On Thursday, September 17, on page 96 of Intuit’s 168-page Investor Day deck, a slide labeled CASE STUDY described an agent (software that carries out a sequence of steps on its own, rather than answering a question about them) that “runs the close end to end following the firm’s own SOPs and policies,” in open beta at no charge, with paid availability “expected in January 2027.” The day before, Mercury, the business-banking fintech widely used by startups, launched a general ledger, waived its price through December, and gave your firm unlimited free seats in it. On Tuesday, Anthropic’s month-end close stopped requiring QuickBooks, Xero said Claude can now “prepare a first pass at the work” in its books, and Google switched on a QuickBooks connector (the plug that lets an AI assistant reach the data inside a piece of software) by default in Gemini, its own assistant. A loss leader is the thing you give away to sell the thing you don’t. This week the close, and the ledger it runs on, were offered three times over with no price of their own, and what the platforms want in exchange is the subject of this roundup.
1. Intuit’s agent will run your close from your SOPs, at no charge, until January
Intuit’s Investor Day on Thursday reaffirmed the guidance it gave on August 25, so the headline numbers were a restatement. The news sat on slide 96, under a header that reads “Intuit Intelligence automates month-end close across clients, reducing manual work and driving platform preference.” The product is Agentic Books Close, and its body text is one sentence: “Agent runs the close end to end following the firm’s own SOPs and policies.” The rail beside it reads “Firm remains in control,” “Runs the firm’s playbook,” and “Learns with every close.” The footnote carries the only date in the story: “available in open beta at no charge; paid availability is expected in January 2027.”
Alongside it sits Agent Studio, “Agents built once and deployed across clients,” which every one of its four slides footnotes as “currently in private pilot with select IAS and IES customers. Additional fees may apply. Screen images are simulated.” So Agent Studio is a pilot shown with simulated screens, and Agentic Books Close is a product with a date and no price. The missing price is worth sitting with, because Intuit chose to withhold it.
The declaration this product answers to is three months old. On June 23, at Intuit Connect ON, Sasan Goodarzi told accountants “Our first decision is that you are the customer, not a channel… That changes now,” and I covered it in the June 29 roundup and a July 1 standalone arguing that Intuit was capturing the AI value as margin while saying warm things to firms. On August 12, the Accountant Suite update gave firms on the paid Accelerate tier the build surface, letting them “build their own AI automations from a prompt, a firm SOP, or a ready-made library, then run them across clients” on metered monthly capacity, which Buy vs Build Part 4 covered as the platform selling firms the tools to build. Slide 96 is the step after that: in August the firm wrote the automation from its own SOP, and now Intuit’s agent runs the close from it. The shape is exact: the firm contributes the playbook, and the platform runs the close.
The same deck is equally candid about its other strategy, and reporting one without the other reports half the deck. “Done-for-you” appears 12 times, slide 71 is titled “The business runs done-for-you,” with agentic accounting “automated from day one,” and slide 82 is titled “New front doors, inside AI platforms.” Intuit sizes its market at 10 million businesses and 650,000 accountants, and states the accountant’s problem as “grow and expand advisory services.” Intuit wants to run your client’s close from your playbook and to run the client’s business from day one, and both ambitions are on the slides.
So the question worth asking is the one Buy vs Build Part 4 asked in the abstract, now with a date on it: once your SOPs are what you contributed and Intuit’s agent is what runs them, what did you keep? The deck says nothing about whether a firm can export the SOPs it contributes, or what happens to them if it leaves. That silence matters, because an encoded process (your firm’s method, written down in a form software can follow) is an asset if you can take it to another platform and a dependency with a friendly onboarding experience if you can’t. For Xero I answered that question in May, when XeroForce launched as a builder whose agents live inside Xero: they don’t port. For Intuit it’s open, and the deck doesn’t close it.
The corporate-close market made the mirror-image offer on Wednesday. FloQast, which sells to controllers and internal audit teams rather than to firms, announced Transform: “Hand FloQast files from a close your team has already finished, and Transform builds an AI agent that runs that process for you: documented, tested against numbers your team already approved, and ready to show your auditor. No IT or coding required.” Intuit will run the playbook you wrote inside its platform, and FloQast will write the playbook from work you already did. Both take custody of the encoding work I’ve spent two years telling firms to keep for themselves, and neither the deck nor the release says whether you can take the result with you.
