Weekly AI Roundup for Accountants: Your AI-native competitors now publish their prices

Weekly AI Roundup for Accountants: Your AI-native competitors now publish their prices

Collective raised $100 million this week to sell formation, bookkeeping, payroll, and tax filing to solo businesses, and its CEO put the pitch in one sentence: "Traditional accounting firms scale by adding people. Collective scales with intelligence." A day earlier, Inkle, which says it's "building the AI-powered Big Four," bought a bookkeeping division of about 100 people for cash. And the AI tools that do this kind of work now come with public prices: a $500-a-month add-on for AI agents (software that works through the steps of a task on its own, the way a staff accountant works a checklist) that close the books across every client, and tax prep from $60 a return.

This week: who the AI-native firms are and what their cost base now looks like, why Intuit Connect shipped its finished features to your clients while the firm tools stayed in pilot and beta, what Anthropic's own SEC filing and its own research say about the person reviewing the agent, and why the permission screen is the part of this week's practice-software releases worth reading.

1. The AI-native firms have names now, and the tools that do the work have prices

Collective's $100 million is a $40 million Series B led by General Catalyst, QED, and Gradient, with Intuit Ventures among the participating investors, plus a $60 million facility for acquiring customers. The release claims revenue up 120% in a year and says AI "resolves 50% of support requests autonomously," but gives no customer count and no price.

Inkle bought Finally's bookkeeping division, about 100 people, in an all-cash asset deal effective October 28. Inkle already has its own ledger and supports QuickBooks, Xero, and others, so what it bought was about 100 people doing bookkeeping. That's scaling by adding people, the thing Collective's CEO says traditional firms do, and it tells you the agents aren't doing the whole job yet.

When I wrote about the four forces squeezing the compliance fee in June, the third one, AI-native firms delivering the same service from a different cost base, had the least evidence behind it. There were few firms to point to, and none buying bookkeeping operations. Now there are names, with funding and acquisitions behind them.

The tools tell you what that cost base looks like. Double priced Autoclose on October 29 at "$500/month, as an add-on billed firmwide" on top of a Double subscription, for agents that pull statements, code bank feeds, post entries, and reconcile on a schedule, then "report back and show every step." Black Ore opened its tax autopilot to small firms on October 27, priced per return with "no five-figure minimum" and e-filing through ProConnect and Lacerte; Accounting Today reports pricing from $60 a return.

Those are tools sold to firms like yours, so the price list is available to you too. None of the AI-native firms has published a fee you can compare to yours, and none has shown it's winning clients from CAS firms. But you can now work out roughly what the software side of a monthly close costs anyone who builds the delivery model around agents.

Run your own numbers: what does a monthly client cost you if close agents cost a $500 add-on across your whole book, plus the subscription and the review time? Then decide whether your fee is for the close or for the judgment on top of it. If it's the close, stop selling it on its own.

One caution on Double. Its "agentic browsing" logs into bank and card websites with stored client credentials to fetch statements where no feed exists. Chase's updated online agreement, in force since October 27, prohibits "External AI, automated tools, or data aggregators" from using a customer's login credentials. Before you turn that feature on for a client, read their bank's terms and check what your WISP (the written information security plan the FTC Safeguards Rule requires of tax preparers) says about stored credentials.

2. Intuit Connect shipped to your clients. The firm tools are still in pilot and beta.

At its September 17 Investor Day, Intuit footnoted Agentic Books Close as "available in open beta at no charge; paid availability is expected in January 2027." The October 27 Connect release says it's "currently in pilot with select IAS customers. Additional fees may apply" (IAS is Intuit Accountant Suite), and Intuit's Firm of the Future site describes it as "in pilot and available to a limited group of firms; sign-up required." An open beta with a January date is now a limited pilot with no date.

