On September 1 an Instagram ad in the UK, tagged #xeropartner and posted from an account with 77,000 followers, said Claude had replaced the accountant its author was paying £800 a month. On September 8 the AICPA said, in its own journal, that a CPA's fee has to price in the liability, not only the time. On September 10 Intuit's CFO said human accountability is the thing AI can't take from assisted tax, and that Intuit sold it last season for $150, and BILL's CEO said that if his software saves you work, you should pay him for it. Four parties answered the same question this week, what is the fee for, and none of them said time. A fifth, California, wrote independence rules for the people who will audit AI and named the profession's standards as the template.
1. A Xero-affiliated influencer told 77,000 followers Claude does the management pack
On Tuesday, September 1, an Instagram ad ran from the account of a UK personal-branding strategist with 77,000 followers, tagged #xero and #xeropartner, with a link for 90% off Xero for six months. She said she had spent more than £120,000 on accountants since 2020, had connected Xero to Claude, and had asked it to build a full management-accounts spreadsheet: "something I used to spend £800 a month having an accountant do for me." Xero had the post removed, so every quote from it is second-hand, via SmartCompany and via Damon Anderson, who used to run Xero's UK operations and wrote the sentence the profession heard: Xero "always sold tools to professionals who did the work and owned the decision," but this "spoke like something else: skip the professional, buy the outcome, here is a discount code."
Xero apologised fast, and twice. Kate Hayward, its UK managing director, replied on LinkedIn within hours: "That ad does not reflect our values or our direction, and it should never have run. I'm sorry." Xero's Angad Soin told SmartCompany the post "was not written or approved by Xero itself, nor was it reviewed in accordance with its usual guidelines, which have now been upgraded."
Then, "shortly after" in Hayward's words, clients who had opened Xero Analytics received an automated monthly analytics email from Xero, sent via Syft, the analytics tool behind that dashboard, from a no-reply Xero address: the business's own numbers, sent by the platform, with no advisor in the loop. Hayward's September 8 blog apologised for that too, said the email "was not intended to bypass the role of the advisor," and pointed advisors at a toggle in each client's analytics settings. By September 10 Xero had paused it by default, but only for Australian customers.
The partner reaction on AccountingWEB is the reaction you'd have: "why should an agent have to visit hundreds of accounts individually to switch off something obviously switched on (without request) centrally by Xero?" I heard the same nerve touched at every event I attended this summer, where the complaint in the room was Intuit moving toward delivering services to accountants' clients directly. I don't think Xero set out to do the same, and Hayward says it didn't: her reply names the intended direction as "designing for the accountant at the centre," and I take her at her word.
But intent isn't trajectory. Both big platforms now have a connector into the ledger (a plug that lets an AI assistant such as Claude read, and in Intuit's case act on, the data in your software), a self-serve analytics tab, and a marketing budget aimed at the owner. This ad is what those three things say when nobody is reviewing the copy.
What makes it the lead story rather than a governance footnote is what the apologies didn't say. Nobody at Xero said it couldn't be done, which is not the same as saying it can: Xero's connector page says "read-only at launch" and, in its FAQ, "No, it doesn't replace your accountant, bookkeeper or JAX" (JAX is Xero's own assistant), the influencer's spreadsheet has never been seen, and nobody has said what the £800 covered. A read-only pull from an unadjusted ledger is a report, not a set of management accounts. But the vendor's answer was to apologise for the message, and the message was that the pack is a prompt away.
I've spent this year arguing that the management-accounts pack is the commodity layer, the rearview mirror (what already happened) that the platform can now install itself, and that the fee has to migrate to the windscreen (what's ahead): the conversation about the next 90 days that a spreadsheet can't have. On September 1 one of the two big platforms' affiliate ads said the first half to small-business owners with a discount code attached, and Hayward's reply said the second half: "some tasks will change. The ones that remain will be more valuable." That's my position in the vendor's words, minus the pricing consequence.
What this means for you. The question the ad asked is the one your clients will ask next, so answer it first. Take your management-accounts deliverable and split it in two on paper: the part that re-presents what's already in the ledger, which a connector can now pull without you, and the part that needs the adjustments, the judgement call about what the numbers mean for the next quarter, and your name on the pack. Then look at the invoice. If it says "monthly management accounts, £800," it prices the half the influencer says she replaced. And if you have clients on Xero Analytics, check the email setting this week rather than waiting for the pause to reach your region, because nothing Xero has published says it applies outside Australia.
The question the ad asked, what the £800 a month was for, gets a full piece of its own later this week.
