Two weeks ago, in the Xerocon roundup, I priced three AI-native ledgers against QuickBooks. This week the middle one, Puzzle, agreed to sell its accounting-firm business to a tax platform and told the trade press it will stop marketing software directly to CPA firms. Meanwhile OpenAI launched GPT-6 with a financial-statement tie-out at the top of the page, Docusign said it will open its connector to every AI agent on September 30, and a services firm priced an assurance engagement run mostly by AI agents at a third to a half below its own rate card.
1. The $60 ledger is changing hands
In the Xerocon roundup I put three AI-native ledgers against QuickBooks on price: Kick at $40, Puzzle at $60, Digits at $65. On Wednesday Puzzle agreed to sell its accounting-firm business, including the ledger and its month-end close products, to Accrual, the tax-first AI platform that raised $75 million from General Catalyst in February and counts Armanino, Aprio, and H&R Block among its customers. Terms weren't disclosed and the deal hasn't closed. It's "expected to close in the coming weeks, subject to customary closing conditions."
Read the two halves separately, because they say different things. Accrual's half is a tax vendor buying a ledger and a close. CEO Cosmin Nicolaescu told Accounting Today "the Close experience is what made Puzzle especially compelling to us," and the release notes Accrual has run its own books on Puzzle since early in its history, so the acquirer was a customer first. Accrual's CAS offering is in early access with firms now, with broader availability planned by the end of 2026.
Puzzle's half is the one a CAS practice should read twice. What remains of Puzzle will, in founder Sasha Orloff's words to Accounting Today, "no longer market a standalone firm-facing technology product directly to CPA firms." In its place, "Puzzle 2.0" is "an integrated accounting platform, combining our technology, our customer acquisition engine, and a small group of ambitious accounting partners," selling accounting plus tax to small businesses at "transparent, fixed pricing." Orloff says the demand came from customers who "wanted more than software," and that the new model puts Puzzle "one step closer to profitability, funded by our revenues and no longer dependent on the next round of venture capital."
So the $60 challenger I priced against QuickBooks is, on Orloff's description, becoming a fixed-price accounting-plus-tax service sold direct to small businesses, with a small group of partner firms somewhere in the delivery. Be careful with that sentence, because the mechanics haven't been published. Nobody has said whether one codebase or two survives the split, who builds the ledger from here, whether Puzzle 2.0 does the accounting itself or fronts it and hands the work to partners, or what any of it costs. Nothing published says Puzzle functionality is being withdrawn from firms. What has been said is who sells to whom, and that Puzzle will stop selling to you.
That is the open question. Nicolaescu says "there is no immediate required change for customers of either company," and on who migrates: "Accrual remains focused on serving Top 100 accounting firms, so, to the extent Puzzle's clients in that category are interested in moving over to Accrual only, we look forward to serving them." That answers it for Top 100 firms. Nobody has answered it for a 12-person practice running 40 clients on Puzzle today, and Puzzle's own customer list includes outsourced-accounting firms like Supporting Strategies and Decimal, which are exactly the firms the split touches.
Two things happened to a CAS practice in one announcement. A new competitor appeared: a fixed-price accounting-and-tax service aimed at the same small businesses you serve, built on an AI-native ledger and funded by its own revenue. And the ledger you might have used to compete moved behind a platform whose CEO says it "remains focused on serving Top 100 accounting firms." The tool went up-market and the competitor went down-market, and a 12-person practice sits between them. That's two walls of the four-way squeeze, the four forces pressing on the compliance fee at once, closing from opposite sides in one deal.
Step back and this is the third time in nine weeks I've written the same sentence. Basis runs the close inside top-25 firms and is waitlist-only for everyone else. Intuit's agent builder is piloting with enterprise firms first. Accrual says Top 100, and now holds the ledger.
The AI-native vendors whose pitch was that they'd level the field are choosing the top of the market, and a firm my size is watching the packaged delivery layer close from above while a fixed-price competitor opens from below. I think that's the most disturbing trend in the profession right now, and I'll say why at length soon. That argument belongs in the next part of the Buy vs Build series, which lands later this week. The short version: the model layer is priced the same for you as for Deloitte, and that's the layer to build on.