Agent Studio, where firms build their own agents, is a pilot for about 20 firms, per Accounting Today. Playbooks, cross-client analysis, and the Intuit Intelligence chat are in beta, and the chat is admin-only. The Accountant Suite MCP server was demoed on stage with Claude pulling client ledger data, but it has no published documentation. (MCP is the open standard that lets an AI assistant read from and work in another piece of software, the way a bank feed lets your ledger read the bank.) Intuit also introduced formal release labels (pilot, beta, and general availability, meaning open to every customer), and noted that "Not all features will be available on October 26, 2026."

On the client side, invoicing from inside ChatGPT and Claude is now available to the public on QuickBooks Simple Start through Advanced, so a client can draft invoices from their own AI assistant. Vendor Direct same-day bill pay is live, and payroll migration from ADP or Gusto now comes with a Payroll Expert review. The firm's one clear gain is a price: the manual Books Close, still labeled beta, now comes at "no additional cost," where April's pricing put it at $8 or $6 per client.

Accelerate's free beta has been "extended through January 19, 2027," the first charge lands January 20, and from January the tier is free only to Premier and Elite ProPartner accountants. Neither of Intuit's Fall update pages states the price after that; April's figure was $149 a month.

So your clients got the finished features and your firm got the pilots and betas. Whether that's sequencing or slippage is Intuit's to answer, and nothing in the release answers it. Before January 19, decide whether Accelerate is worth paying for on what's available today rather than what's in pilot, and ask your Intuit contact which ProPartner tier you'll be in and what you'll pay on January 20.

3. The reviewer is only a check if the judgment is their own

Anthropic's own researchers ran an experiment this week that every firm reviewing AI-prepared work should know about. They gave 50 professional data analysts an analysis task to complete with Anthropic's AI agent, and secretly had the agent lean one way on four judgment calls, chosen by coin flip. The analysts went with the agent's lean 86% of the time, where chance would be about half. More than three-quarters repeated its arguments afterwards, and only 15% said the agent had influenced them.

That's your review process. When someone at your firm reviews an AI-prepared close, the agent hasn't just done the work, it has explained why, and a confident explanation is what pulled these analysts along. They weren't accountants, they were new to the subject, and the agent leaned harder than it usually would, so 86% isn't a number for your firm. The direction is the point, and the authors put it plainly: "a human reviewer is only an independent check if their judgment is their own."

Anthropic's IPO filing, made public on October 26, says the same thing in investor language. It warns that reviewers may suffer alert fatigue, may lack the context to judge what an agent did, or may not get to it in time, and that firms face pressure to give agents more autonomy for the productivity gains. It also says the law on who's liable when an agent acts is unsettled. That's Anthropic's exposure, not yours. Yours is settled: your firm's name is on the work, whoever prepared it.

A tax authority is heading the same way. The ATO said on October 30 that "AI told me" isn't reasonable care, with draft guidance due around early December. That's Australia, not the IRS, but it shows where the standard of care for AI-assisted work is going.

The audit habit of forming an expectation before you look works for an analytical review, but nobody writes down an expected answer for 400 coded transactions. For bookkeeping, the fix is to separate the agent's work from its argument. Check the coding against the source document and the client's prior months, not against the agent's note on why it chose that account. For the real judgment calls (capitalize or expense, business or personal, which period), have the agent flag the item and set out the options without picking one, so your reviewer makes the call. The study's authors recommend exactly that kind of design: an agent that asks for the expert's view first, or argues each side.

Then track how often your reviewers change the agent's work. If that number sits near zero month after month, treat it as a reason to look harder, not to relax. The provenance checks from September still cover the inputs, and they're the kind of check that catches fabricated references like the ones found in an A$166,000 Deloitte report on a Western Australian university merger.

4. Your practice software can now write, so read the permission screen first

TaxDome released a public API (the published route one piece of software uses to talk to another) and an MCP server on October 29. Firm owners choose which team members can connect, set read-only or read-and-write "for each area of TaxDome," and "a connection can never do more than the person who set it up." More than 200 firms ran the beta, rollout to all firms is "over the coming weeks," and no price was given.