2. The AICPA just made the value-pricing argument in public. Does your invoice make it?
In June the IRS Office of Professional Responsibility wrote the sentence I quoted in the June 29 roundup: under Circular 230's ban on unconscionable fees, "Cost savings should be passed on openly, with billing practices that reflect the efficiencies gained from the use of [generative AI]," and practitioners should "fairly credit to the client's account any cost reductions." On September 8 the AICPA answered, in the Journal of Accountancy. Formally it's a request for "clarifying language and FAQs" rather than a refusal. The argument it used to ask is the pricing pivot we've been talking about for months.
Eva Simpson, the AICPA's vice president for tax and advisory member value: "The IRS statement implies that any efficiencies gained through AI should be directly passed on to the client. That is an overly simplistic view that ignores the full economics of AI adoption, including software licensing costs, implementation expenses, governance requirements, and the significant investment needed to train professionals to use these tools responsibly and effectively. It also overlooks that professional services are increasingly priced based on value delivered, not just time spent." Melanie Lauridsen, on the FAQ work: "value pricing is something that is just part of business, and you have to be able to include the liability that you're taking." CEO Mark Koziel called the OPR language not authoritative and possibly "overstepping a little bit," and chair Jan Lewis said "we need to bill for that service based on the value provided."
Lauridsen's phrase is the one to keep. "The liability that you're taking" is the wringable neck, the person whose neck the client can wring when the return is wrong, offered as the pricing defence: the fee prices the standing-behind, not only the minutes. One scope note before the practitioner point. OPR's reach is representation before the IRS, and since the Loving and Ridgely cases arguably not ordinary return preparation at all, which is part of why Koziel can call the language non-authoritative; for a CAS owner the relevance is the argument, not the rule.
The same day, Inside Public Accounting put a number on how far the profession has actually moved. Early results from its 2026 survey of 605 North American firms: hours-times-rates fell to 63.1% of revenue from 67.1%, fixed-fee and value-based billing rose to 33.7% from 29.3%, CAS grew 16.7% to 10.9% of the average firm's revenue, and revenue per FTE, about $226,000, "has failed to keep pace with inflation for five consecutive years." Net income still grew 10.4%, so firms are converting slower growth into profit. But nearly two-thirds of revenue in that sample is still billed on time, and the tax slice of it is what OPR's sentence is aimed at.
And on September 9, at an invite-only event in Manhattan run by Accrual (the tax-AI platform that bought Puzzle last week), the leaders of firms much larger than yours said what the AICPA's argument requires. Armanino's CEO: "There's so much wound around time. That creates the wrong incentives going forward that we have to stare down to fundamentally change… It goes to the core of how we price, how we structure engagement." H&R Block's Scott Manuel: "if you're still promoting based on productivity… people are going to do productivity… Technology can create the capacity, but us as leaders have to figure out how to apply that capacity to valuable things." Take the three together and the fee fight is three-sided: a regulator reading the fee rule as cost-plus, a professional body reading it as value, and a market in which two-thirds of revenue is still billed on hours.
What this means for you. The AICPA's argument only works for a firm whose invoice already says so. A tax practice billing hours-times-rates for AI-assisted work has no answer to "time that was not actually spent," because the invoice concedes the premise. A firm on fixed fees, with the accountability and the review priced in, has the AICPA's answer ready-made, and it's the same answer that works for a client asking why the fee didn't fall when the software got better. If you're in the 63%, the FAQ the AICPA is negotiating won't save you. The invoice will, and I'll come back to it at the end.
3. Intuit's CFO says the moat is human accountability, and Intuit's price for it was $150
Intuit's Investor Day is Thursday, September 17. On September 10 its CFO, Sandeep Aujla, gave the preview at a Goldman Sachs conference, and four of his sentences matter for this audience.
First, the channel: "70% of mid-market customers have an accountant, and accountants are a massive influencer in terms of what they adopt… We declared accountants as a customer, not just a partner, several months ago." Two days earlier Intuit gave that declaration a person. Dixie McCurley, who built Cherry Bekaert's national CAS practice, is now Intuit's first Managing Partner in Residence, with a published brief that includes "Define what's next for CAS: Explore how AI, automation, and agentic workflows can change the way firms deliver CAS." (Agentic just means AI that takes a sequence of actions on its own rather than answering a question.)
Second, the product and what Intuit measures. Asked about reconciliation: "They log in. We say, 'Hey, Intuit Intelligence has done all this job for you. Is this right? Anything you want to change?'… and that's saving them 12 to 14 hours a month." Then the metrics: "how many people are logging in, how frequently they are logging in, and most importantly, how many people are allowing our Intuit Intelligence to take action on their behalf. Because it is not just about giving them advice, but taking action, and that is the key differentiator."