What this means for you. If you run clients on Puzzle, nothing changes this month, and you should ask this month which of the two companies you'll be a customer of by year-end and on what terms. If you were evaluating an AI-native ledger as your escape from Intuit's price rises, the escape route changed owner mid-evaluation; moving to a challenger ledger is a rebuild decision, and this week showed the rebuild target itself can move. And find out whether Puzzle 2.0's partner programme is open to firms your size, because it's the only published route by which a non-Top-100 firm gets this ledger at all. A fixed-price service that brings its partners customers is either a channel you join or a competitor you price against, and you should decide which before it publishes a price.
2. The tie-out is now on the launch page
Three frontier launches in three days: Anthropic's Claude Fable 5.1 on Tuesday, Google's Gemini 3.8 Flash on Wednesday, OpenAI's GPT-6 Astra on Thursday. Frontier model just means the most capable tier each lab sells. The story for this audience isn't the models. It's what OpenAI chose to put at the top of Astra's launch page: a financial-statement tie-out.
The customer story is Legora, a legal-AI company now moving into audit, tax, and compliance. Its agent "completed the tie-out across 41 documents in a single run" and "found all four errors Legora had planted in the accounts, including a £500,000 gap hidden in the revenue note." Legora's own engineer says the manual version "can take an entire evening, sometimes days." That's a workflow every reader of this roundup has done by hand, and a frontier lab decided it was the best advertisement for its new model.
The detail underneath that number matters more than the number. Legora reports Astra improved "nearly 40%" on the financial-statement workflow, against about 3% across its own suite of legal tasks. It's one vendor's test on a launch page, so give it that much weight and no more. But it lines up with something I've argued all year: narrow, repetitive, high-volume work is where the models improve fastest, and the tie-out is exactly that kind of work.
Read the other finance claims from the same week with one distinction in mind. Hebbia, Samaya, and Google all published finance results too, and those are investment-analyst tests: pulling figures from SEC filings, computing growth rates, comparing guidance to actuals. The task isn't yours, but the skills are: find the number in a long document, compute from it, and reconcile one document against another. That's what a tie-out is made of, so gains on those tests are gains on your work by another route.
The one caution is Google's claim that the $0.75 Gemini 3.8 Flash beats frontier models on that test: it's Google's own run, and the benchmark owner hadn't published a score for it as of this weekend. Legora's result is the one this week that measured your workflow directly.
Now the price, because the two frontier models landed at an identical list price: $10 per million input tokens and $50 per million output tokens. A token is roughly three-quarters of a word, so that is a price per word in and per word out. The price war I tracked in July stopped at the top of the range this week; neither frontier vendor cut the headline number. What moved is underneath it.
Two settings decide what you actually pay. The first is the cache read: the model re-using material it has already been sent, your standing instructions, a client file, a skill, rather than reading it fresh every turn. Anthropic's pitch rests on it: Fable 5.1 "will cost an estimated 25% less than Fable 5 for typical workloads," and up to "approximately 45%" for agentic work (work where the AI takes a sequence of actions on its own), because a cached token now costs $0.25 per million instead of $1.00, a 75% cut. The headline rate didn't move; the rate on repeated material did, and in agentic work repeated material is most of what the model reads each turn.
The second is the effort setting, a dial that controls how long the model thinks before it answers. Independent tester Artificial Analysis (which discloses it helped Anthropic with pre-release evaluation) found that at maximum effort Fable 5.1 costs $3.76 per task on its test suite, 20% more than Fable 5, because it produces about 1.7 times the output. One notch down, at extra-high effort, the same model scored higher than Fable 5 at maximum and cost less, at $2.72 per task. Anthropic says the same thing from the other side: at low or medium effort the new model matches or beats the old one "at a much lower cost," and the defaults in its own products are high and medium.
So the two figures aren't a contradiction; they measured different settings. Turn the dial to maximum and the new model costs more per task than the old one. Leave it at the default and it costs less.
