Karbon put the same question to business clients four different ways this week and got four different answers about money. Make the existing work faster and 37% of them want a fee cut. Deliver something they have never had before and 54% will pay more. That gap arrived in the same week the model layer got 80% cheaper and both major ledgers put their prices up. The saving is real, and it is not landing on your invoice.
Your clients will pay for capability. They won't pay for speed.
Karbon published The Future of Client Trust in the Age of AI on July 27. Be clear about what it is: a practice-management vendor publishing research that supports its own product story. Read it anyway, because the design is the clever part. Four scenarios, all the same technology described differently, each asking what should happen to the fee.
Scenario one: AI makes your existing work faster. Clients split three ways: 21% would pay a premium, 42% say the fee should stay the same, and 37% say it should come down. It is tempting to add the last two together and report that 79% want flat or falling fees, and that would be wrong. Karbon reads the 42% as a value-pricing position, clients paying for the result rather than the hours behind it, which is the opposite of the fee compression we have been tracking all year. The honest headline is narrower: when AI only makes existing work faster, 37% expect a discount.
Now read that 37% next to something the IRS already published. In June the Office of Professional Responsibility issued its first guidance on AI in federal tax practice, and on fees it is blunt: cost savings and administrative efficiencies from AI "should be passed on openly," with billing practices that fairly credit the reduction to the client's account. Billing manual hours you did not actually spend can be an unconscionable fee under Circular 230 section 10.27. That is narrower than it sounds, because it binds federal tax practice rather than all of your CAS work. But it means a third of your clients are now asking for something your regulator has already written down.
Here is where I will go past the data. Karbon caught clients at a moment when AI speed is still novel enough to be worth arguing about, and that will not last. My read is that speed stops being a premium and stops being a discount, and becomes the price of being considered at all: the client will not pay extra for the three-day close, they will simply stop tolerating the three-week one. The survey cannot see that yet, because it asked people to put a price on a change they are still noticing. Table stakes are invisible by definition, so build for the version of this where nobody thanks you for being fast.
Scenario two changes the answer completely. Deliver something the client never had before, real-time dashboards or monthly cash flow forecasts, and 54% will pay more. Scenario three goes further. Give the saved time back as attention, where AI handles the routine work and you call every month for a strategy session, and 81% say you are more valuable. Karbon's own framing is the line worth stealing: clients want the roughly 21 hours a month AI hands back "returned to them as attention rather than as a discount."
One finding nobody else will pull out of this has a deadline attached. The window for charging for new capability closes with client tenure: 74% of start-ups would pay more for something they have never had, against 39% of mature businesses. What reads as an upgrade to a younger client is a baseline expectation to a long-tenured one, which is the same table-stakes drift arriving early. So sell forecasting or a live dashboard to your newer and mid-market clients first. Your twenty-year clients are the hardest sell in the book.
Open weight models just made the capability 99% cheaper, and your clients are already using it
Three things happened in the model layer in five days, and they only make sense together.
First, Kimi K3's weights actually landed on July 26. Open weights just means a model anyone can download and run on their own hardware instead of renting it through a vendor. K3 now edges out the closed frontier on two benchmarks shaped like our work: 33.4 against Claude Fable 5's 26.8 on banking tasks, and 34.8 against 34.7 on spreadsheets. Nobody reading this is going to self-host it, and that is not the point. The point is what a free, benchmark-competitive model does to the price of every paid tier above it.
Second, three days later, OpenAI answered. On July 30 it cut the price of Luna, its mid-tier model, by 80%, from $1 and $6 per million tokens to $0.20 and $1.20. Read its own claim carefully, because it is the sentence that matters for firm economics: on a professional agent benchmark, that cheap model now beats Claude Fable 5 at an estimated cost per task nearly 99% lower.
The practical read is that "which model should we use" is now the wrong question. OpenAI says the right one out loud in its own release: use the expensive model to resolve uncertainty and define the plan, then use the cheap one to execute the well-specified work. Applied to a practice, the expensive model writes the procedure and reviews the exception, and the cheap model does the volume. Most firms are currently paying frontier rates for both halves of that.
