I spent last week on the floor at Xerocon Denver, and the press releases only tell you half of what happened there. On stage, Xero assembled the whole stack: embedded payroll, payments, expense management, an AI client-chaser, and a month-end close agent. In the hallways, practitioners were asking a different question, and it wasn’t about features. This week’s roundup is a first-hand report on both conversations, plus the stories from outside Denver that complete the picture.
Xero assembled the whole stack, and left the doors open
The announcements came fast. Xero Payroll, powered by Gusto, is now built directly into Xero for US businesses: federal, state, and local filings, direct deposit, contractors, and benefits. Melio Expense Management tracks and categorizes card spend in real time. A new Melio API opens Melio’s payment rails (ACH, real-time payments, checks, wires, virtual cards) to developers in closed beta.
And from the stage came an announcement that hasn’t made it into any coverage I can find: standard ACH bill payments through Melio will be free. That matches Intuit, which made standard ACH free across its Bill Pay tiers this summer, so free money movement is now table stakes on both platforms.
Two agents headlined, and neither is in your hands yet. Casper, Melio’s new AI client manager for accounting firms, will proactively identify missing information, contact your clients, and collect what’s needed to keep the bookkeeping moving; it isn’t yet available. Client-chasing is the least defended and most hated slice of a CAS engagement, and it just got a named product. The other was the new month-end agent in XeroForce, still in closed beta, which Xero says will work document reconciliation status through to manual journals for prepayments and amortization, then hand you a complete account of its actions to review and accept. Notice where the vendor drew the human boundary: the agent produces, you review and accept, which is exactly the accountable review layer we’ve been arguing is your durable work.
Now the witness report, because I was in the room. Everything above was demoed live except Casper, and a conference demo is the easiest artifact in software to produce. The month-end agent, the most interesting thing announced, is in early access with a select group of users in Australia, and it isn’t new: it was announced at Xerocon London a few months ago, and a version of it was previewed even earlier. The distance between the keynote and your practice is measured in re-announcements, and the buzz in the room was matched by exactly that frustration: the excitement is real, and so is the wait.
Here’s the observation I haven’t seen anywhere else. Xero is assembling the stack without closing the perimeter: Melio still connects to QuickBooks, and Casper works across ledgers. Syft and Hubdoc, both Xero acquisitions, still serve competing platforms, and Xero’s own blog goes out of its way to call the new expense product “accounting software-agnostic.” Compare that with Intuit’s walled garden and you’re looking at two different bets on how a platform wins: Intuit bundles to keep you in, Xero absorbs the stack while selling pieces of it to its competitor’s customers.
What this means for you: the close and the chase, two of the stickiest labor lines in a CAS engagement, are being productized by your own ledger vendor. Don’t rebuild your workflow around a keynote; the demo-to-delivery gap is long, and Australia gets it first. But do price forward. When the agent drafts the journals and the chaser collects the documents, what’s left on your invoice is the review, the judgment call, and your name on the result.
Everyone in Denver wanted a cheaper platform. The incumbents don’t sell one.
The other hallway conversation was price. QuickBooks raised prices on August 1, with reported increases taking Plus from $99 to $140 and Advanced from $200 to $340, and Intuit explicitly framed the rise around its new AI capabilities. We covered both ledger repricings in the August 3 roundup, “Nobody is paying you for speed.” What I can add from Denver is what those numbers did to the room: a meaningful share of the practitioners I spoke to were there because they wanted an alternative. Then Xero announced its own US increases days before the conference: from October 1, Early goes to $27, Growing to $59, and Established to $97, with the same justification about product investment and features that are “smarter behind the scenes.”
The honest read is that among the incumbents there’s no cheaper platform to run to, and there isn’t going to be one. As AI gets embedded in every ledger and you actually start using it, the vendors’ own compute bills grow with every transaction their models touch. Those costs get passed through, so shopping between the majors buys you a different logo on the same trend line.