Google said the quiet part on its own launch page: "At times, the model might use more tokens to maximize performance, especially at higher effort levels." And OpenAI added a cost most firms haven't budgeted for: Astra ships with a production safety monitor, and "extra safety checks can sometimes slow, pause, or stop legitimate work... In the API, the task will stop." A paused job still has a bill.
What this means for you. If the vendors are right about the tie-out, the defensible human work is deciding what counts as an error, testing the model against that standard before you trust it, and signing off on what it finds. I've been calling that the review layer for two years, and this week a frontier lab put it at the top of its launch page. On cost, stop quoting AI per seat. The bill now depends on the effort setting, how much of your standing material gets reused, how many helper agents a task spawns, and whether a safety monitor stops the job halfway, and none of those is on a price sheet. Measure cost per accepted result instead: what did it cost to get a piece of work you signed off?
3. Every vendor now wants to be a tool call inside your agent
In July I wrote that every vendor wanted to be your only vendor. Six weeks later the pattern has inverted, and this week it arrived from three directions. Intuit put QuickBooks into Perplexity Computer (Perplexity's agent product rather than its search app), so US QuickBooks customers can follow up on overdue invoices, set up recurring invoices, send payment links, and pull payslips and payroll information, signed in with their existing QuickBooks account. That's the same set of actions Intuit shipped into Claude and ChatGPT in July, now on a third assistant, and it's the same move Xero described at Xerocon, where MCP connections into the ledger were up 10-fold and a ChatGPT connector was announced as coming soon. One vendor adding one more assistant isn't news. Most of the tools you already use racing to be reachable from every assistant is.
The dated event is Docusign, which announced on Thursday it "will open its Model Context Protocol (MCP) Server to every AI agent on September 30." MCP is a standard plug that lets any AI assistant call a piece of software directly, the way a USB port lets any device connect to any computer. Docusign's will be "callable natively from Claude, ChatGPT, Gemini, Copilot, Slack, and any MCP client," with admin controls at the account level. If your engagement letters go out through Docusign, an agent could be sending them from the 30th, once your admin switches it on.
The accounting-tool layer is doing the same thing in miniature. Numbersgame now offers "one controlled MCP for the client books it manages" across QuickBooks and Xero, with permissions scoped per client. AuditFile shipped MCP access so a firm's other approved AI tools can task and query the audit file, and Bizora prices audit research at $0.65 per query through the same kind of plug. Ledger, e-signature, audit file, research: four categories of vendor, one week, all building the same plug.
This is the opening. I've told firms to leave Zone 3, automation inside the platform itself, until last, because getting in meant a fragile browser script that broke on every update. The platforms are now taking that wall down themselves. The gap workflows I've been telling you to build (client communication, exception review, workpaper assembly, engagement letters) can now reach into the ledger, the e-sign layer, and the audit file through the same agent, using the vendor's own connector and, for Docusign and the accounting tools, the vendor's own admin controls. For a small firm that is a step change in what you can automate, and it arrived from three directions in one week.
One item goes on the list, and it is a permissions item rather than a reason to hold back. The Intuit route runs under the client's own QuickBooks login, with no firm admin layer in between, so the client's credentials exercised from inside a Perplexity agent can now send a payment link from books you reconcile. The client could always do that in QuickBooks; what's new is that an agent you never see can do it too. Numbersgame's framing is the right shape: a firm-controlled plug with scoped permissions per client, rather than "an all-or-nothing permission model for each client."
What this means for you. Pick one gap workflow this month and wire it to a sanctioned connector instead of a screen-scrape; the engagement-letter flow through Docusign on September 30 is the obvious first candidate. Then ask three questions about your client list: which ledgers are reachable from a consumer AI account, under whose credentials, and does the engagement letter cover actions you didn't take? The answer is a paragraph to add, and it shouldn't slow you down.