That's the point where the software does the work and the human clicks accept, the boundary I quoted from Xero's stage in August, now stated by Intuit's CFO as the KPI. The 12 to 14 hours is a fireside-chat figure with no method behind it; the KPI is the news.
Third, the front door: "We are broadening the front doors. Getting these customers in through QuickBooks Lite, QuickBooks Free." Neither tier is new (I covered the $0 tier two weeks ago), but hear the strategy: a $0 ledger with no accountant access on its feature list, built so that "as you grow, you very easily get into Simple Start because it's on the same platform." For a business that starts on the free tier, the software is the only accountant it meets, and the tier is built so that stays true through Simple Start.
Fourth, the one worth reading twice. Asked whether Intuit felt AI-native tax competitors this season: "We did not." His reason is one this audience will recognise: "88% of the market is in assisted tax… What people are attracted to assisted tax is outsourcing to human accountability. That is something that AI is not going to be able to go after." (The 88% is by his measure; DIY software files far more than 12% of returns, so read it as Intuit's view of where the money is.)
Then he turned it into a weapon: "On the assisted side, our strategy is to be price disruptive… we had an offer out there for $150, we'll get your taxes done in assisted method… using AI, having the ability to be price disruptive is actually on strategy for us and is actually an accelerant." The $150 was a past-season offer, not a list price. He added that it "pains me tremendously" when a simple filer at $50,000 of income pays a competitor $400 or $500.
Put the four together. Intuit says the accountant is a customer, has hired one to prove it, measures success by how much autonomous action the client permits, is widening a free front door for businesses too small to have an accountant, and believes the human-accountability layer is defensible. Then it priced that layer at $150 when the human is Intuit's. The wringable neck is the framework I've used all year to say what survives, and Intuit's CFO agrees; he has also decided what it costs.
What this means for you. "Accountant as customer" is good news with a price attached: a vendor that sells to you builds for you, and also decides what you're worth. Do two things. For the simple returns in your book, the ones a W-2 and a couple of slips would cover, put your price beside $150 and write down what your version of accountability includes that Intuit's doesn't: the planning conversation, the entity question, the letter you answer in March. If the honest answer is "the same thing, for $400," that's the gap Aujla says he's targeting. And watch Thursday for one thing only, which is what Intuit unpacks about the accountant channel; his tease was "you'll see next week when we unpack even more data on the big bets."
4. Your vendors want to price what they remove. The labs are pricing what you use.
Buy vs Build Part 4 opened with BILL's stated intent to move from per-seat to consumption pricing. On September 10 its CEO, René Lacerte, gave the numbers behind it at the same Goldman conference. "1 year ago, we would have had close to 2,500 employees, and now we are around 1,500. That is close to 40% in 1 year… We created a flatter, leaner organization… we felt that AI required that." (The only filed figure is a May restructuring of "up to 30%"; the 40% is the CEO's framing across a longer span.)
On the product side, "over 20 agents" now touch roughly 95% of risk decisions, the W-9 Agent has 40,000 customers and over 240,000 W-9s, and 60,000 customers use the Invoice Coding Agent. On the accountant channel, "close to 10,000 accounts," and on adoption, "it's not a question of adoption."
Then the pricing sentence, which a CAS owner should read slowly because it's the pricing pivot from the vendor's side, in accounting vocabulary: "I am an accountant at heart, so I like the matching principle. If we're saving you work, you should pay us… to be in a position to actually change from just the subscription pricing model we have today to something that actually more closely matches and aligns with the value that we're creating." He said it was early days, and no price change has shipped. But the intent is stated: the AP platform wants to price by the work it removes from your firm.
The labs moved the same week, for a different reason. As of September 10, OpenAI is "temporarily pausing new sign-ups and upgrades to the ChatGPT Pro $200 plan," including upgrades from every cheaper tier, with no end date; "once your Pro $200 subscription ends, you cannot purchase it again until the pause is lifted." The stated reason is demand for its new model.
Two days earlier a refiled class action, filed by two former FTC officials acting as plaintiffs' counsel, alleged that Anthropic's "5x" and "20x" Max plans mislead because the multiple applies to five-hour windows under a weekly cap. Neither is a price rise. Both are what happens when the thing being sold is capacity rather than a seat, and the flat-rate plan is the one that runs out. My reading: if the top flat plan is closed, the only route to more capacity than the mid tiers allow is the metered one, paying per token, which is roughly per word in and per word out.
