On July 21, CliftonLarsonAllen put something unusual for a firm of its size in a press release: it is going to train its own AI model, rather than license one or bolt one onto its existing stack. A model trained on what the firm itself knows.

The announcement pairs CLA, a firm with more than $2 billion in annual revenue, about 9,000 people, and 120-plus locations across the US, UK, and Europe, with Digits, the AI-native ledger company whose platform and model-training technology will do the encoding. The plan is to roll the model out first inside CLA’s client accounting and advisory services practice, which the firm says represents roughly 15 percent of its revenue, then deploy it to thousands of clients over three years. Accounting Today reports the offering will be optional for clients within CAAS service levels.

Digits founder and CEO Jeff Seibert put the mechanics in one sentence of his statement: “By encoding decades of judgment and knowledge into a firm model that benefits clients, Digits empowers CLA’s professionals to spend their time where it matters most.” The release is just as direct about how the model keeps learning: every transaction CLA clients and professionals review, categorize, or correct in Digits sharpens the intelligence behind it.

Encoding decades of judgment is the story here, and it reaches well beyond one firm’s product launch. The rest of this edition is about what that phrase actually commits a firm to.

For two years, the AI story in accounting has been firms buying capability from outside. We have assessed that markettool by tool: platforms that run the close, agents that reconcile, ledgers rebuilt around AI. In nearly every case, the intelligence belonged to the vendor, and the firm rented it.

The interesting exceptions pointed in one direction. Ramp built Stack so that firms teach it their procedures, which the company frames as the firm’s own operating knowledge captured as standard procedures that stay with the firm. Basis raised $100 million in February at a $1.15 billion valuation and says its agents already run inside roughly 30 percent of the top 25 US firms. Digits priced its firm plans on how much work its agents handle without human touch. The vendors understood before most firms did that the scarce ingredient is not compute. It is the accumulated, corrected, industry-specific judgment sitting in a firm’s files and its partners’ heads.

CLA’s move takes the logical next step: if the judgment is the asset, own the model it gets encoded into. CEO Jen Leary framed it as access: “AI should not belong only to the biggest companies and the biggest firms with the largest budgets.” Take the framing seriously, because there is a real version of it. A $2 billion firm building a model that its small-business clients use through a CAAS subscription is, in fact, AI reaching businesses that would never buy an enterprise platform themselves.

But notice what the framing quietly concedes. If firm-owned models are how AI stops belonging only to the biggest, then the firms without a model of their own are on the wrong side of the sentence.

We read the press release, the Accounting Today coverage, and the trade write-ups. Three questions that matter to practitioners, and to CLA’s clients, are not addressed in any of them.

First, consent. The model improves, per the release itself, on what clients and professionals review, categorize, and correct. Client transactions are the raw material. The release does not describe how client consent works, what an engagement letter for a model-training relationship looks like, or whether a client can use the service while opting out of sharpening the model. To be fair, the rollout is described as optional, and nothing suggests CLA is skipping its professional obligations; the AICPA’s confidentiality rule and its third-party-provider interpretation exist precisely for arrangements like this. The Code gives a firm two clean paths: a binding confidentiality contract with the provider, carrying reasonable assurance about its safeguards, or the client’s consent up front. A model-training arrangement stretches both paths, because the disclosure is not a one-time handoff. It is continuous, and it improves a commercial asset. But an announcement this public, about training on client work, that never says the words consent or confidentiality, is a gap other firms should not copy. When the profession’s flagship experiments set the template, the template should show its controls.

Second, ownership. The model is described as CLA’s own, proprietary, firm-specific. It is also built, hosted, and trained on Digits’ platform, with Digits’ technology, under terms nobody has published. The question every firm considering this path should ask is the one we raised when we covered the AI-native ledger shift: what happens at the exit? If the relationship ends, does the model, and everything those decades of judgment taught it, leave with the firm, stay with the platform, or dissolve into ambiguity? “Proprietary” is a legal claim, not a technical description, and none of the public materials say who holds what. Financial terms were not disclosed.

Third, the moat. The release describes a flywheel: more clients, more corrections, a sharper model. Flywheels compound. A firm pushing thousands of clients’ transactions through the platform, with 9,000 professionals correcting the model’s mistakes, learns at a rate a 30-person firm cannot approach. Leary’s line about AI not belonging only to the biggest is true for CLA’s clients. For CLA’s competitors, the mechanics run the other way: this is a scale advantage that gets harder to catch every quarter it runs. The honest read is that both things are true at once, and which one applies to you depends entirely on which side of the model you are standing on.

What follows is our read, not reporting. Firm-owned models are about to become a category. The pieces are all in place: platforms like Digits and Ramp have industrialized the training infrastructure, the large firms have the data volume to make it work, and this announcement gives every managing partner a board-meeting slide to point at. Expect more of these deals within the year, and expect the marketing to run ahead of the disclosures, because it already has.

The firms that navigate this well will treat it as two separate decisions. One is technological: whether to encode your firm’s knowledge into a model, and on whose infrastructure. The other is fiduciary: what you owe the clients whose engagements are the training data. The first decision is exciting. The second is the one your professional obligations, and eventually your regulators and insurers, will grade. Underwriters have a way of asking exactly the questions a press release skips.

If you run a mid-size firm, the uncomfortable news is that waiting is now also a strategy choice. You will probably never train a model alone, but your accumulated judgment, your niches, your corrected workpapers, are exactly what the platforms want, and what the consolidators are buying. Before you sign any platform agreement, read the training and data clauses: who learns from your firm’s corrections, who keeps that learning if you leave, and whether your clients’ data feeds anything beyond your own engagements. Those clauses, not the demo, are the deal. And they cut both ways: a platform that cannot answer them clearly has just told you everything you need to know.

If you work in CAAS or outsourced accounting, your daily corrections are now, at some firms, literally the product. That makes your judgment more valuable, not less, and it is fair to expect your firm to be able to explain where those corrections go. Keep your own record of what you catch and fix, too. In a world where corrections train the product, the professional who can show their judgment has receipts.

If you are a client of a firm heading down this road, ask the plain questions: does my data train your model, can I opt out and still get the service, and what do I get back for helping make your model smarter? A good firm will have answers on paper. The release that started all this does not, yet.

Seibert’s phrase is the one to keep: encoding decades of judgment. A model can absorb every categorization a firm ever corrected, and that is real, compounding value. What it cannot absorb is the thing the judgment was for: standing behind the next decision, the one that has not happened yet, with a name and a license attached. AI can do the work. It cannot sign the work. CLA seems to understand that, which is why the model ships inside an advisory practice full of humans rather than instead of one. The firms that copy the model without copying that understanding will discover the difference in front of a client, and the ones that never ask who owns their encoded judgment will discover it in front of a lawyer. Your firm’s judgment is becoming an asset class. Decide who gets to capitalize it.

— Footnote

Footnote is an independent publication. It is not professional accounting, tax, or legal advice. Our analysis and opinions are based on the company announcements and reporting linked above; company statements are their own, and we have not reviewed CLA’s client agreements or the commercial terms between CLA and Digits, which are not public. Details are current as of July 28, 2026. We have no consulting relationships with any company named in this article.