4. A firm just used AI to undercut its own rate card, on purpose
In July I wrote that the question is no longer who does the work. This week a professional-services firm answered it anyway, in public, with a price attached. Cherry Hill Advisory now performs external quality assessments of internal audit functions, an engagement governed by the IIA's Global Internal Audit Standards, "primarily" with two named agents: "Renata," an assessor trained on the firm's completed engagements, and "Tomasz," a QA lead that pre-clears the work against the IIA's quality manual. Every rating, conclusion, and final opinion is reviewed and signed by a Certified Internal Auditor. CEO Mike Levy, in an interview with Accounting Today: "On average we are seeing these assessments cost 30% to 50% less than our traditional assessments," and about 50% faster.
Look at what the 30 to 50% is measuring, because it isn't a discount. The comparison is against the firm's human-performed assessments: the same engagement, the same standard, the same signature, with the assembly and documentation done by agents instead of staff. The lower price is the cost of the delivery model, not a promotion on the old one.
Levy's stated reason for passing it on is reach rather than share: "My goal is to make this accessible to every function, regardless of budget, since there are many functions that can't do this work because of budgetary constraints today." In June the IRS's Office of Professional Responsibility said AI savings on tax work "should be passed on openly." Here is a firm doing that voluntarily on an assurance product, and using the savings to reach buyers it couldn't serve before.
The detail I'd frame and hang on the wall is that the engagement file "records reviewer agreement/disagreement" with the agents. That is demonstrable quality control, the review record that shows a human checked the machine, produced as a by-product of the delivery model rather than as extra work. Weigh it properly: one firm, one engagement type, self-reported, no client quoted. But I haven't seen another firm publish the price, the sign-off model, and the review record together.
The big firms moved the same way. AuditFile's launch copy for its agent suite is "the firm that staffs itself: agents do the work, humans apply judgment," with a bring-your-own-model option so the firm's existing AI subscription runs the agents. EY US put $100 million into rewards for staff who "harness advanced technologies," with a top award of $25,000. Deloitte stood up a global practice to build on open-weight models (models you can download and run on your own hardware) and is hiring "forward deployed engineers" to run it, the closest thing the Big Four has to a job title for the Champion role.
What this means for you. Cherry Hill's move raises a specific question for a CAS practice. When agents do the assembly and a human signs, the cost of delivering the engagement drops, and you have to decide where that saving goes. Holding the fee flat and hoping nobody notices is one answer, and it's the one Intuit's toggle and Puzzle 2.0's fixed price are both designed to break. The better answer is Levy's: price the agent-delivered version on what it costs to deliver, aim it at the clients who couldn't afford you before, and let the engagement file record who reviewed what. The signature is what you're selling now, so price it like it.
5. Two-thirds of preparers use AI for research. 93% have no tool for finding the planning opportunity.
Last week I set Singapore's national AI programme against an American salary survey that never asked about AI. This week the profession's other flagship survey did ask. The Journal of Accountancy's 2026 Tax Software Survey polled 1,808 AICPA members who prepared 2025 returns for a fee, and 65% said they use AI in tax research, followed by client communication at 32%. Only 16% have no plans to use AI in their practice, down from 24.3% a year earlier.
The adoption gap is closing in the workflow that feeds a court filing, and the same issue carries the case study. In Clinco (T.C. Memo. 2026-16), Judge Holmes found three apparently hallucinated citations in a taxpayer's filing, cases that don't exist at the volumes cited, in a dispute over more than $2.2 million of underreported Schedule C receipts. The memo doesn't say what produced the citations, and the filing was the taxpayer's, not a preparer's. But two-thirds of preparers researching with AI and a Tax Court memo on invented authorities written up in the same issue is the review-layer argument in one journal.
The number to keep is a different one. Asked whether they use any tool to review returns for planning opportunities, 93% said no. They don't lack planning software; 37% of the same respondents own it, with Planner CS and Bloomberg Income Tax Planner the named leaders. Read it precisely: the planning tools run a scenario the practitioner has already spotted, and almost nobody uses anything that looks at a finished return and says where to look. Of the 20% of respondents expanding services, 75% named tax advisory and 55% named CAS.
What this means for you. The pivot to advisory is happening and the tooling for it isn't. The white space in your tax practice is the instrument that tells you where to look, and you already have the raw material: several hundred finished returns from last season sitting in your tax software. A defined checklist of planning triggers (entity choice, retirement plan headroom, estimated-payment exposure, reasonable compensation, state nexus) run by AI across those returns, inside a tool your firm has approved for client data, is a Zone 2 workflow: export, review, act.