Set BILL beside the labs and you have two vendors leaving the seat for different reasons. Your AP vendor wants to charge you for the work its agents remove; your model supplier is rationing the flat plan and billing the rest by the token. The seat is still the only unit most firms budget AI on, and it's the unit both sides are walking away from.
The useful half of the week is that the measuring instruments shipped too. Anthropic's Smart reports (beta, Enterprise only, so out of reach for a 10-person firm today) report "Cost per session by type of output" and count "sessions that produced nothing usable… with their cost." That's the first cut of a cost-per-accepted-result number built by a vendor, though it records what was produced rather than what you signed off, and the point for now is that the category exists.
Anthropic's September 8 cost post also names a habit that costs money if you're paying per token: "Instructions like 'double-check your work' or 'verify twice before responding' are often taken literally by frontier models and can waste tokens." In its benchmark, which planted that kind of anti-pattern and then removed it, removal cut costs 14.6% and raised accuracy 5.3%. And Julien Simon did the arithmetic nobody prints: at the same $10-per-million-input and $50-per-million-output list prices, Anthropic's $0.25 cache read against OpenAI's $1.00 (a cache read is the model re-using material it has already been sent, such as your standing instructions) means a 20-turn agent job costs about $3.50 on one and just over $6 on the other.
What this means for you. Three things follow. First, assume the vendors in your stack are watching BILL, and reprice your own services before they do, because "if we're saving you work, you should pay us" is the sentence you need to say to clients before your AP vendor says it to you. Second, stop budgeting AI per seat; the plan can be paused, the multiple can be litigated, and the metered bill depends on the cache and the effort setting (how long the model thinks before it answers). Third, if you're paying per token, delete "double-check your work" from your standing instructions. Verification belongs in a separate pass with a record, which is the provenance check I wrote up on September 4, not an incantation that costs tokens without adding a check.
5. California just wrote the job description for an AI auditor, and put CPAs in it
Two bills signed on September 9 make California the first state to register the people who audit AI. They define an AI audit, write independence rules for the auditors, and, in the enrolled text of AB 1405, let a registered CPA satisfy the report and independence requirements by following the profession's own standards instead.
The mechanics first. SB 813 defines a "covered AI audit" as "an audit conducted to assess internal controls, processes, or systems implemented for an AI system or model that are necessary for compliance with state law," and tells the state to align its procedures, to the extent practicable, "with existing professional and regulatory audit and assurance standards." AB 1405 creates an AI Auditor Registry by January 1, 2029, after which nobody may "offer, sell, or conduct a covered AI audit unless the person is registered," and it copies the profession's independence rules almost clause for clause: no auditing your own work, no negotiating for a job with the auditee mid-engagement, a 12-month cooling-off period, and no fee contingent on the result. Then the carve-out: a registered AI auditor holding a California Board of Accountancy licence, and a firm with a permit to practise public accountancy, "shall be deemed to satisfy" the report-content and independence paragraphs, provided they comply with the Accountancy Act, the AICPA Code of Professional Conduct, and "the attestation standards promulgated by the AICPA." Complaints against a CPA auditor go to the Board of Accountancy.
Be precise about what it isn't. It's a registry, not a licence, and it's open to anyone who registers; the CPA carve-out is a side door, not a monopoly. CPAs still have to register. Nobody is required to buy one of these audits; SB 813 says so explicitly.
The registry is 27 months away, and it's one state. OpenAI endorsed both bills the same day and called for "mandatory, capability-based national regulation," and a lab endorsing a state audit regime is a tell that it expects one nationally.
Now the reason it's a numbered story and not a quick hit. I've been reluctant to write about AI assurance because it isn't CAS work, and most of you won't do it. But look at the shape of what California built and compare it to SOC 2.
SOC 2 exists because the market needed an independent opinion on a service provider's controls, the profession already had the attestation standards and the independence rules, and so CPA firms own it. AB 1405 describes the same shape for AI: a market that needs an independent opinion on the "internal controls, processes, or systems" around a model, and a statute that names the profession's standards as the template. The difference is that California left the door open to non-CPAs, so the profession's edge here is the one it already has, which is 100 years of independence habits and a standard-setter that already exists. Two weeks ago Schellman, a CPA firm, became the first accredited auditor under a private AI standard, and now the largest state has built the public version.
Here's the connection to everything above. Stories one through four are about fee compression on the work you do now, and the answer I usually give is advisory. AI governance and attestation is the other answer: a service line whose price is set by liability and independence rather than hours, which is the pricing pivot with the profession's own rules as the moat.
