Weekly AI Roundup for Accountants: Intuit put QuickBooks reconciliation tools in Claude and hasn't announced it

Weekly AI Roundup for Accountants: Intuit put QuickBooks reconciliation tools in Claude and hasn't announced it

Intuit has put a second QuickBooks connector in Claude’s marketplace, and this one lists tools to reconcile bank accounts, post journal entries, and switch between client books in a single conversation. Intuit hasn’t announced it. This week TaxDome also stopped selling per-seat plans and started charging by the client, and it shipped the first tool I’ve seen that lets a client’s own AI assistant talk to their accountant’s practice software.

Below: what the new connector lists and what’s still unknown, what TaxDome’s per-client pricing does to a firm that grew without hiring, why your client’s agent (an AI assistant that can go and do things for them, like checking a status, rather than only answer questions) can now reach your firm, and how firms are starting to cap their AI spending. Then six quick hits, from audit to the data-center vote.

1. QuickBooks reconciliation now works inside Claude, unannounced

Anthropic’s marketplace of Claude connectors now lists “Intuit QuickBooks Labs (Beta),” published under Intuit’s name and marked “Added October 2026.” (A connector is the plug-in that lets Claude read from and work inside another program, the way a bank feed lets your ledger read the bank.) The listing describes 120 tools, each one an action Claude can take, and the description reads like a bookkeeper’s task list: “categorize and match transactions as they arrive, reconcile each account to its statement… record bills, expenses, payments, deposits, transfers, and journal entries.” It adds: “If you manage more than one company, including client books, switch between them in a single conversation.”

The tools include a full reconciliation set (begin, mark items cleared, finish, undo, and history), bank-feed categorizing and matching, switching between companies, and a tool named for closing the books. Reconciliation and bank-feed matching have never been in QuickBooks Online’s public API (the published route other software uses to work in your ledger), which is why outside tools could prepare a reconciliation but couldn’t finish one inside QuickBooks. The original Claude connector, live since April, still can’t reconcile, and Intuit’s own open-source connector works with one company at a time.

Everything else is missing. There’s no Intuit release, blog post, or help article, and Anthropic hasn’t announced it either. There’s no price, no published word on who’s eligible, no date for leaving beta, and nothing on how Claude’s reconciliations show up in the audit log. Public snapshots of the marketplace place its arrival between October 16 and October 30, a window that included Intuit Connect, where, according to AccountingWEB, Intuit demoed Claude pulling client ledger data through an “Accountant Suite MCP server” (MCP is the open standard these connectors are built on); nothing yet says whether that demo and this listing are the same thing.

Three weeks ago I wrote that a ledger connector has to write and switch between clients before it’s much use to a firm, and that Digits had done both. I’ve now connected QuickBooks Labs myself, and it does everything the listing says, including the part that matters most to a firm: one connection that moves between client books. Reconciliation, bank-feed matching, and a multi-client connection in one place is exactly what accounting firms need, and it’s exciting to see these functions finally open to an AI assistant.

The catch is where it lives. QuickBooks Labs is a beta connector listed only in Claude’s marketplace, and the reconciliation functions still aren’t in QuickBooks Online’s public API; I checked Intuit’s own developer catalog. Firms build on more than one assistant and more than one platform, so these functions need to be available through the public API and standard MCP servers that any tool can use. I appreciate that this is an experiment, and I hope it’s the start of these functions becoming common rather than a Claude-only feature.

It’s also a direct challenge to Xero, which has said it won’t do this. In May, Xero’s developer team wrote that “we will not be adding the ability to reconcile bank statement lines via the API,” citing a policy decision not to act as an intermediary for raw bank statement data, and the request, with 284 votes, has been marked declined since 2017. Xero, your biggest competitor just did it. You are now officially behind.

Intuit’s own lineup pulls two ways. Its Agentic Books Close, which Intuit says handles reconciling accounts, is in pilot, “free during the pilot, and a paid add-on is planned afterward.” Meanwhile the reconciliation controls sit, unpriced and in beta, inside a general-purpose assistant. That’s platform absorption, the first of the four forces I’ve argued push the compliance fee down: if a client can ask Claude to reconcile their own books, your monthly fee has to stand for more than the reconciliation.

Decide now who in your firm may connect it, because nothing stops a staff member with a Claude plan and client access from doing it today. Try it on a test company first, and before you let it near live client books, compare the reconciliation report and audit-log entries from a Claude reconciliation with one done in the QuickBooks screen, because Intuit hasn’t said how the two differ.

2. TaxDome now charges by the client, and your own AI’s access is extra

On November 2, TaxDome stopped selling per-seat plans after nine years. The new plan is $1,500 a year for your first 150 active client accounts and up to 20 team members, with each extra account at $15 as it becomes active, or $6 to $12 each if you buy a pack at renewal (the bigger the pack, the lower the price). New firms are on the new plan now, and existing firms move at their first renewal on or after January 1, 2027.

