On August 24, Thomson Reuters, the company that sells Checkpoint to this profession, announced it had trained its own AI model. Most of the AI tools in your stack are an app on top of a model somebody else trained; this one, called Thomson, has Thomson Reuters' own content built into the model itself, and the company plans to roll it out across its tax and legal products. The program cost $40 million over several years. The last step, three weeks of rented computing time that turned a starting model plus Thomson Reuters' archive into Thomson, cost under $450,000, which tells you the expensive part was the archive and the people, not the training.
It was cheap because Thomson Reuters didn't start from nothing. It started from Qwen, an open-weights model built by Alibaba, meaning a finished model published for anyone to download and build on, and ran it through 200 billion tokens (roughly 150 billion words) of additional text drawn from public sources and its own Westlaw, Practical Law, Checkpoint, and Reuters archive. This series has spent four parts telling you the intelligence layer, the model doing the thinking underneath every AI app you use, is a commodity you rent from the frontier labs (OpenAI, Anthropic, and Google, the companies making the most capable general models). That's still true. What Thomson Reuters did is what a commodity looks like when somebody with a proprietary archive decides to build on it.
Two things before you read further. First, Thomson Reuters still rents most of what it runs on: its own model powers one feature inside one product, and its CTO says other vendors' models may be used where they're better. Second, on Thomson Reuters' own results table, Claude Opus 4.8 already scored higher than Thomson the day the report came out, and the report describes a "model factory" built to be run again, so the company has signed up for a retraining schedule. No firm your size should do that. The useful part of the story is what Thomson Reuters chose to build on, because you own the same thing.
What went into the model is what you're sitting on
Strip the engineering away and Thomson Reuters built on two things: its published archive, and, in its CTO's words, "work created by its own subject-matter experts." Note what it left out. It bought no data, and it used no customer data.
Your firm has the same two things, and the same line between them. The archive equivalent is your methodology: the review note that says "not that account, this one, because the client leases the vans," the standing rules you apply to a monthly close or an annual return, every correction you've made to an AI draft this year. The customer data equivalent is the client files themselves, and they stay where they are. What goes in front of the AI is the rule, written as a rule, not the return with the client's name on it.
The market already prices the written-down version. Apron, a UK bookkeeping-automation vendor, shipped plain-English standing rules this month ("Invoices from Vodafone are always Telephone and Internet, with 20% VAT"; "Never publish documents over £1,000 automatically") to every user of a product priced from £5 a client. Thomson Reuters put its knowledge inside the model. Apron put it on top, as instructions. Same knowledge, two floors, and only one of them needs a $40 million program.
Why now: the packaged version isn't coming for you
I said in the September 7 roundup that I'd explain at length why I think the most disturbing trend in the profession right now is where the vendors built around AI from the start are choosing to sell. Here it is. The finished products, the ones that run the close or the tax workflow for you, keep arriving for the largest firms first.
Basis runs the close inside top-25 firms and is waitlist-only for everyone else. Intuit's agent builder is piloting with enterprise firms. Accrual agreed to buy Puzzle's ledger and says it remains focused on the Top 100.
Here's why that's worse than it sounds. The pitch of every AI-native vendor since the first of them launched has been that a 12-person firm could finally run like a 200-person one: the same close, the same tax workflow, without the headcount. What's shipping is the opposite, because a vendor sells the finished product to the buyer who can fund the implementation, and that's a Top 100 firm with a technology team and hundreds of clients on one platform. A 12-person practice is a support ticket by comparison, so it gets the waitlist. The capacity gap these tools were supposed to close is being widened by the way they're sold.
And it's closing from two sides at once. In the same deal that put Puzzle's ledger behind a Top 100 platform, what's left of Puzzle became a fixed-price accounting-and-tax service sold straight to small businesses, with "a small group of ambitious accounting partners" somewhere in the delivery. That's the shape to watch. The packaged AI layer moves up-market to the firms that can pay for it, and a packaged AI service moves down-market to your clients, priced by the month. A firm in the middle can't buy its way to a different position, because the only AI it can buy off the shelf is the same bundle its competitors and its clients get.
The base model underneath all of them is available to you at the same public price the largest firms start from. The packaged layer above it is not, and the packaged version for a 12-person practice is not being built by anyone but you. That's the reason to act on what follows now rather than at the next renewal.
What this means for you: two moves
First, write the methodology down, this month, for one engagement. Pick the engagement you repeat most, whether that's a monthly close, a quarterly review, or an annual return. Write the standing rules you apply to it that currently live in your head or your review notes: what each client's recurring items mean, which positions you always check before you sign, what counts as an error, who signs. Keep it to a page of plain text, save it in a format you own rather than inside a vendor's memory feature, and put it in front of whichever AI you already use, as a project instruction or a saved procedure (part 3's portability rule: plain files move with you).
That page does the job Thomson Reuters paid a training run to do, at the scale you can afford. It isn't free of maintenance: when the model underneath changes, you re-test the page and edit a line. Thomson Reuters re-runs the factory. Your maintenance is a text edit, and the difference in cost is the whole argument.
Second, get ready for your vendors to do what Thomson Reuters did. Harvey, the legal AI vendor, raised $550 million on September 9 with "professional services firms" written into its market and its own model trained on an open-weights base. Deloitte opened a practice this month to build on open-weight models. Your ledger, capture, and tax vendors will follow, and at your next renewal the pitch will be a model of their own, trained on their data, as a reason for the price.
Ask two questions before you pay for it. Show me what your model does on my work that a frontier model with my own rules in front of it doesn't. And when the model underneath changes, how will I know? A vendor that can't show you a version trail hasn't answered the second one, and that's part 1's maintenance list landing on their side of the contract.
The question moved, and the answer got cheaper
Part 4 asked what's bounded enough to own. Thomson Reuters answered it for a company with $40 million to spend: own the layer trained on your archive, and the retraining bill that comes with it. For a firm your size the answer is one floor up. Own the knowledge, written down, on top of whichever frontier model wins the month. Owning a layer means owning its maintenance, and the layer you can afford to own is the one whose maintenance is a text edit.
That closes the series. Five parts ago the question was whether your apps were wrappers, an app's own screen on a rented model. Now the question is whether your firm's knowledge is written down anywhere an AI can read it.
The Encoding Loop Starter Kit is the on-ramp: the correction-capture template turns each review note into a rule in the format above, the vertical-narrowing worksheet picks the engagement, and the 90-day roadmap takes you from one page to a loop. Download it free at theaiaccountant.ai/encoding-loop-starter-kit and start the first page this month. Thomson Reuters has used less than 10% of its archive so far. How much of yours is written down?