Financial Cents added AI tools on October 28 that create, send, charge, and void invoices on its Scale plan, though the changelog doesn't say whether a person confirms before a card is charged. Puzzle's MCP went live, read-only for now. Kick added check and deposit capture, free, where the assistant proposes the GL account and the user reviews it before it posts.

Ignition launched Parkes, an AI admin assistant that finds underpriced and out-of-scope work and drafts repricing and collections, with firm-level settings of auto, approve, or disabled and a full action log. It's open to firms that don't use Ignition.

Two weeks ago the roundup's test for a ledger connector was whether it could write. Practice management, billing, and capture can now write too, which makes the permission model the decision. TaxDome's rule has a consequence: if the firm owner sets up the connection, it can reach up to every client the owner can see.

Start read-only everywhere and turn on write access one area at a time, where there's an action log and a scope limit. Set connections up from a staff account scoped to the clients and areas the work needs, not from the owner's login. And don't give an assistant the ability to charge a client's card until you know whether it asks first.

Quick hits

Copilot for finance reads your clients' ledgers but doesn't write to them. Microsoft's October 29 finance release puts accounting skills into Excel and Cowork (Copilot's agent workspace), in its early-access Frontier program, that compare balances, prepare reconciling items, and draft journal entries in the workbook. The partner plugins are read-only where Microsoft documents them: Xero's lets you "Check your P&L, cash position, and top customers," and Mercury's "Check balances, transactions, and statements." QuickBooks was named in the announcement but isn't on Microsoft's plugin list yet, and your IT admin has to enable any plugin before staff can use it. Nothing Microsoft documents posts back to the ledger, so treat it as review prep in Excel.

Rotate the keys, update the phones, and plan for February 1. LiteLLM, an open-source tool developers use to route requests to AI providers, had a compromised release, version 1.107.0, that carried a credential stealer on October 21. If anyone who builds tools for your firm installed it, LiteLLM's instruction is to rotate "every provider key stored in the LiteLLM proxy DB" (the keys that let a tool spend on your paid AI accounts). Microsoft Authenticator for Android had a token-theft flaw rated 8.6 out of 10 (fixed in version 6.2606.4246), and Microsoft's own SMS and voice codes for Entra ID (the sign-in behind Microsoft 365) retire for most users from February 1, 2027, so update staff phones and move everyone to passkeys or the Authenticator app before then. In Rightworks' survey of more than 500 firms, 81% have a WISP but only 44% can remote-wipe a lost device.

CBIZ and Grant Thornton are set to close today. CBIZ shareholders approved the merger with Grant Thornton on October 27, 96.9% in favor, at $55 a share, with closing expected November 2. CPA Trendlines counted 37 firm acquisitions in the third quarter against 42 a year earlier, so consolidation continues at roughly last year's pace. Read it next to Inkle: firms are still buying firms, and the AI-native firms have started buying bookkeeping operations too.

Wolters Kluwer still hasn't given a date. At CCH Connections on October 26, Wolters Kluwer added fixed-asset depreciation flowing into the return, an AI tax planner, agentic audit risk assessment, and K-1 handling to last week's agentic 1040, and again gave no date, availability label, or price for any of it. Last week I suggested CCH Axcess firms might get those answers at Connections. They didn't, so keep asking your rep.

What this week adds up to

The software side of the work now has a published price: a $500-a-month add-on for close agents across a firm's whole book, tax prep from $60 a return, while AI-native firms raise money and buy bookkeeping operations. What doesn't have a price is the judgment and the accountability on top, and this week's evidence says that layer is fragile unless you design it. The vendor's own filing says reviewers may suffer alert fatigue, and its own study shows reviewers adopting the agent's view without noticing. If a competitor can run the software side of your clients' closes on a subscription plus a $500 add-on, what exactly is your fee paying for, and could you show a client the review that proves it?

If you want to work through what your firm's fee is for once the software side has a price, I'm happy to talk it through with you. Book a free consultation at theaiaccountant.ai/consultation.