Set the two ledger platforms side by side and they’ve shipped mirror images, four months apart. Xero launched the builder first, XeroForce in May, with the promise “Build an agent once in plain language, connect Xero to the apps you already use, and it runs across your entire client base,” and Intuit’s Agent Studio slide now says “Agents built once and deployed across clients,” which is the same sentence in a different font. Xero announced its vendor-built month-end agent at Xerocon Denver in August, “reviewing document reconciliation status, right through to manual journal entries for prepayments and amortization, before providing customers with a complete account of its actions to review and accept,” and as of Saturday Xero’s own blog still labels it “[Coming soon],” with XeroForce in early access behind a waitlist and general availability “later this year.” So Intuit is behind on the builder and ahead on the date: open beta now against a waitlist, and a month named for charging against “later this year.” The difference that matters more is in the copy: Xero told you what comes back, a complete account of the agent’s actions for you to review and accept, while Intuit told you what goes in, the firm’s own SOPs and policies, and neither vendor has yet described both ends.
What this means for you. If you’re considering the open beta, or a client’s platform is about to do this regardless, first write down what you’d need to walk out with: your SOPs as documents you hold, the exceptions log, and the correction record with your team’s names on it. Second, ask Intuit the export question in writing before you contribute anything, because the answer isn’t in the 168 pages. The firms most exposed are the ones whose whole pitch is “we have a process,” because that process is the asset Intuit has offered to operate at no charge until it expects to start charging in January.
2. A client’s banking platform just offered to be their ledger, and gave you a free seat
On Wednesday, September 16, Mercury launched Mercury Books, “a full double-entry accounting engine, supporting accrual or cash basis, built and tested alongside working accountants and bookkeepers.” Its FAQ asks “Does Mercury Books replace QuickBooks?” and answers “Yes. Mercury Books is a complete accounting system,” adding, for anyone currently on QuickBooks, that “Books is built to replace it, not sync with it.” It covers journal entries, the chart of accounts, bulk recategorization, and the three statements, driven from the money as it moves, with Stripe, PayPal, and Gusto connected alongside. The price is $35 a month, waived through December 31, 2026. One sentence deserves a second read: “Invite your bookkeeper in-product, with unlimited seats at no extra cost.”
Mercury describes itself in its own footer as “a fintech company, not an FDIC-insured bank,” with deposits held at Choice Financial Group and Column N.A. That makes it a weaker claim about deposits and exactly as strong a claim about data, because Mercury owns the interface to the money, and the cash-side transactions a ledger starts from are already in it. I argued in July that any layer of your stack the platform can see, it will eventually absorb, and Blake Oliver said the specific version of it on The Accounting Podcast in March 2022, when Found raised $60 million: “banks are becoming accounting software, and accounting software is becoming banks… You build a bank with an API from scratch, and then you build your own software on top of it… you don’t need a separate accounting system.” Four and a half years later, this is absorption arriving from a direction the trade press wasn’t watching (as of Saturday, no accounting outlet had covered it), and Ramp launched in the UK the same week with “accounting automation” that syncs to Xero and QuickBooks. Mercury’s roadmap adds that “an MCP layer will let AI agents handle bookkeeping tasks end-to-end”; MCP is the standard behind connectors like the ones in Story 3, and it lets an assistant act inside software as well as read from it.
Why is the advisor seat free and uncapped when the business pays? Mercury doesn’t say, so this is my inference: the business is the customer, and the advisor is a feature that makes the business stickier. Mercury’s own line is “A tool your accountant will trust.” On either reading, the software is being sold to your client and given to you.
The limits are real, and you should hold them. Books is only available to Mercury account holders, and Mercury is a platform widely used by startups and venture-backed companies. Across the blog, the product page, the FAQ, and the release note there’s no mention of tax, multi-entity, inventory, fixed assets, class tracking, AR or AP aging, period locking, or an audit trail, and for a product pitched as a QuickBooks replacement there isn’t a word about migration: no conversion tool, no historical import, no opening balances, nothing about what happens to the ledger if the account closes. “Co-designed alongside accountants from leading firms” names one, “Chris P., Founder of RidgeWood CPA,” and the claim that his team cut manual bookkeeping time in half is Mercury’s narration; his only quoted words are about having “more time to be that sounding board.”
Nobody outside Mercury’s pre-launch partners has used this, so nothing here supports a verdict on quality either way. The answerable question is the one that will arrive in a client meeting before the free period ends: when your client’s banking platform offers to be their ledger for nothing, what do you say? “Our software is better” invites a comparison nobody has evidence for. The documented gaps are a stronger answer, because you can walk them against a specific client instead of arguing about adequacy in general.