A client counts as active in a plan year when you invoice them, open a job, or send a chat, or when they log in or upload, and a household counts once. Work inside a client stays unlimited; what’s metered is your client count.

That changes who pays. TaxDome’s old plans ran $800 to $1,200 per seat a year, according to Accounting Today. On the $1,000 tier, a 10-person firm with 300 active clients paid $10,000 and would pay $3,300 at list price after its first renewal in 2027 ($1,500 plus 150 accounts at the $12 pack price), while a three-person firm with 600 active clients paid $3,000 and would pay $6,225 ($1,500 plus 450 accounts at $10.50). Both figures are before any discount and leave out the AI add-on. TaxDome’s Ilya Radzinsky told Accounting Today about a 50% discount on required accounts for smaller firms, but no TaxDome page states it, so don’t count on it until your renewal quote does.

In August I wrote about the pricing wrapper nobody reads: per-client fees make you weigh every small client against the fee, and per-user fees punish the part-timer and the admin seat. TaxDome has traded one for the other. With up to 20 team members included, the seasonal preparer and the admin who only touches billing no longer cost a seat, and Accounting Today reports TaxDome sold more than 40,000 seasonal licenses last year. The small-client question comes back in a new place: because sending a chat or opening a job makes a client active, the question becomes whether a lapsed client is worth a $15 reminder, which is small money but the same adjudication.

Seat pricing rewarded the firm that grew its client list without hiring, which is what AI is supposed to let you do, and client pricing takes that reward away. Radzinsky told Accounting Today that seat pricing “has effectively been deciding who got to be on a platform,” and TaxDome’s argument is that under seats one firm “pays 40% more than the other” for the same client. He also said the change wasn’t designed to raise revenue.

The AI moved separately, and it didn’t wait for renewal. For US firms, a $200-a-month AI add-on now covers AI document renaming and tagging, AI help in tax return delivery, and API and MCP access, which is the route your own AI assistant, CRM, or notetaker uses to reach your TaxDome data. TaxDome’s help center says that starting November 5, those features need the add-on or a free trial, so a firm that had document renaming switched on under its old plan now pays to keep it. AI checklist matching, AI checklist generation, and tax return delivery itself stay in the base subscription, and TaxDome says Atlas, its built-in AI agent, is coming to the add-on.

Put that next to the next story. Connecting your own AI assistant to your own firm’s data in TaxDome now costs $2,400 a year on top of the base plan, while your client’s assistant connects to their side of your portal for free. Marc Staut of Boomer told Accounting Today this week that firms are “noticing which vendors now charge more for full API or MCP access to the firm’s own data,” and TaxDome is now one of them.

The add-on is also where August’s other question applies: who carries the cost risk, and who holds the option to reprice? A flat $200 a month per firm, “every team member included,” with “technical rate limits only” on API and MCP, means TaxDome carries the cost of how heavily you use it and keeps the right to change the price, and with Atlas on the way, what the add-on contains will change too. A firm that wires its own agents into TaxDome pays less per use the more it uses it, while a firm that only wanted document renaming back pays $2,400 a year for that.

So what does $2,400 a year buy? TaxDome’s tool catalog shows an assistant that can run the pipeline: tell you which jobs are overdue and what each one is waiting on, create and move jobs, deliver a finished return, and chase clients by chat, email, or text. For now it’s a dispatcher, not a preparer: it can find every return waiting on a signature and chase it, but it can’t open the W-2 a client uploaded or carry its numbers into your tax software, because the catalog says files “can’t be downloaded or read yet” and email bodies aren’t available either. It can’t create an invoice or an e-signature request. And there’s no test mode, so moving a job runs that stage’s automations for real, including, in TaxDome’s own words, a real invoice “payable with a live card” or an engagement letter to real signers.

At that price, for that functionality, I think the value is questionable. TaxDome doesn’t prepare returns, so the hand-off a tax practice most wants automated is getting the client’s documents into the return, and that’s the one the MCP can’t do. For a firm with fewer than about 225 active clients, the add-on also costs more than the plan itself. If you want to try it anyway, the add-on bills monthly and you can cancel it any time, so run the 14-day trial on a test client in January and pay only for the months of tax season.

Run your TaxDome numbers this week: count last year’s active clients (households once), subtract the 150 included, price the rest at the pack rate for that volume, add $2,400 a year if you want the AI add-on, and compare the total with your current bill. Packs start at 50 accounts and unused ones expire at renewal, so size the pack to a count you’re sure of. Check your renewal date, because a renewal that falls before January 1 keeps your current plan for the next term. And if your team relied on document renaming, check whether it switched off on November 5.

