Basis announced on August 5 that it's "the first platform to take responsibility for the entire first pass" of a tax return. Four days earlier, OpenAI published 10 solved mathematics problems and did the opposite: it refused to claim authorship of work its model did alone, taking responsibility only for whether the answers were right. Both sentences are about who's answerable when a machine does the work. Only one of them is being sold to you, and it's the one that gets the distinction wrong.
Two companies claimed AI accountability this week and only one of them meant it
Basis launched its End-to-End Tax Platform on August 5, claiming it deploys agents across the whole tax return rather than automating one step at a time. An agent here just means software that carries out a sequence of actions on its own instead of waiting for a prompt at each stage: it gathers evidence from the client, prepares the return, reviews it, and brings the accountant in when something needs judgment. There's no pricing, no availability date, and the only way in is a waitlist.
Read that claim again, because the verb is doing something the product can't. A software company can't take responsibility for a tax return; it can only take the work. Responsibility has a license behind it, an insurer behind that, and a person whose name goes on the file. We've argued before that accountability is what no client's agent can swap for a cheaper alternative, and this is the first time a vendor's marketing copy has quietly claimed it.
OpenAI's version is the same distinction handled properly. Its unreleased internal model resolved or advanced 10 long-standing open problems in mathematics, at a compute cost it puts at roughly $2,000. OpenAI then wrote that "claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system's contribution and the nature of genuine human intellectual work," while taking responsibility "for their correctness." The machine did the work, humans stand behind whether it's right, and those are two claims that mustn't be collapsed.
Then on August 2 the EU AI Act's transparency obligations took effect, and disclosure of AI-generated content became a legal duty rather than an ethics preference. If you serve a client in the EU you're inside its reach, because it catches firms in third countries where the output gets used in the Union. Being in scope isn't the same as owing something: the marking duty sits on the model vendors, a client deliverable isn't published to inform the public, and published commentary is exempt where a human reviewed it and someone holds editorial responsibility for it. What's left exposed is unreviewed work going out with nobody standing behind it, which is the same sentence as everything else this week.
Buy vs build, answered three different ways by three firms in one week
Between August 3 and August 5, three large firms made the same decision and resolved it three different ways. Deloitte built its own and productized it, launching ControlCatalyst.AI across the internal audit, SOX, risk, and controls lifecycle. Aprio, "the 20th largest business advisory and accounting firm in the U.S.," co-built with a vendor: its own professionals are building and refining audit agents alongside Fieldguide, spanning planning through fieldwork, review, and financial statement preparation, inside a $300 million five-year commitment. Citrin Cooperman did neither and hired the builders, engaging Ode with Anthropic to run its AI work across tax, accounting, and advisory.
Be precise about that third one: Citrin Cooperman's announcement is not an audit story and never uses the word agent. What makes the cluster worth your time is that build, co-build, and rent are all on the table in one week, priced at the top of the market, and there's a version of each at your size. This is the argument the buy versus build series has been making since the pricing wrapper piece, and Part 3 finishes it on Wednesday.
Notice what none of them outsourced. All three bought capacity to do the work. Not one bought someone else to be answerable for it, because that isn't a thing you can purchase.
The productivity gain has a default owner and it isn't the worker
Four data releases landed in three days last week and they tell one story. Productivity rose 1.4% in the second quarter while output rose 1.7% and hours moved 0.3%. Workers' share of the pie, meaning the portion of output going to wages and benefits rather than profit, fell to 52.9%: the lowest in a published record that starts in 1947. Output went up, hours didn't, and the difference went somewhere.
It isn't going into your profession's payroll. Financial activities employment fell another 14,000 in July and is down 121,000 since a recent peak in May 2025, in a profession that calls itself short-staffed. A panel of economists surveyed in July expects downward wage pressure on college-educated workers at nearly twice the rate it expects for workers without a degree, 57% against 34%, and names banking and finance among the occupations it expects to lose most jobs to AI.
