
Quick hits
Your client can join a waitlist for an assistant that checks their books. Anthropic opened a US waitlist for Claude Duet on October 21, an assistant that "checks your connected accounts every few hours," with QuickBooks, Gusto, Stripe, Gmail, and DocuSign among the apps it names. It shows its plan and waits for approval before it sends a message, makes a payment, signs a document, or deletes anything, and it refuses to trade securities or move money between accounts even with approval. It's for Pro and Max subscribers only (Team and Enterprise "are not included yet"), so your client on a $20 plan can join the list while a firm on a Team or Enterprise plan can't. Anthropic's one QuickBooks example is checking for unpaid invoices and drafting reminders, and it hasn't said what else the connection can change.
Dext's AI connector is live, and I tested it. Dext launched its AI Connector on October 22 for practice accounts, with Practice Advanced partners able to request access; it isn't available for business accounts yet. It lets you query your Dext data from Claude, ChatGPT, or Copilot, and it's read-only: Dext's CPTO, Stephen Edginton, says it's "built to retrieve, not to decide." I connected Claude to my practice's account, and within three minutes it had built a live dashboard of the inbox backlog across all 67 active clients, flagging items over 90 days old and clients who'd stopped uploading. The limit is that Dext supports it only through those three assistants, so you can't yet plug it into other MCP-enabled tools, such as Ramp's Stack or an app you've built yourself; ask Dext when that changes.
The profession built an assurance engagement for AI agents. The AICPA and AIUC announced on October 19 a single CPA-led engagement that produces both a SOC 2 report and an AIUC-1 certification, with live testing of agent behavior across "hundreds of risk and attack combinations" and quarterly retesting. Several CPA firms are delivering initial engagements (Accounting Today names PwC), with a full launch in early 2027. It's addressed to AI developers and enterprises deploying agents. For a CAS firm the use is as a buyer: ask the vendors whose agents can write in your clients' ledgers whether they plan to get one.
Patch your website, and find your API keys. PHP, the programming language behind WordPress and a large share of websites, disclosed two critical flaws on October 21 (rated 9.2 out of 10) and patched them, with the PHP Foundation's warning: "If you run PHP, it is imperative that you update now." If your firm's site or client portal runs on WordPress or another PHP platform, ask your host whether the October 21 patch is applied. In a separate breach, OpenCode, a coding tool, said stored OpenAI and Anthropic API keys may have been exposed, and users reported charges they didn't make. An API key works like a password that spends money, so know where your firm's keys are stored and set a spending cap wherever the provider offers one.
The newest models in Office will be metered. Microsoft said GPT-6 Astra and Claude Fable 5.1 will be the first models in Copilot Chat, Word, Excel, and PowerPoint that are charged by use rather than included in the Copilot license, generally available in early to mid November. Your staff won't see them unless an admin turns on usage-based billing, which draws down credits your firm buys. Decide whether you'll turn it on, and with what cap, before staff ask. Separately, Gmail's AI Inbox, which summarizes "priority work updates across your emails" and surfaces action items and deadlines, reaches Google Workspace accounts on the Scheduled Release track from October 26 (US, English only) with "no admin control for this feature," so tell your team what they're seeing before they ask.
What this week adds up to
The agents are taking the work that taught the job. The work left is reviewing what they produce and answering for it. Anthropic now hires for that, Thomson Reuters wants its software to do part of it, and Wells Fargo has put the liability for it in a contract. My bet is that the firms in good shape in two years will be the ones that designed how a junior learns judgment once the volume work is gone. Who in your firm is learning to review the agent's work, and on what?
If your AI Champion is the person who'll redesign how your firm trains its next hire, the next Practice Transformation Program cohort starts Wednesday, November 11. It runs four live three-hour sessions and nine modules, and ends with a transformation plan your firm actually executes. US$2,997. Enrollment closes Sunday, November 1: theaiaccountant.ai/transformation.