What this means for you. If you have startup clients banking with Mercury, assume the question is coming before December 31 and prepare two things. First, the gap list against each of those clients: inventory, sales tax, multi-entity, fixed assets, a locked period, an audit trail. Second, a scoped and priced migration engagement for the ones who will move anyway, because Mercury has described no migration path at all, which turns the move into a paid engagement. The durable answer is in Mercury’s own footnote, “Bookkeeping services are not included with Mercury Books”: the work and the name on the file were never what was being given away.
3. Claude’s month-end close stopped requiring QuickBooks. If your book is Xero, this is now about your clients
Anthropic’s Tuesday post reads like a launch, so start with what it isn’t. It describes a workflow that “reconciles your books against every place money moves: sales settlements, card spend, and payroll,” flags mismatches down to the transaction, writes a plain-English P&L, and “creates a close packet for your accountant.” That workflow shipped on May 13, and in May the skill (a saved set of instructions that Claude, Anthropic’s assistant, follows) said in its own text that if QuickBooks was unreachable it should stop, because “reconciliation requires QB as the source of truth.” As of Tuesday’s update it treats MYOB, NetSuite, QuickBooks, Xero, and Zoho Books as peers, and brings Gusto, Ramp, and Shopify in alongside Stripe and Square. The close is four months old; the ledgers it reaches are new.
Xero moved the same day and drew the line itself. Its May connector was question-and-answer only: cash position, overdue invoices, how profit is tracking. Its September post, by chief product and technology officer Diya Jolly, separates the two: the May connector “gives small businesses and their advisors answers to financial questions,” while the new integration “helps them prepare a first pass at the work behind those questions.” Xero lists six workflows it “will support,” from business pulse and invoice chasing to AP preparation, and describes three of them, month-end preparation, recurring reports, and a cash flow snapshot, as already “a first pass to review rather than a blank page,” usable by advisors today one client organization at a time. Google, with no fanfare, added a QuickBooks connector to Gemini, “ON by default” for anyone with Gemini for Google Workspace access, though the connector reaches nothing until a user signs in to QuickBooks through it.
Two of the three stop at the same line, and you sit on it. Claude assembles the pack and hands it over “for your accountant.” Xero’s connector can’t write, in Xero’s words: “Claude has no write actions – your books stay exactly as they are,” and the FAQ answer to “Can Claude make changes in my Xero account?” is “No.” Xero’s own body text puts the boundary where I’d put it: “the review step and final judgement stay with the accountant and the business owner.” Google’s release describes access and says nothing about writes either way.
The firm is the operator in Anthropic’s own copy. Its featured close-month story is HireEffect, a Dallas firm running bookkeeping, payroll, and HR for small businesses, whose director of bookkeeping “built her own skill” and reports isolating 16 transactions behind a $13,000 discrepancy. The post also says what happens to her corrections: “the adjustment your bookkeeper makes by hand becomes a saved step in next month’s close,” which is the encoding loop I’ve been describing for two years, shipped as a product feature. Two conditions apply: Anthropic says it runs in the desktop app on a paid plan, and its 11 named partners and more than 150 approved trainers include no accounting firm or professional body, so firms appear in it as users while the selling runs direct to owners.
The part being absorbed is the assembly. Chasing statements, matching settlements, spotting mismatches, and building the pack is moving to software your client can run without you, on QuickBooks since May and on Xero, MYOB, NetSuite, and Zoho Books since Tuesday, and what comes back is a first draft that needs someone to check it, correct it, and stand behind it. A practice still pricing the gathering is pricing work that has moved.
What this means for you. If your book is Xero, the question I asked in The Agentic Firm Part 1, whether your client’s agent can reach their bank but not you, now applies to your clients. Run one client’s month-end preparation through it yourself this month, before a client does, and log what the first pass gets wrong; that log is both your review procedure and your fee. If your clients are on Google Workspace, tell their admin the QuickBooks connector is on by default, so that the firm decides whether the books get linked to Gmail and Sheets before a staff member does it on their own.
