On August 19, BILL told the analysts on its year-end earnings call that it’s pivoting to an agentic platform (agents being AI that carries out multi-step work on its own, rather than answering questions), and that per-seat pricing is on the way out, replaced by a platform fee plus fees based on consumption: the metered work its AI actually does. Eighteen days earlier, Intuit had folded equivalent bill-pay capability into QuickBooks Advanced at no extra cost. Your AP layer is repricing from seats to work done in the same month your ledger vendor started giving the same capability away.
Two weeks before that call, part 2 of this series looked at the same possibility and deliberately hedged: usage-based pricing was something to watch for, not something that was coming. The hedge lasted fourteen days. We closed this series in mid-August expecting the market to take its time supplying evidence. Instead it supplied a month of it, on the record, layer by layer. So this part re-runs the five layers from part 1 against what August actually did.
The intelligence layer keeps repricing itself
Start with the layer the wrapper critique got right. On August 21, OpenAI cut its flagship model’s API price by 20 percent on input and 33 percent on output. That’s not a one-off: some frontier model gets cheaper almost every week now, and the cheap challengers underneath keep getting better. Part 3 said a nearly-as-good model arrives every few months. Revise that to weeks.
What’s new isn’t just the price chart. It’s that routing (sending each job to the cheapest model that can pass your checks) is turning into ordinary practice, with its own how-to guides. Nate Jones, writing in late August, put the operating rule in one line: “the most expensive model no longer gets every job by default. It has to earn the work.” When the intelligence in your stack is routable, any subscription priced as if the intelligence were the value has a problem.
The interface prediction shipped as a product
Part 3 argued the durable interface is a review queue: the transaction screens empty out as agents take the work, and what survives is the surface where a human reviews the output and stands behind it. In August, Xero built exactly that boundary into a product. Its new month-end agent, in early access now, works reconciliation through to draft journals and then hands the practitioner, in Xero’s words, “a complete account of its actions to review and accept.”
The vendor drew the human line precisely where this series said the durable value sits. When you’re paying for an interface, that’s the half worth paying for.
The harness and the janitor both got cheaper to supply
Part 1 called the harness (client scoping, confidentiality, keeping one client’s data out of another’s work) the layer that most honestly earns a vendor’s margin. In August, the model vendors started absorbing pieces of it. OpenAI reaffirmed zero data retention for eligible API customers and previewed misuse monitoring that runs on keys only the customer holds: “OpenAI personnel do not have a copy of those keys, so they cannot access the underlying content.” The harness is still a real value layer, but part of it is migrating into the model layer itself, priced at zero.
The janitor’s tools grew up the same month. In one week, Anthropic moved skills (encoded, reusable procedures an AI can follow), computer use, and file handling out of beta, with version pinning. Pinning matters more than it sounds: it’s the difference between a demo and a procedure you can hand an agent every month, then regression-test when the model underneath changes. The janitor list from part 1 didn’t get shorter. But the mop got much better, and that lowers the build-readiness bar part 3 set.
Two verdicts came in
First, the capital verdict. Rillet, an AI-native ledger two years out of stealth, raised $100 million at a $1 billion valuation on August 17, its third round in about a year. And Aprio, a top-25 firm, launched a venture arm to invest in AI companies, per trade press reports. The money has decided that owning this technology beats renting it.
Second, the practitioner verdict, and it’s the one that changes this series’ closing question. Xero published first-of-their-kind ecosystem numbers in late August: new app registrations up four times since 2025, usage of its MCP server (a standard connector that lets an AI plug straight into the ledger) up ten-fold in six months, and “one in five connections into the Xero ecosystem is now a custom app from outside of the Xero App Store.” That’s the vendor’s own measurement, and the definition is loose enough to catch simple integrations. The direction isn’t loose at all: firms are already building.
The same week, Jason Staats devoted an issue to when firm-built software makes sense, and Kick shipped a free, open-source collection of accounting agents that runs on nothing more than a Claude or ChatGPT account. Even the platforms want in: Intuit’s new Custom AI Automations lets firms on its Accelerate tier build their own AI automations from a prompt or a firm SOP and run them across clients, on a metered monthly capacity. Read that carefully: the platform isn’t fighting the builders, it’s selling them the build surface. Staats’s caution is the one to keep taped to the monitor: “When every aspect of every tool is negotiable, you stop running your firm and start running your tools.”
The question moved
Put the month together. The pricing wrapper is being rewritten toward consumption, exactly as hedged. Intelligence reprices weekly and is now routable. The vendor interface is converging on the review queue. Pieces of the harness are free, the janitor’s tools are production-grade, and both verdicts came in for building.
Every one of those moves the answers to part 3’s renewal checklist, and all of them move in the same direction: toward more layers your firm can honestly supply. So the question this series closed on has shifted under it. It’s no longer whether firms like yours will build; the ecosystem data says they already are. It’s what is bounded enough to own: which problems are small enough, checkable enough, and stable enough that your firm should hold the mop, and which still deserve the packaging premium. Staats’s warning and part 3’s build-readiness test are how you draw that line honestly.
One more thing happened in August. Thomson Reuters, the company that sells research to this profession, built its own AI model and published what it cost, and that story doesn’t live in any of the five layers: it moves the buy-vs-build question down a level this series hasn’t touched yet. That’s part 5, next week. Before it lands, re-run part 3’s one-question-per-layer checklist against this month’s prices. How many of your five answers still stand?
Drawing that line is an operating decision, not a software one, and it needs somebody inside your firm who owns it. That is what the Practice Transformation Program builds. Enrol your champion at theaiaccountant.ai/transformation: four live sessions, nine modules, and a transformation plan your firm executes rather than files, for $2,997.