This week, two of the biggest players that touch your practice — the IRS and Intuit — made a move on the same thing: the savings AI is creating inside your firm. One says those savings belong to your clients. The other is quietly raising its prices to keep them. If you price your work as a markup on software and hours, the ground just shifted under that model.
The IRS just decided your AI savings belong to the client
On June 24, the IRS Office of Professional Responsibility issued its first guidance on AI in tax practice (Alert 2026-19). Most of it is sensible: you stay responsible for AI output, you verify it before it goes out, you keep client data off public systems. The line that matters for pricing sits in the fee section. OPR reads Circular 230's ban on charging an "unconscionable" fee (§10.27) to mean that "any cost-savings and administrative efficiencies derived from [AI] should be passed on openly," and that you must "fairly credit to the client's account any cost reductions." Billing full manual hours for work AI did in seconds, or double-billing for AI-assisted tasks, "may violate § 10.27."
Two clarifications. This is OPR's interpretation, not the regulation itself — but OPR enforces Circular 230, so it's the enforcer telling you how it reads the rule. And the language reaches past the hourly biller: "credit to the account" doesn't care about your billing model, so "I charge fixed fees" isn't automatic cover. The market is already moving the same way — Ignition data this month shows the billable hour is now the minority in tax prep.
What this means for you: don't assume a fixed fee puts you in the clear — that's where the trap hides. The expectation OPR just put in writing is that AI's savings flow to the client, and that expectation doesn't care how you bill. It also doesn't stop at the regulator: clients read this too, and "the IRS says AI should lower the cost" becomes a reason to ask why their invoice hasn't moved. Your defensible position isn't a billing model; it's being able to say what the fee buys beyond the production time AI just compressed. We opened a three-part series on exactly this squeeze last Friday.
Intuit wants those same savings — going the other way
At its Connect ON event on June 23, Intuit CEO Sasan Goodarzi told accountants "you are the customer, not a channel" — a warm reframe after years of treating firms as a route to small-business customers. Read that warmth next to the pricing news from the same broadcast: Intuit is raising prices on QuickBooks Essentials, Plus, and Advanced "to reflect the value" of the AI it has built in. So in one week, the platform captures AI value by charging you more for it, while the IRS tells you to hand your AI value to clients. That's the vise — squeezed from below as AI compresses the cost of the close, and from above as the platform raises its rake.
The question Connect ON forces is story one's question, sharpened: if part of what you bill is a markup on "software fees," and the software keeps getting more capable and more expensive, what exactly are you charging for? There's a bigger platform-consolidation story underneath — QuickBooks Online Accountant sunsets at the end of 2026 and every firm moves onto Intuit's AI-native suite — which we take apart in a full-length piece on Wednesday.
Claude went multiplayer — and it lives in your group chat
Anthropic launched Claude Tag, which drops its AI assistant into Slack as a shared teammate. Until now, working with an AI assistant has been a private, one-to-one affair — you, a chat window, your own files. Claude Tag makes it multiplayer: anyone in a channel can tag the AI into a task, watch it work in the thread, and pick up where a colleague left off. It runs on Anthropic's servers rather than your computer, it's in beta on the Team and Enterprise plans, and it keeps a shared memory of the channel's work.
For a firm, the interesting part isn't "AI is in Slack now." It's that the unit of AI work moves from one person's private window to a shared, visible team surface — which is exactly where firm knowledge usually lives or dies.
Two cautions before you try it. It can't see your local files — the thing that made Claude Cowork useful for accountants — so everything arrives through cloud connections. And its memory is shared across channels by default, a client-confidentiality problem if you're not deliberate, the same scoping issue we worked through in the access-paradox roundup. Our read: worth a contained internal pilot in a channel with no client data, not yet ready for client work. It's the first release, and it will change.
When any receipt can be faked, whose job is it to catch it?
A new Emburse survey found about 40% of US employees have used AI to generate a fake expense receipt — invent a purchase, inflate a real one, or recreate a "lost" one. The detection data is worse: by mid-May, AI-generated fakes made up roughly 71% of all flagged fraudulent documents, according to AppZen. These aren't clumsy forgeries you'll catch by eye. The credible defenses — Ramp's, for instance — read a file's metadata for AI fingerprints rather than looking harder at the image.
Here's the question for your practice. You've historically processed expenses and invoices, not approved them — you record what the client hands you. But when a convincing fake takes 10 seconds to make, clients may quietly assume their bookkeeper is the backstop, whether or not that was ever the deal.
The work is shifting from producing the record to standing behind it — faster than the conversation about who's responsible for it. So have that conversation now, explicitly: where does verification responsibility sit, for expense reimbursements and supplier invoices alike? Catching what's wrong, and being clear about whose job that is, is becoming the service.
Also worth tracking
Washington has started gating which models you can use. The White House asked OpenAI to limit its next model, GPT-5.6, to a small set of government-approved partners — the first time the US has restricted an American model before release. Nobody compelled OpenAI; it agreed, apparently to avoid the kind of export-control order that forced Anthropic to pull its Fable 5 model earlier this month. That's the part worth noting: one enforcement action has produced voluntary restraint across the field — effective government oversight where none officially exists. The takeaway is the one we keep returning to: if your workflow depends on a single frontier model, whether it's available next week is no longer entirely up to the vendor. We covered the supply-chain version here.
The model carousel kept turning. Google's Gemini 3.5 Flash went generally available with "computer use," and OpenAI retired GPT-4.5. Computer use is the part worth flagging: letting a model click around a screen the way a person does is a real step toward AI that operates your software directly, not just answers questions about it. And it points to the deeper pattern — these models converge. A capability one of them ships this quarter shows up in all of them before long, so the tool you standardized on last quarter is rarely the best one this quarter.
The week in one question
Notice the through-line. Every story this week is a fight over the value AI creates and who keeps it — the savings, the trust, the verification, the firm knowledge. The IRS says the savings belong to clients; Intuit says they belong to the platform. The receipts story says value is moving from making the record to verifying it, and Claude Tag says the work is moving into shared spaces you now have to govern.
The firms that come out ahead won't have the best tools — they'll be the ones who can say, out loud, what their clients are actually paying for. So answer it: strip out the software markup and the hours, and what's the one sentence that explains your fee? If you can say it without flinching, this squeeze is survivable. If you can't, that's the work.
That's the work we do in the AI Practice Transformation program — turning that one sentence into a fee model that holds. The next cohort starts July 10; if you want help, join us here.