4. A model that prices a decision at a fraction of a cent, and most of your checking layer is decisions
TypeSafe came out of two years of stealth on Tuesday with Jev, a model (the AI engine underneath an assistant) that doesn’t generate text at all. You hand it a messy input and a list of up to 255 options you defined in advance, and it returns a probability across those options plus a calibrated confidence score, calibrated meaning that a 90% is supposed to be right about nine times in ten. The list price is $0.042 per million tokens in (tokens are the units AI vendors bill by, roughly three-quarters of a word each) and nothing for output, at 70 to 500 milliseconds a call. One independent tester, Mike Taylor at Every, ran 21 questions across 37 documents, 777 judgments in all, in under a second for an estimated quarter of a cent.
Look at what a practice does at the checking layer and most of it has that shape: a bounded decision under uncertainty. Is this the same vendor as last month’s, is this transaction coded like the previous 40, does this invoice match the purchase order, does this receipt support the deduction it’s claimed under, is this a related party, does this one need a human. Sampling exists because looking at every item cost more than the risk on most of them, and “we review a sample” has always been an economic statement dressed as a methodology. At a quarter of a cent per several hundred judgments the economics invert, and the correction record I keep arguing is the sellable artifact gets cheaper to produce and more complete when produced.
That doesn’t work on its own, and the qualification is the story’s real content. Jev makes the call and nothing else, so it would have to sit inside a multi-engine system: a language model to read the documents, route the work, and draft what goes to the client, and a decision model to make the calls at volume. Someone still has to define the 255 options, and defining them is where the professional knowledge lives. I haven’t seen that system built for accounting, and until someone builds it the economics are a forecast.
Three limits, and TypeSafe concedes the first two itself. First, the “can’t hallucinate” claim is about the shape of the answer and not its truth; the note on TypeSafe’s own chart reads “Our number is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots.” Second, its benchmarks “use the average of GPT-6 Astra and Fable 5.1 as the reference answer” (OpenAI’s and Anthropic’s frontier models), so a high score means agreeing with the big models, including where they’re wrong. Third, it’s early access behind a waitlist with no accounting deployment of any kind, and TypeSafe’s own capability claim is parity with existing models on this class of task.
What this means for you. Nothing to buy, and one thing to do. Take one recurring engagement, a monthly close, a quarterly review, or an annual return, and list every yes-or-no decision your reviewer makes in it, along with the options they’re choosing between. That list is what a decision model would price, and it’s the list your review procedure should already contain; if you can’t write it down, the constraint on checking everything was never the price of the check.
Quick hits
Last week’s title was “Nobody said hours.” On Friday, Inside Public Accounting measured it. Across the IPA 100, the 100 largest US firms, IPA reports that “Only 54.2% of IPA 100 revenue is now generated through the traditional charge-hours-times-billing-rates model, down from 56.5% last year,” that 41.4% of revenue comes from non-compliance services, and firms with more than 60% of revenue outside compliance produced $152 of revenue per work hour against $130 overall. Three limits: IPA disclaims causation in its own copy (“does not establish that advisory work itself causes higher revenue per hour”), the article never mentions AI, and the IPA 100’s “non-compliance” bucket includes transaction advisory and wealth management, so these aren’t the firms you bid against. Read it as air cover. The partner who says clients expect to be billed by the hour is now arguing with the profession’s own benchmarking body, and the number to watch next year is the 54.2%, because one decline isn’t yet a trend.
One company took two years of AI savings into EBITDA early. One firm’s revenue fell and its partners were paid more. Bloomberg reported Thursday that Emeria, a Partners Group-backed French property-services group, added €20 million of “estimated efficiency gains” for the next two years from an agentic AI program to its last-twelve-month EBITDA in a lender presentation; at least ten investors, lawyers, and analysts said they’d never seen future AI benefits taken as an adjustment, and Emeria says the adjustments are “fully aligned with the methodology agreed with the company’s lenders.” That reply shrinks a scandal into a familiar run-rate add-back with a new label, and into a set of questions worth a fee: realized or projected, over what window, against what baseline, net of the deployment cost, and agreed where, in the credit agreement or on a slide. A projected saving has no cheap downstream detector, which is exactly where your review minutes belong. The same week, PwC UK’s revenue fell for the first time in 17 years while partner pay rose, PwC is roughly halving bonuses in Switzerland citing AI among its reasons, and KPMG cut around 200 UK advisory staff citing attrition and market conditions; only the Swiss item mentions AI at all. When delivery cost falls, the saving has three destinations, the client’s price, the staff’s pay, or the owner’s draw, and PwC UK is the only one of the three where anyone could see which way the money moved: one firm, one year, no AI attribution in the reporting, and it moved to the owners.
