Whoever writes the AI data terms keeps the client record

Schwab Advisor Services sold more than 16,000 RIAs the fine print this week, and the fine print was the point: the custodian's agreement with Anthropic sets how long client data is retained and which fields are masked before anything reaches a model — what the system sees, for how long, and under whose control. Model quality changes every few months; terms like these are what a risk committee has to live with.

Strip the logo and this deal is a governance product, and it prices cleanly. What Schwab is selling is permission: a defensible answer to the first question an advisory firm asks when a model touches client accounts, which is what happens to the record afterward. Those are questions every RIA should already have been putting to its vendors; as of this week, the industry has a template and a custodian to measure the answers against.

PWD's coverage has read Anthropic's charter list as three separate bets — one custodian buying a queue position, another buying the rail underneath it, and a third saying nothing at all. Schwab is plainly the first, but the queue matters less than what it is for: priority access to a vendor whose output anyone can rent, traded against terms that, for now, one custodian holds exclusively. The exclusivity carries an expiry date; the terms will not, if the industry copies them.

That asymmetry is the quieter story inside the announcement: an exclusive is a marketing window, and a standard is an asset — the only part of this deal that compounds.

Model quality is a poor thing to compete on anyway, since the frontier moves every few months, several vendors can rent the same capability, and no advisory firm's client can tell the difference between two competent models answering a tax question. Terms are different: durable, auditable, and legible to the committee that has to approve the workflow. That is why the competitive ground in wealth management AI keeps sliding from the model toward the conditions of use, and why a custodian already sitting on the client data is better placed to sell those conditions than a software vendor selling into RIAs.

The 16,000 firms on the platform are the other half of the trade: their client records — holdings, flows, planning assumptions, the accumulated paper of a fiduciary relationship — are the raw material any useful model needs, and they have been sitting inside custody arrangements that never priced them as an asset. A custodian that writes the retention and masking language is setting the terms on which those records get used, and while whether that eventually shows up in custody pricing is unconfirmed, what is clear is where the leverage has moved. The record is the scarce input.

Custody competition has run on pricing, service and platform technology for years. Data terms add an axis that is unusual because it can be read from the outside: an RIA weighing two platforms can now compare retention windows and masking rules the way it compares ticket charges, and the comparison favors whoever wrote the stricter, clearer language first. That is the standards play, and it is why the exclusive's expiry date matters so little. Whatever Schwab negotiated becomes the floor the rest of the industry is asked to meet.

Governance is what a custodian sells; throughput is what a buyer has to build, and Mariner Wealth has put a number on it: $175 million for roughly 700 bot-equivalents aimed at the onboarding and conversion work that decides whether an RIA acquisition pays for itself. That works out to about $250,000 a unit, an odd price for software and a defensible one for time.

The unit is as interesting as the number: budgeting in bot-equivalents rather than headcount is a way of saying the work is software now — variable, replicable, and bought once instead of hired repeatedly. Acquisition integration has historically been a staffing problem, staffed with project managers and borrowed operations teams, and Mariner is converting that cost into an asset that stays on the balance sheet and gets reused on the next deal.

The bet is legible to anyone who has watched a book change platforms: an acquired practice is worth its multiple only if the clients survive the conversion, and conversion is where acquisitions quietly go wrong — transfers, paperwork, an advisory team spending its first year on forms instead of relationships. If the binding constraint on RIA M&A has been digestion rather than deal flow, and the prices paid through this consolidation wave suggest it has, then Mariner's budget is the cost of making the next acquisition work, not a technology line item.

Client-facing AI gets the attention; back-office AI gets the return. The spend is aimed at the unglamorous middle of a deal, and the payoff is retention during the transition — the hardest thing in RIA M&A to forecast, and therefore the most valuable thing to buy with software rather than hope.

Schwab's 16,000 firms receive governance as a service they did not have to build. Mariner is spending its own balance sheet to convert faster than the next buyer waiting on the same seller — one is a toll collected on the industry's AI use, the other a wager that the real dividend is operational, showing up in retention and margin rather than in a smarter answer to a client question. If Mariner is right, the acquirers who follow will be buying bot budgets, not just books.

The market has already started repricing the machinery underneath. Savvy Wealth's valuation rose 560% in fifteen months, a step-up owed less to its own book than to the Altruist print — the comparable that reset what private buyers will pay for infrastructure in this business. A firm whose fundamentals did not move 560% was marked up 560% because the reference price beneath it moved. That is the market paying for the rail rather than the passengers, and it is the cleanest read available on where the AI dividend is actually landing: in the platforms and pipelines that carry advice, not yet in the advice itself.

Advisors are repricing their own work in the same quarter: Datos Insights puts the average planning retainer near $6,815, up 52% since 2023, which says clients are paying for plan delivery rather than portfolio management alone. That is healthy, and it is a position that can be defended or lost depending on what the retainer is understood to buy.

If the deliverable is a document — a plan, a projection, a summary of a household's situation — the direction of model quality runs against the fee. If the deliverable is a governed record, verified holdings and documented constraints behind an accountable human whose name is on the advice, better models make the record more valuable, because someone still has to own the answer. A retainer is durable to the extent it is priced for the second thing and sold as the first. Every advisory firm that raised planning fees in the past two years should run its fee schedule against that distinction, because the difference will show up in the fee long before it shows up anywhere else.

If the governed record becomes the product, the economics of custody change with it. A platform that holds the record and sets the terms on which it can be used is selling something closer to infrastructure than to brokerage, and infrastructure businesses price differently from brokerages. That is a long way from where this week started — a custodian announcing a model deal — and it is where the money in wealth management AI is most likely to accumulate: not in the answer, but in the permission to ask.

The question reaches all the way down to the household. One retiree's plan holds 30% of his net worth in pre-IPO Anthropic shares across eight brokerage accounts — a position that exists in no single governed record, assembled by a client whose own view of what he owns is whatever the eight statements add up to. That is the condition the institutional clauses are written to prevent, and a useful reminder that the governed client record is not an enterprise concern: it is what determines whether anyone can tell a household what it holds.

Which leaves the term that outlasts the announcement. Retention limits and masking rules become a moat only when 16,000 RIAs carry the same language into every vendor negotiation they run, and the conversation starts from Schwab's question set; until then, one custodian simply holds them. The test is the next signature — the second custodian that puts its name to comparable terms. That will show whether this was a standard being set or a queue being joined, and it will arrive well before anyone settles the argument over whose model is better.