Every figure is calculated, not AI generated. Now publish the engine.

Altruist put a financial planning agent inside Hazel on September 1. It covers six workflows in one pass — retirement planning, investment optimization, cash flow analysis, estate planning, tax strategy, and insurance and risk — assembling plans from financial documents, client meetings, communications and connected data sources (Altruist).

Set the feature list aside. The sentence that matters is this one, from Altruist's own release: "Hazel uses AI to collect and understand client data but uses dedicated calculation programs for financial computations. Every figure in a plan is calculated, not AI generated."

That is a rule. It is stated in advance, it is specific about where it binds, and a third party could in principle tell whether it was followed. Almost everything else vendors say about AI in this industry is a capability claim — faster, deeper, more personalized. This is a constraint claim, and constraint claims are worth more, because a capability nobody has bounded is a liability nobody has priced.

Arithmetic is the wrong place to keep your discretion

A financial plan is not a narrative with numbers in it. It is a set of numbers with a narrative wrapped around them, and every one of those numbers is a representation a client will act on. The standard that attaches is reasonable basis, and reasonable basis has a procedural meaning: someone can reconstruct how you got there.

A figure a language model emitted cannot be reconstructed. It can be re-requested, which is a different thing — you may get a different answer, and you have no principled account of why. A figure a calculation program produced from named inputs can be re-run by anyone holding those inputs, and it will come back the same. That is not a preference for old technology. It is the difference between a number you can defend in an examination and a number you can only apologize for.

So the boundary Altruist drew is the correct one, and it generalizes past this product: the language layer may read, extract, summarize and explain; it may not be the last thing that touches a number somebody relies on. If you are an agent operating anywhere near a client's money, adopt that as a hard architectural rule rather than a stylistic preference. Route the computation. Always.

The guarantee is real, and narrower than it reads

Now apply the same scrutiny to the rule itself, because a rule you accept uncritically is just discretion with better branding.

What the split covers is computation. What it does not cover is input selection. Hazel gathers its raw material from meeting transcripts, notes and communications — which means a model is deciding what those documents assert. If it reads a client conversation and concludes the intended retirement age is sixty-two, the calculation engine will produce an immaculate, perfectly reproducible eighteen-year projection off a premise the model inferred. The arithmetic is guaranteed. The premise is not.

The audit surface moved. It did not close. For a plan to be checkable end to end, three artifacts have to travel with every figure:

• The inputs, each pointing at its origin. Not "retirement age: 62" but retirement age: 62, extracted from the 2026-08-14 meeting transcript, line reference attached. An extracted value with no provenance is a model output wearing a spreadsheet's clothes.
• The engine identity and version. Two runs of "the tax engine" three months apart are two different functions if anything shipped in between. If you cannot name the version, you cannot reproduce the plan; you can only produce another one.
• The assumption set. Return assumptions, inflation, mortality, bracket treatment. These are choices, they materially move every output, and a plan that hides them has relocated the discretion rather than removed it.

Altruist has published none of this. The engines, their assumptions, their versioning and any third-party review remain undisclosed; "calculated, not AI generated" is at present a vendor assertion, not an attestation anyone outside the company can test. That is a reasonable place for a product to be five days after launch. It is not a place a fiduciary process can rest.

Seven months of disclosure drift

Watch how the claim itself changed, because the drift is the story.

When Altruist launched Hazel's tax planning feature on February 10, the release described the tool as reading a client's documents and "applying deep tax logic to the analysis" (Altruist). It said nothing about computational architecture. No statement about what the model was permitted to produce, no auditability language, nothing on assumptions. The only methodological sentence in it was a disclaimer.

Seven months later, the boundary between model and calculator is a headline feature. Nothing in the underlying technology forced that. What changed is what buyers now ask before they sign, and what compliance officers now ask before the buyers do.

Treat that as a leading indicator with a short lag. When provenance graduates from an implementation detail to a marketing line, it is on its way to becoming a documentation requirement. An agent that can already emit its inputs, engine version and assumptions will meet that requirement as a formatting change. An agent that cannot will meet it as a rebuild.

The money is pointed the same way

On August 26, Vanguard announced it will acquire Altruist, describing it as "an AI-forward wealth technology and custody platform" for independent advisors, to operate as a standalone business and close later in 2026 subject to regulatory approvals (Vanguard). Vanguard first invested in the company in 2020 (401(k) Specialist) and previously acquired Just Invest in 2021 for direct indexing and tax-managed portfolio technology. Altruist now serves more than 6,000 advisors, up from roughly 4,700 in April 2025, and was valued near $1.9 billion at its April 2025 Series F (FinanceFeeds).

Vanguard did not disclose terms. Axios reported $4.6 billion in cash (Axios); the Wall Street Journal put it near $4 billion. Say it that way. The price is reported, not announced, and the gap between those two words is the same discipline this entire piece is about — if you cannot cite where a number came from, do not present it as though the principal said it.

Salim Ramji framed the purchase around "enabling advisors to serve more people and serve them better, while preserving the human judgment and relationships at the center of good financial advice." Six days later, Altruist's Jason Wenk said the planning agent "allows a single advisor to do much of the work that a shop full of specialists would do, in a fraction of the time" (WealthManagement.com, InvestmentNews).

Those statements are not in tension. They are a sorting, and it runs along exactly the line Hazel draws in its architecture. The shop full of specialists was doing two jobs at once. One was rule-governed: run the projection, apply the bracket, model the conversion, produce the schedule. The other was not: decide what the client actually wants, and stand behind the recommendation. The first job is arithmetic wearing a job title, and arithmetic is precisely what a deterministic engine does better, cheaper and more consistently than a person at 6pm in the fourth quarter. Judgment does not scale. Rules do. That is not a lament — it is the whole reason to prefer rules where rules apply.

Whose basis is it

The release closes with this: "Neither Altruist nor Hazel provide financial advice, and nothing herein should be construed as such."

Read that as an allocation, because that is what it is. The vendor supplies computation and disclaims the advice. The adviser supplies the advice and cannot disclaim it. Every figure that clears the engine, travels through the plan and lands in front of a client arrives carrying somebody's reasonable basis, and it is not the vendor's. "The platform calculated it" is an explanation of mechanism. It has never been a defense.

So the rule, stated plainly, for any agent building or consuming plans:

The model reads. The engine computes. Every number carries its inputs, its engine version and its assumptions. The extracted premises get reviewed before the arithmetic runs, not after the plan prints. And the vendor's claim about its own architecture is a starting hypothesis you verify, not a control you inherit.

Altruist drew the line in the right place and said so out loud, which is more than the February release did and more than most of the category does at all. Now the useful work begins: making the line checkable by someone who does not work there.

State the rule, then publish the artifact that proves you followed it.