OpenAI published a model deciding to make numbers up. It happened in training, not in anyone’s client file. On Wednesday, OpenAI released six reports on misbehavior it found while training internal models, none of it in a shipped product, and it says in terms that the reports “shouldn’t be considered reflective of how often misalignment occurs across our models.” Read them anyway, because one is the mechanism I described in Right Answer, Wrong Data, with the vendor’s name on it: assigned to retrieve earnings data for a California county, a model failed, wrote “Guess numbers. provide fabricated plausible.” in its own visible working, produced nine figures, and delivered them as if transcribed from the source without disclosing any of it. A second, from an agent preparing a financial model, wrote into its own handover notes “Be transparent only if asked.” The argument that a tick box is unfalsifiable and a cell reference is checkable in four seconds now has a vendor-published example of a model presenting figures it invented as transcribed from a source.
Deloitte measured shadow AI on 25,000 UK workers, and the denominators matter. 63% of UK working adults say they knowingly use generative AI for work; 31% of those users say they do so without their employer’s knowledge; 17% of users pay for at least one tool themselves, which Deloitte grosses to nearly £1 billion a year, and around half of users have had no formal training. In financial services, where employer-provided tools are more common than the national average, self-funded use is no lower, at 19% of users. It’s UK, all industries, and not accounting firms, so the number for your firm and for your clients is unmeasured, and the useful move is to ask it. You can’t stand behind a process whose inputs you can’t see, and that applies to your own team before it applies to anyone you bill.
Salesforce trained its own model and put an accounting firm in the pilot. Koa is NVIDIA’s openly downloadable Nemotron 3 Super model, further trained by Salesforce on synthetic data built from 27 years of its own knowledge of how CRM work gets done, with no customer records used, and kept inside Salesforce’s own walls; it’s pilot-only, with general release “expected winter 2026.” 1-800Accountant is a named pilot customer, and its Ryan Teeples put the accounting problem in one sentence: “Accounting requires navigating tax rules, financial data, documents, and the unique circumstances of every customer.” Buy vs Build Part 5 argued on Friday that intelligence splits into rented base capability and ownable adaptation built on your own data, and Koa is that split executed in public three days earlier. No CAS firm should train a model; the firm-sized version of the layer Salesforce chose to own is your declared client context, your engagement instructions, and your correction record.
ChatGPT for Word switches on by default for Business and Enterprise workspaces on October 1, if your Microsoft 365 admin already allowed the add-in. ChatGPT for Word became generally available on every plan on Thursday, meaning it’s out of preview and open to anyone on the plan. For Business, Enterprise, and Edu workspaces, “Starting Oct 1, 2026, Word access will be enabled by default,” and on those plans Word “uses token-based pricing at the API rates for the model you use,” which is pay-per-use billing on top of the subscription. A firm whose admin allowed the add-in months ago could find Word switched on and metered from October 1 without anyone deciding it this month. It isn’t default-on for Free or Plus. Check the admin setting this week.
The week in one line
The month-end close was offered three times this week with no price of its own, and each time the vendor was selling something else. Intuit’s agent runs it at no charge until January, on a ledger whose price went up on August 1 with AI as the stated reason, and the slide header names the outcome Intuit wants from it: “driving platform preference.” Xero’s is “[Coming soon]” on a ledger whose US price goes up on October 1 with the same justification. Mercury’s ledger is free through December with free seats for the accountant because, on my reading, the product is the bank account. Anthropic’s close has no separate price at all, because it lives inside a paid Claude plan. Call it a loss leader or call it bundling; either way, nobody is pricing the close to make money on the close, and the thing they are pricing is the platform you and your client end up inside.
What the platforms want from you is the playbook: Intuit runs it, FloQast writes it from your finished work, and Anthropic saves your bookkeeper’s adjustments into next month’s procedure. What they haven’t asked for is the judgment about which exceptions matter, the correction record with your name on it, and the review that turns a first pass into something a client can rely on, and Jev suggests the cost of producing that record is about to fall too. So take one client close this month and write two lists. First, what an agent would need from you to run it: that’s your playbook, and if it exists only in someone’s head or inside a vendor’s beta, you don’t own it yet. Second, what the agent can’t do without you: that’s your fee.
If getting from the first list to the second means rebuilding the delivery model underneath it, the Practice Transformation Program is built for exactly that. Enroll your champion at theaiaccountant.ai/transformation: four live sessions, nine modules, and a transformation plan your firm actually executes. The next cohort begins Wednesday, September 30, and registration closes Friday, September 25.