Here's the part that transfers straight to a 10-person practice. A productivity gain has no natural home: it goes wherever your pricing model already points it. Bill by the hour or by transaction count and a faster close cuts the invoice by itself, so the client captures it without either of you deciding anything. Hold a fixed fee and those same three days land on your margin. Put nothing into the three days at all and the capacity just dissipates, which means nobody gets it.
There's a fourth claimant, and this profession has barely started talking about them. Real hourly compensation fell 3.1% in the second quarter, the steepest quarterly rate of decline since late 2022, while those same workers produced more. The senior who absorbed a third more clients this year can do that arithmetic without reading a productivity release. Here's where I'll go past the data: the IRS has already told tax practitioners that AI savings should be passed to clients, your margin wants those same savings, and the person carrying the heavier load is the only claimant at the table without an advocate. That's a retention problem arriving dressed as a margin improvement, and it will surface as resignations long before it surfaces as a policy discussion.
Two capabilities went free, and one of them is your client's
Luna, whose 80% price cut we covered last week, is about to become the default model for free ChatGPT users. OpenAI said on August 6 that it lands this week, and none of it is confirmed live yet. The number worth your attention is buried underneath: on an internal evaluation of financial, medical, and legal prompts requiring factual detail, responses containing at least one factual error were about 62% less common with Luna than with the version it replaces. That's OpenAI's own measurement and nobody can reproduce it, because the prompt set has never been published.
Every client who's ever asked ChatGPT a tax question is about to get a version materially less likely to be wrong, at no cost, without being told it changed. That doesn't remove the review layer: 62% fewer answers containing an error is not zero errors, and the residual is where you earn. But your client's baseline confidence is about to move whether your firm's tooling does or not, and that shows up in fee conversations first.
The second one is quieter. At the end of July, Accounting Seed shipped a hosted MCP server at no additional cost, so Claude, ChatGPT, or Gemini can now pull balance sheets, review cash flow and aging, and route payment proposals for approval against live data. MCP is just a standard that lets an AI assistant call another system directly instead of a human copying between them. It's Salesforce-hosted, so most of you will never install it. The direction is the point: the ledger is becoming something an agent can call, and that's the precondition for everything downstream.
Quick hits
Three reminders that you're building on someone else's balance sheet. Thomson Reuters is retiring FileCabinet CS, with support and updates ending December 31, 2027, and its own bulletin warns it "can't guarantee perpetual access" to your documents after that. Vic.ai rebuilt its payment rails, and customers must now complete a banking application and open a new business bank account in their own name to keep using a product they already pay for. Baker Tilly pulled a roughly $3 billion leveraged loan that would have funded a dividend of up to $1 billion, after lenders wanted wider pricing. Vendors die, vendors migrate you, and the capital funding your competitors' roll-up has a ceiling.
Ambrook raised $30 million to be the ledger for businesses nobody builds for, growing from roughly 2,500 customers a year ago to more than 8,000 across agriculture, trucking, construction, and property management. Thomson Reuters Ventures is on the cap table, so the parent company backing a challenger ledger is also the one telling FileCabinet customers it can't promise access to their own files. Those are exactly the verticals where repeat work makes firm-specific automation compound, so the question is whether you build an advisory practice on top of that or wait to be told which ledger your clients moved to.
The only thing nobody gave away this week
Look at what actually moved. A vendor offered to take the first pass of your tax returns, a model resolved 10 open problems for about $2,000, and the free tier your clients use is about to get more accurate. Every one of those sits on the doing side of the ledger.
Nothing moved on the other side. Nobody shipped a product this week that will sign a return, carry the liability, or take the call when the notice arrives, and the regulator moved the opposite way. The labor data shows what happens at national scale when nobody decides: output rose, hours didn't, and the difference landed on the side of the table that sets the price. In your own practice, that side is you.
So the question isn't how much the machine did. It's whether you can show what you did after it, and whether you've decided where the saved capacity goes rather than letting your pricing model decide for you. Three days out of a close is worth real money to somebody. The only bad answer is that nobody chose.
That's what the AI Practice Transformation program is built around: not the tools, but the delivery model that turns saved capacity into something a client will actually pay for. Enrolment for the next cohort closes on August 21 at theaiaccountant.ai/transformation.