The professional rule that limits it is independence, and it should be said plainly: you can't attest on the AI governance of a client whose books you keep. It's a California registry today, and a 10-person CAS firm won't be issuing AI audit opinions in 2027. But the demonstrable-QC record I keep telling you to build for your own AI use (who reviewed what, what was corrected, who signed) is, in California's statute, the thing that gets audited. The firm that has one can be audited, and knows what it's looking at when a client asks who audits theirs.
What this means for you. There are two moves, and neither is "become an AI auditor." Build your own record now, because the statute describes what a compliant one looks like, and the clients of yours who sell into larger companies may be asked for the equivalent in a vendor questionnaire long before 2029. And decide now which clients you could never attest for (the ones whose books you keep) and which assurance partner you'd refer the rest to, because the profession is about to be asked, and the firms that answer first will set the price.
Quick hits
10,980 Americans put "a small business's monthly bookkeeping" second on the list of what AI can do soonest. Anthropic's economists published three scenarios for 2030 on September 9, and the survey behind them asked 10,980 US adults when AI could do eight tasks as well as a skilled professional. Ordered by how hard respondents thought each was: routine business emails first, a small business's monthly bookkeeping second, an accurate personal tax return from someone's documents third. That's lay perception, not capability, and "personal tax return" was undefined, so read it as the simple W-2-and-slips return. It's still the client-perception wall of the four-way squeeze (the four forces pressing on the compliance fee at once), measured, with the two commodity products of a CAS practice sitting second and third on the public's list.
In the middle scenario, the one closest to the public's median expectation, wages in cognitive occupations (which include business and financial operations) are flat by 2030 while other workers gain 5.9%. Jack Clark's line on NPR is the adoption gap stated by the lab itself: "it will get really, really good. But it will make its way into the economy more slowly."
Apron put a per-client price on AI, and Caseware put an agent in Excel at no extra charge. Apron, a UK bookkeeping-automation vendor, is giving its William AI assistant to every user and, from December 1, pricing it per client entry at £5, £11, and £21 tiers with a £199 cap for existing top-tier partners. Its "Guidance 2.0" rules are written in plain English ("Invoices from Vodafone are always Telephone and Internet, with 20% VAT"; "Never publish documents over £1,000 automatically"), which is the coding layer of a client context file (the standing instructions that tell an AI how this client's books work) shipped as a product and priced per client. Caseware made Verity for Excel generally available at no additional licence for firms that already have Verity, with a Response Validation Agent that checks client documents against the request before they reach the workpaper while "the decision to accept a submission always rests with the auditor." And Microsoft's Excel Copilot can now answer "What changed that caused this formula to break?" and undo a colleague's AI-assisted edits while keeping the manual ones. The review layer is arriving inside the tools, and the price sheet for client context is per client.
Your API key is the target. Anthropic's threat report this week says AI access itself is now "the sole objective of multiple criminal groups": fake installers "spoofing as popular AI harnesses including Claude Code," a "cheap Claude" reseller whose traffic "was in fact silently proxied to a different AI model," and keys lifted from public GitHub repositories. Its one customer-facing rule is the one to print: treat AI keys and agent integrations "with the same level of seriousness as they do production credentials… An alleged discount that requires routing traffic and credentials through an unknown intermediary introduces tremendous risk to user data and systems." OpenAI shipped the boring fix the same day: expiry dates and a maximum lifetime on API keys. For a firm, install from the vendor only, refuse discounted access through a third party, and treat every key like a bank login. And if Claude Cowork stopped reaching your files on Windows this week, it's the September 8 Windows update, not you; Anthropic says it's investigating.
The week in one line
Four parties answered "what is the fee for?" this week. An influencer with a Xero discount code said the pack is a prompt away. The AICPA said the fee has to carry the liability. Intuit's CFO said it's human accountability and sold it for $150. BILL's CEO said it's the work removed, and California, for good measure, wrote independence rules for the people who'll audit the machines and named the profession's standards as the template.
Not one of them said hours, and the IPA survey says 63% of revenue at the 605 firms it measured still does.
You've heard me make this argument for months. What changed this week is who's making it: the platforms, the profession's own body, the AP vendor, and a state legislature, in their own words and for their own reasons. When the market and the profession agree on where the fee lives, the firm still billing on time isn't holding a position. It's holding an invoice whose only remaining defender is a regulator that wants the savings handed to the client.
So pick one engagement letter that renews this month and rewrite the fee line so it says what you're accountable for, with no hours on it. If you can't write that line, that's the work.
If rewriting the line means rebuilding the delivery model underneath it, the Practice Transformation Program is built for exactly that. 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.