I checked the challengers before letting that sentence stand, and they complicate it in a useful way. The AI-native ledgers now undercut the incumbents on list price: Kick starts at $40 a month, Puzzle’s Core plan is $60, and Digits starts at $65, against QuickBooks Plus at a reported $140. Rillet, this week’s unicorn, went the other direction entirely: no public pricing, sold to mid-market ERP budgets, not yours. But the gap between an incumbent and an AI-native ledger isn’t a price shop, it’s a platform bet: a migration, a younger product, and a different risk profile. And notice the meter on the insurgent side: Puzzle prices its AI agents in monthly credits, which is consumption billing arriving at the challengers before some incumbents have even named it.
BILL made the same point in structural form on its earnings call four days before Xerocon. According to transcript coverage, CEO René Lacerte described a strategic pivot to an agentic platform, meaning software where AI agents do the work by default rather than waiting to be invoked, and management signaled a move away from per-seat pricing toward a platform fee plus usage-based consumption fees. The adoption numbers behind that pivot are the strongest any accounting-adjacent vendor has published: over 175,000 businesses using BILL’s AI agents, its W-9 agent at more than 40,000 organizations, and its invoice-coding agent at over 60,000 companies, where BILL says it has eliminated about 90% of the coding steps on a multi-line invoice. Meanwhile BILL’s customer count actually slipped, to 479,300, while revenue grew 14%. Fewer customers, more automation, more revenue per customer: that’s what a platform monetizing work instead of access looks like.
Hold that against the other price move of the week. On August 21, OpenAI cut its flagship model’s API price by 20% on input and 33% on output. The raw intelligence in your engagement gets cheaper every quarter, while the platforms wrapping that intelligence get more expensive and re-meter from seats to consumption. The margin between those two curves is being captured by somebody, and the vendors have told you their plan.
What this means for you: an incumbent-to-incumbent switch buys no margin relief, and the real price gap, incumbent to AI-native, is a rebuild decision rather than a procurement one. Either way, the margin decision lives in your own pricing. If your fees are built on labor hours while your vendors reprice around work delivered, the squeeze lands on you from both directions, and the answer is the one we keep returning to: price the accountable outcome, not the time.
The builders weren’t a keynote theme. They were the audience.
Three sources converged this week on the same claim: firms have stopped just buying software and started making it. Xero published first-of-its-kind numbers from its own ecosystem: new app registrations are up 4x since 2025, custom apps built outside the App Store now make up one in five of all connections into Xero, and usage of its MCP server grew ten-fold between December and May, passing one million API calls in June. MCP is worth unpacking once: it’s the connector standard that lets an AI assistant like Claude or ChatGPT read and act on your ledger directly, so a firm can build its own workflow on top of Xero without being a software company. One in five connections is no longer an experiment; it’s a measured fifth of the ecosystem, standing against roughly 1,000 certified apps.
I watched this happen live. Across the two days I counted half a dozen sessions on building against Xero, including Regan Ashworth and colleagues building working, Xero-connected apps on stage with Lovable, a tool that turns a plain-language description into working software. Some sessions took ideas shouted from the crowd and built them in the room in under an hour. The energy in those rooms was the real story of the conference for me: the vendor published the data, but the audience supplied the proof.
Jason Staats devoted the same week’s issue to when firm-built software makes sense, and his framing is the useful counterweight. He green-lights four bounded categories (integrators, custom reporting, spreadsheet enhancers, and custom MCPs) and names the honest motivation: “the motivation is rarely efficiency. It’s usually shame… Shame about paying $19/month for an app you could make yourself.” His caution deserves equal billing: “When every aspect of every tool is negotiable, you stop running your firm and start running your tools.”
And Kick shipped the starter kit: Agent Builder Club, a free, open-source collection of accounting agents, including a close agent and an advisory agent, that runs with nothing more than a Claude or ChatGPT account. Community contributions are already arriving, including a Profit First assessment and a chart-of-accounts cleanup agent.
One caution before you hand your team a build mandate. The same week, security researchers reported a malicious-skill campaign with 1.7 million installs across the AI agent supply chain, and OWASP published the first security standard for agent skills. Skills and agents are software, and software has provenance: who built it, who reviewed it, where it came from. A firm AI policy that covers “which model” but not “which skills, from where, reviewed by whom” is governing the smaller risk.