3. Your client’s agent can now reach your firm

In July I wrote that your client’s agent can reach their bank, but it can’t reach you. Everything in a client’s financial life was becoming reachable by their AI except the firm, which still answered through a portal login, an email, or a wait. This week the second half stopped being true for TaxDome firms, and the first half got easier.

On November 4, TaxDome released a Client MCP, which TaxDome calls “an industry first”; it’s the first I’ve seen of a practice platform letting a client’s own AI talk to the firm. Your client connects TaxDome to their own Claude, ChatGPT, or another assistant that supports MCP and asks what’s still outstanding or whether the return is ready, and signing and paying hand off as links. One connection covers all of a client’s entities, and it’s free for every firm, “independent of the firm’s AI add-on.”

That’s the contact I described in July as the kind nobody enjoyed in the first place, like “Can I get a copy of last year’s return?” Now the client’s assistant can ask at 2am and get an answer, without a staffer hunting for a document the client already had access to. Ryan Bakke, a CPA quoted in TaxDome’s release, puts it in tax-season terms: “The slowest part of every tax season is waiting on documents, and now we help clients set this up and what we need just shows up.” The release names reads and links rather than uploads, so read that as the client’s assistant telling them what’s missing, not sending it.

The bank side got easier the next day. On November 5, Anthropic opened Personal Finance in Claude, a beta for US subscribers on its individual Pro, Max, and Ultra plans that reads up to two years of bank and card transactions through Plaid (a service many finance apps use to connect to bank accounts). It’s read-only, and Anthropic’s help center allows a second login at the same bank “such as a joint or business account.” It isn’t offered on Team or Enterprise plans and can’t run in scheduled tasks, so for now it’s a conversation the client starts, but Anthropic says people are best served by Claude “working alongside the licensed professionals and regulated firms they trust.”

In July I said a scoped connection that “reaches only what it’s allowed to and logs every request” is, done properly, the safest channel a firm has. TaxDome has done the scoping: a connection “sees only what the client already sees.” The release doesn’t yet say whether you get the log, can switch it off, or are notified when a client connects. July’s second question, who owns the front desk once a vendor builds it, has a first answer for TaxDome firms: TaxDome built this one and set its scope, and hasn’t said what control you get.

If you’re on TaxDome, put those three questions to your rep in writing now. Once you have the answers, pick a handful of clients who already use Claude or ChatGPT, help them connect before tax season, and track which questions stop arriving by email. If you’re on another platform, use July’s filter at your next renewal: who gives my client one front desk an agent can actually reach? And keep the question from part 2 in view: once the agents handle the friction, decide which conversations you still want to have yourself.

4. Firms are capping what AI costs. None of them measure what it produced.

Accounting Today got five firms to talk about AI spending this week, and two put numbers on it. Charlie Apigan at LBMC runs a $25-a-month soft cap per person on his team, which he raises in steps, up to a $500 hard cap: “what if that costs $100? And it [produced] nothing, so we want to stop it at 25 bucks.” Marc Staut of Boomer described a CIO who “caught a single user running up $600 in AI costs in one day by iterating on the same deliverable,” and the same task costing “50X more on one model than another.” PKF O’Connor Davies and Wiss are absorbing the cost, and Wiss is holding off on a tech fee because “the clients are not ready for it.”

The vendors moved their meters too. Microsoft delayed default usage-based billing for new Copilot Business subscriptions bought through partners from December 1 to a rollout between January and March 2027, with a default cap of 4,000 Copilot Credits per user per month, which works out to $40 at the one cent per credit in Microsoft’s Credits Guide. Anthropic added project tags to Claude Enterprise in beta, so usage can be grouped by a cost code such as a client or matter number.

Every control on that list caps or tracks spending, and none of them measures what the spending produced. That second number is the one your fee depends on. A $600 day is a problem if it produced a deliverable nobody used and a bargain if it replaced two days of a senior’s time. What you need is cost per accepted deliverable, by client: what the AI cost for that client this month, against the work that went out without being redone, and none of the firms quoted has put a number on it.

If you’re on Claude Enterprise with usage billing, tag projects by client, keep a simple count of deliverables accepted without rework, and look at the ratio after one month-end; tags count usage-billed spend, not what your seats already cover. If you’re on seat plans, log AI time against the client in your time tracking instead. If you’re buying new Copilot Business seats through a Microsoft partner, decide before January whether to leave usage billing on and where to set the cap, rather than letting a default decide it.