Third, and this is the one to sit with. OpenAI published its own usage research on July 27, drawn from more than 800,000 US workplace messages, and 43.5% of occupation-specific use turns out to be work that belongs to somebody else's job. Financial calculation is among the three most common outside tasks in every one of the seven non-finance groups measured, and the crossover runs higher in small organisations than large ones.
Your clients are not planning to replace you. They are doing small pieces of finance work themselves, at the moment the question arises, because the handoff is no longer necessary. That is the erosion of the entry point rather than a capability threat. If your advisory relationship was sustained as a byproduct of those small calls, it now has to be designed in deliberately.
Both ledgers raised their prices in six weeks, and both blamed AI
Intuit's increase takes effect today, for renewals on or after August 1. It was announced back on June 26, and the reported plan moves are steep: QuickBooks Online Plus from $99 to $140, and Advanced from $200 to $340. Confirm those figures against Intuit's own price table before quoting them to a client, because the numbers are not in the text of the announcement. Xero announced its own rise late last month, effective October 1, at a gentler 8% or so across Early, Growing, and Established.
Both cite AI investment as the reason. So here is the squeeze from a direction we have not covered: not the client pushing your fee down, but the platform pushing your cost of delivery up, while the model layer that supposedly drove the increase just got 80% cheaper.
The buried item in Intuit's release matters more than the price. QuickBooks Advanced now ships Continuously Clean Books, where transactions are "categorized, verified, and posted according to your business preferences" and Intuit's own experts "perform regular quality checks." That is the demonstrable quality control layer we have spent five months naming, now sold inside the subscription. Note the activation language: "For client-billed subscriptions, this service is ready to turn on immediately. If you are on a firm-billed subscription, please coordinate with your accountant."
Three things to do about this before October. Run a stack cost review across every client on either ledger. Write down the rule for how platform price increases flow into client pricing, so you are not improvising it one client at a time. And make a decision on Continuously Clean Books now, before a client-billed subscriber switches it on without telling you.
Quick hits
Grant Thornton is buying CBIZ for $5 billion. All cash, $55.00 a share, roughly a 54% premium, private equity funding it, creating the fifth-largest US tax and advisory provider at about $7.5 billion globally and more than 34,500 people. It is the profession's largest deal since Price Waterhouse merged with Coopers & Lybrand in 1998. The read for a small practice is not "get bigger." It is that the competitive set your prospects compare you against just added a very large, very well-capitalised member with a technology agenda, and what survives that comparison is the relationship.
Basis and Braintrust published an open standard for supervising an agent. Basis is an AI-native accounting vendor, so this comes from inside the profession rather than from a lab. A behavior spec is a written file defining how an agent must behave, kept in version control, and the design decisions are the interesting part: plain imperative English, never shown to the agent, and when the spec and the live context disagree, the spec wins. That is the encoding loop with a file format, and a better skeleton for your own procedure files than anything we have published.
The close and audit stack went agentic in one week. Workday, BlackLine, Workiva, and Zuora all shipped agents into the close and audit layer, and Workiva went further, opening a gateway that lets clients bring their own AI while Workiva keeps control of permissions and data lineage. All four automate preparation, and Workday is explicit that audit testing is still future work, so the review residual holds another release cycle. The reconciliation is no longer the deliverable. The correction trail on top of it is.
Three walls of the vise moved in one week
We have been describing compliance fee compression as a four-way squeeze: the platform absorbing your work from below, the regulator directing AI savings to clients, price challengers undercutting you, and clients repricing the relationship on AI delivery. Three of those four moved at once this week. Intuit put quality-checked books inside the subscription and raised the price of the box they sit in. The regulator's position on passing AI savings through is already in writing. And clients told a vendor's researchers that if all you do is go faster, they would like the difference back.
Speed is the one thing in this week's news that nobody will pay you for, and it is what most firms are quietly counting on to justify their AI spend. My read is that it does not even stay neutral. It becomes the floor.
So the question for your own week is not how much time AI saved you. It is what you did with the hours. If they went into the same deliverable produced sooner, you have built the argument for your own fee cut and handed your client and your regulator the same sentence. If they went into something the client has never had, or a conversation they have never been offered, you have built the only thing this week's data says they will pay for.
That is the work the AI Practice Transformation program is built around: not the tooling, but the delivery model that turns saved capacity into something billable. Enrolment for the next cohort closes on August 21 at theaiaccountant.ai/transformation.