What this means for you: this is the week the ecosystem data caught up with the argument. Our Buy vs. Build series, which concluded this month with “The thing no vendor can sell you,” made the case that the real question was never whether to build but what’s bounded enough to own; Xero’s one-in-five figure says your peers have already decided. Start with Staats’s categories, treat every install like software procurement, and remember the quote Xero put on its own stage, from practitioner Rob Collings: “For the first time in history, there is this period where accountants can start to build things themselves because of AI.”
The client-data question got asked three times in one week. OpenAI answered.
Accounting Today told practitioners to audit their tech stacks for terms that let vendors train AI on firm and client data. Two days later Digits founder Jeff Seibert argued in the same publication that clients’ data belongs to clients. Then OpenAI published a post aimed squarely at organizations handling “financial records, health data, confidential business plans,” reaffirming zero data retention for eligible API customers, meaning prompts and responses aren’t kept after the request is processed, and previewing something further: misuse monitoring that runs on infrastructure the customer controls, with content “encrypted using keys controlled by the customer” so that OpenAI personnel “cannot access the underlying content.” A technical white paper is promised for September.
When the trade press, a vendor founder, and the model provider all land on the same question in one week, the question has become a market. The conversation about where client data goes is now being had for you, whether or not your clients have asked yet.
What this means for you: get to a one-sentence answer per tool in your stack: does anything in this contract permit training on client data, and where does the data go? A firm that can answer that in writing, for every tool it runs, holds a governance artifact its competitors don’t have and a ready answer for the first client who asks. The vendors are engineering themselves out of the “we could see it” position; you should be able to show your clients you checked.
Quick hits
The money priced AI-native accounting this week. Rillet raised a $100 million Series C at a $1 billion valuation, led by ICONIQ, its third round in twelve months, with more than 600 customers and a stated mission of “accounting superintelligence.” The juxtaposition worth savoring: the same vendor whose benchmark showed frontier models are either reliably right or clearly out of their depth just raised $100 million on the thesis that it can manage that boundary. In the same week, top-25 firm Aprio launched Aprio Ventures to invest in emerging AI companies: a firm deciding the returns on AI accrue to owners of the technology, not just users of it.
The two frontier labs showed you trust and money in the same week. OpenAI disclosed it paused its largest planned training run and slowed scaling for two weeks after preliminary evidence that an upcoming model may cross its own critical threshold for cyber capability, with monitoring now consuming roughly 20% of the inference compute being watched. A lab throttling itself at measurable cost is what vendor self-governance looks like when it binds. Meanwhile Bloomberg reports Anthropic’s revenue run rate passed $65 billion ahead of an IPO it expects to be the largest ever, with talk of a $2 trillion valuation; none of it confirmed by Anthropic, and the public debate over whether that run rate is real is, at bottom, an argument about revenue quality and recognition. That’s your home turf, and your clients will ask you about it.
Ramp built a router to cut AI bills. The spend-management company launched Router, a single endpoint that sends each AI request to the cheapest model that meets the required performance level, with coverage reporting internal savings around 40%. A finance platform, not a tech vendor, just built infrastructure for managing AI spend, the same month the Big 4 started budgeting AI usage like a cost center. “What does your AI actually cost per unit of work” is becoming an answerable, and billable, advisory question.
The platforms are assembling. The firms are building.
Here’s the pattern from one seat in Denver. From above, the platforms are pulling payroll, payments, expenses, the chase, and the close into the subscription, and repricing that subscription as they go. From below, the firms in the audience are building their own apps, agents, and connectors at a pace the vendor’s own data now measures at a fifth of its ecosystem. Both movements were on the same stage, in the same week, and they’re not in conflict: they’re competing for the same territory, the work between your ledger and your judgment.
The platforms will keep absorbing, and their prices will keep rising with their compute bills. The models underneath will keep getting cheaper. The firms that capture the gap between those two curves are the ones building their own capability on the falling curve instead of renting everything on the rising one. Which curve is your practice on?
If you don’t like your answer, that’s a conversation worth having before your vendors have it for you. Book a free consultation at theaiaccountant.ai/consultation and we’ll work through where your practice sits between the two curves, and what to change first.