Quick hits

Audit is saying it too: the reviewer can’t hand judgment to the tool. Last week I wrote that a reviewer is only a check if the judgment is their own, and the audit side has now said it in its own words. In the Journal of Accountancy, AICPA chief auditor Jennifer Burns said “we can’t abdicate our responsibility to AI,” Carl Mayes said “Entry-level staff are relying on the output of the bot without challenging it,” and a draft AICPA practice aid on AI says “the auditor retains responsibility for supervision of the tools.” BDO’s survey found 99% of finance leaders use AI for at least one estimate their auditor reviews, while 49% can document all of their AI-assisted outputs. The AICPA’s SOC 2 exposure draft proposes AI governance, prompt manipulation (instructions hidden in a document or email), and data poisoning (corrupted data an AI learns from) as new points of focus, so expect AI controls to start appearing in the vendor SOC 2 reports you rely on.

The AICPA’s advice on selling your files for AI training: decline unless you can clearly show it’s permissible. “Offers to buy, receive, or otherwise use firm data for AI model training are rising,” the AICPA’s Professional Ethics Division staff wrote in the Journal of Accountancy on November 2 (training means feeding data to an AI so it learns from it), and two days later Going Concern printed one such pitch, aimed at firms “with at least 10 employees.” The staff guidance isn’t a blanket ban, but it sets a default of declining unless you can clearly establish the deal is permissible, and it closes the usual escape route: “redaction does not, by itself, turn confidential client information into nonconfidential information.” For US firms, Section 7216 adds possible criminal exposure, and the guidance warns against assuming an ordinary engagement letter is enough. The simplest way to meet that default is one line in your firm’s AI policy: no sale of client data for AI training.

Prosecutors say a $129 million fraud used fake executive voices and fake lawyers. On November 5, federal prosecutors in Manhattan charged three men over a scheme that emailed employees about “a confidential financial transaction,” then mimicked the CEO’s voice or image on calls while co-conspirators posed as law-firm partners; the government puts it at more than $129 million, with victims including a Rhode Island bank and a Massachusetts mutual-fund company. Tell your clients’ payables staff that any new or changed payment instruction gets confirmed on a second channel, using a number they already have, before money moves. Urgency, secrecy, or a lawyer on the call is a reason to slow down, not speed up.

The buyers of small firms now say AI is the plan. Numica raised $21 million to buy practices billing $700,000 to $5 million from owners nearing retirement and move about 75% of their workflows onto its own ledger. Fortune reports that “a CPA still reviews and signs the final work” and that “clients are generally not told AI is doing much of the work,” and Numica’s CEO asked, “Do you care if a calculator or Excel was used to make calculations?” Grant Thornton Advisors closed its CBIZ acquisition on November 2 at a $5 billion enterprise value and tied it to a $1 billion AI investment. If you plan to sell within five to seven years, a buyer like Numica plans to have AI do the work and a CPA sign it, so decide what you’d tell your clients about that before a buyer decides for you.

The first model to clear 50% on Vals AI’s corporate-tax benchmark isn’t one you can use. On November 5, Vals AI’s tax test, in which an AI works through 193 US corporate-tax questions on its own, had its first model above half: Google’s Gemini 4 Carbon at 51.78%, still available only to “a subset of cyber defenders.” The best models you can actually use sit just under half, with Qwen 3.8 Max at 49.66% and Claude Fable 5.1 at 49.19%. That’s corporate-tax research rather than 1040 prep, but the best models on the market getting about half the questions right is the case for keeping your review step.

Data-center limits are spreading. AI prices haven’t moved yet. Voters favored limits in 40 of 41 local data-center ballot measures on November 3, according to Ballotpedia, and Texas’s incoming governor has promised a moratorium on data-center approvals. On November 5 the Nasdaq fell 5.5%, its worst day in 19 months, with chip stocks down 12%, which some analysts read as a bet that the AI build-out will be slower and more expensive. No lab has raised a list price, and the latest moves from Anthropic and OpenAI were cuts. But Anthropic’s IPO filing says computing capacity (the data-center time AI runs on) bought on short-term commitments “currently costs significantly more” than its multi-year deals, and tight capacity tends to show up as lower usage limits and caps before it shows up on a price list, so watch your firm’s plan limits first.

What this week adds up to

This week the connections arrived from both directions. Claude can now reach QuickBooks reconciliation tools, possibly across your client list, on terms Intuit hasn’t published. Your client’s agent can now reach your firm through TaxDome, on terms TaxDome set, and what you control is which clients you help connect and which conversations you keep for yourself. In July I asked whether your firm could answer the door if your client’s agent came knocking. Some firms now can, so when it knocks on yours, what will it find?

If you want a plan for how your firm answers that door, the next Practice Transformation Program cohort is a week-long intensive program from November 30 to December 4, and enrollment closes November 27. Details are at theaiaccountant.ai/transformation.