Two things follow from the survey. First, that return-scanning workflow is the cheapest source of advisory conversations you'll build this year, and the survey says 93% of preparers have nothing like it, so you'd be ahead of almost everyone. Second, if your team is among the 65% using AI for tax research, add one line to the review checklist before the next filing: open every authority the AI cited and confirm it exists and says what the memo claims. Clinco is what happens when nobody does.
Quick hits
OpenAI plans to cut Cursor off on November 12. OpenAI told SpaceX, Cursor's new owner, that it intends to stop supplying its models to the coding tool "because we cannot be confident that SpaceX will use our technology within our terms of service." Nobody using Cursor did anything; the model they built their workflow around leaves the app on a proposed date set by a dispute between two other companies. Nate Jones wrote the version for this audience: "a model can disappear from the app where you learned to use it, even when the model and the app both remain available separately," and the asset at risk is "everything the model has learned about your work."
If your provider changed the deal on a Friday, could another model continue Monday's close from what you hold outside it? If not, the client context you've spent 2026 encoding lives in someone else's building. Rent the surface, own the knowledge.
AI fell to the fourth-cited reason for layoffs in August. After five months as the leading stated reason for US job cuts, AI was cited for 3,462 cuts in August, its lowest monthly total since December, according to Challenger, Gray & Christmas. Year to date it's still first at 116,175, about 22% of all cuts. Before you read that as AI cooling, read Challenger's own caveat from July: naming AI in a layoff announcement "can win over investors while pushing current and prospective employees away," and as regulation takes shape "companies will be even more careful in their announcements."
The series measures what companies say, in both directions. The better number for a client conversation comes from Ramp: the median business on its platform spends $11.95 per employee per month on AI, which is the adoption gap in dollars.
NVIDIA is buying Hugging Face for $12.93 billion. Last week it was a rumour; this week it's an 8-K. Hugging Face is where open-weight models are published and downloaded: 3 million models, 18 million developers, the place Thomson Reuters published its model two weeks ago and the platform OpenAI's agents broke into in July. Jensen Huang's pledge is that "NVIDIA compute will not be required to build on or deploy through Hugging Face," with close expected in the first half of 2027. For a practice this is a footnote to the data-sovereignty thread: if you ever run a model on your own hardware for confidential client work, this is where it will come from, and the chip vendor now owns the shop.
Washington's AI bills now name a tax base. Rep. Greg Casar's bill would levy an excise on AI tokens at a rate that rises with unemployment, Senators Wyden and Warren have data-centre and energy-based variants according to Fortune, and on Thursday Sanders and Casar introduced a bill to ban "artificial superintelligence" outright and pause advanced development until a federal regulator exists. Axios says the ban isn't expected to advance, and nothing suggests the tax bills move this Congress. It's here because a token excise is a consumption tax on the exact input in story two, and it would land on the same line as the cache discounts.
Your clients will ask what an AI tax means for their bill long before anyone asks you about superintelligence. Have a one-paragraph answer ready.
The week in one line
Every headline number in this week's news is true, and not one of them tells you what matters. $60 a month was Puzzle's price, and the story is who sells it to whom. Four errors found is the headline, and the story is who decided what counted as an error and who signs the result. $10 per million tokens is the price, and your bill depends on a dial.
Docusign opening to every agent is the headline, and the story is whose credentials. 30 to 50% cheaper is the headline, and the story is the review record. 65% of preparers use AI for research, and the number that should change your practice is the 93%.
The vendors, the surveys, and the press will keep handing you the headline figure. Reading underneath it is the job, and it's the same job you already do on a client's P&L: the revenue line is true, and the story is in the notes. So take one number you've been quoting about AI in your own firm, cost per seat, hours saved, tools adopted, and ask what's underneath it. If you can't answer, that's the number to replace this week.
If the number you find underneath is that your firm is bolting AI onto the delivery model it already has, the Practice Transformation Program is built for exactly that gap. Enrol 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.