Your CAS clients are already asking a chatbot about their own numbers. Bill just measured what that does to their expectations of you: 82% of firm leaders say the bar moved. Here is the close that clears it.
Bill surveyed more than 200 accounting firm leaders for its 2026 AI Ambitions report, and 82% said AI has raised what clients expect from their firms. Speed tops the list at 79%, then better advice at 67%, data privacy at 66%, cybersecurity at 65%, compliance at 63%, and transparency at 51%. Behind the percentages: more rush requests, more fee scrutiny, and clients willing to leave when answers come slowly.
The speed number needs translating. Your clients are not timing your replies against other firms. They are timing you against the chatbot they used at lunch. A restaurant owner who asks Claude why her food cost jumped gets an answer in eight seconds. It is generic, and it is wrong about her rebate timing, but it is instant. When your explanation of the same line item takes eleven days to arrive in a PDF she has to log into a portal to read, the comparison does its own damage.
The wrong response is typing faster. The survey's own data points somewhere else: firm leaders say AI is what lets them expand services, into tax planning (50%), client advisory (39% moving in now, on top of the 38% already offering CAS), and business strategy consulting (38%). Kimberly Blascoe at CPA.com framed it cleanly: compliance and operations work generates the data, advisory turns that data into guidance, and AI closes the gap by making tailored advice scalable.
For a CAS practice the concrete version is the monthly package. Statements plus a standing offer to call with questions is now the thing a chatbot can imitate. What it cannot imitate is context: the covenant sitting at 1.19 against a 1.25 floor, the revenue dip that is milestone-billing timing rather than churn, the owner who needs to hear "hold off on the second truck" from someone who knows there is a truck. The deliverable that survives is the narrative, why it moved and what to do next, landing days after close instead of weeks.
So treat the 79% as a cycle-time problem, not a typing problem. The distance between "books closed" and "client understands the month" is the number to shrink, and most of that distance is waiting, not working. This issue's Workflow Lab takes the biggest slice, the commentary draft, down from hours to a review pass.
Anthropic's July update to Claude for Financial Services added Excel plug-in connectors for S&P Capital IQ, Morningstar, PitchBook, and Daloopa, plus write access to Outlook, OneDrive, and SharePoint. The May agent templates, including a month-end closer and a GL reconciler, now have more to work with inside a Microsoft 365 tenant.
Accounting Today's July AI Thought Leaders survey asked which processes get automated this year. The consensus: transaction-level work goes to the machines, while judgment calls like revenue recognition structuring stay human, because AI has no professional skepticism and no license on the line. Staff accordingly.
Smart Communications' June benchmark found 82% of consumers worldwide say it matters that companies disclose when AI is part of the interaction. If Claude drafts your client emails or close commentary, one disclosure line costs nothing and feeds the transparency expectation the Bill survey put at 51%.
The Rise 2040 vision project got top billing around AICPA ENGAGE this summer: a profession-wide look at what accounting becomes over the next fourteen years, with AI at the center of it. CPA Practice Advisor's Technology Lab podcast has the debrief if you want the early read on where the profession's own leadership thinks this goes.
Read it next to the 67% who say clients also expect better advice, and the trap becomes visible: clients want faster and smarter at the same time, which rules out both heroic overtime and unreviewed AI output. The firms that clear the bar are changing the cadence of the deliverable. The firms that meet the speed demand with something a chatbot could have written will discover their clients can get that for free.
What it is: An AI-native general ledger built for firms that serve startups and lean small businesses. Double-entry accounting rebuilt around agents that draft categorizations, reconciliations, and month-end close steps, with what Puzzle calls governed automation: nothing posts to the ledger until a human approves it. Native feeds from Stripe, Mercury, Brex, Ramp, Gusto, and Bill.com, with wider bank coverage through Plaid.
What it does well: The close. The agent drafts, you review, and the cycle compresses; firms running client books on it report cutting close time by as much as half. Keeping cash and accrual views live on the same ledger is the quiet killer feature for CAS work: founders think in cash, covenants are written in accrual, and Puzzle stops you from maintaining two workbooks to bridge the difference.
What it doesn't do well: General practice. Inventory-heavy clients, complex multi-state payroll, or anyone welded to a QuickBooks app stack will not map cleanly. AI actions meter through a credit system with small allowances on the lower tiers, so model the real monthly cost per client before you move a book. And it is young: review the agent's work the way you would a first-year's, which is the point of governed automation but still takes minutes per close.
Pricing: Free entry tier for low-volume entities, then roughly $25 to $150 a month per entity as features and volume climb, with AI credits metered on top. Cheap enough to pilot on one startup client this quarter and time the close against your current stack before deciding anything.
Anthropic ships a month-end closer agent template now, aimed at enterprise finance teams with Claude wired into their systems. Most CAS shops need the useful half of that, not the infrastructure. I get it from a plain Claude Project, one per client, and it drafts the monthly commentary before the coffee is done. Create a Project named for the client, then put the things you would brief a first-year on into the Project instructions: variance thresholds worth flagging (say 8% or $2,500, whichever is smaller), chart-of-accounts quirks, billing timing that routinely fakes a trend, the covenant or KPI watch list, and how the owner likes to be spoken to. That context is the entire advantage. It is exactly what the client's own chatbot does not have.
(1) Drop in the trial balance or P&L export (QuickBooks, Xero, or Puzzle) plus last month's commentary.
(2) Prompt: "Draft this month's variance commentary. Flag every account past threshold. Where the cause is visible in the data, propose a one-line explanation. Where it is not, write [CHECK] and a question for me."
(3) Answer the [CHECK] lines yourself. That is the advisory part.
Why the [CHECK] convention matters: Claude is good at arithmetic on what is in the export and dangerously fluent about what is not. Forcing it to separate "the data shows" from "ask the human" keeps the draft useful and keeps your name on the judgment. My review pass runs about twenty minutes against the two hours a blank page used to take.
Standing rules: client-identifying exports go through a Team or Enterprise workspace, tax-software parity as always. And read every number against the source before anything leaves the building. A confident draft carrying one wrong assumption is worse than no draft at all.
Time the gap between "books closed" and "commentary sent" for one CAS client. Then build the Project above for that same client and run it on the July numbers you are closing right now. Compare the clock. Setup is thirty minutes; if the draft does not hand you back at least an hour, reply and tell me where it fell short, because that is worth knowing too.
P.S. What did a client ask their chatbot before they asked you? Reply with the question, no client names needed. I am collecting them for a future In Practice on what clients actually do with AI between meetings.
