Mariner's $175 million bot budget is a supervision purchase

Mariner's $175 million commitment to bots, at $250,000 per bot across five years, amounts to $50,000 a year — junior-hire territory rather than a software subscription — and divides into roughly 700 seats. No firm staffs 700 licenses; it staffs 700 supervised workflows, drafting power acquired inside a business that still has to review everything the machine produces. The model is the cheap part of the arrangement and always was. The review is what scales with use — the first thing a regulator will ask about and the line item almost nobody in the adoption data has priced.

RIAs doubled their AI use this year, retiring the old question of whether firms are experimenting and replacing it with a harder one: the experiments have not produced money. Boston Consulting Group found that only 6% of companies have seen AI reduce costs or grow revenue, a payoff rate low enough that the adoption curve and the return curve now describe two different businesses running at once. The distance between them is where supervision spend lives, and it is why a $175 million bot budget reads as a wager on control rather than on capability.

That gap points to something else: the cost of running AI well sits where a license fee never covered — in the review step, the data handling, and the record that survives an audit. A practice can turn on a drafting tool in a week, but building the supervision around it takes longer, costs more, and produces nothing a client can see — which is exactly why it gets deferred and exactly why examiners reach for it first.

How practices intend to staff that control is visible in the hiring data: 73% of surveyed RIA firms said they want to add junior advisors, and 15% said they want compliance staff specifically around AI adoption. The first figure is capacity; the second is the check on it, and set side by side they describe a plan built on production-line arithmetic — three-quarters of firms hiring at the drafting end, roughly one in seven hiring at the inspection end. That spread assumes supervision is a light lift, and it reads more like a plan that has not yet had its first examiner conversation.

The regulatory line, when it lands, comes through a door that is already open. The SEC's AI priorities turn on who signs the review, and the existing fiduciary and disclosure duties already reach AI-drafted advice the moment it sits in front of a client, so nothing has to be newly written for the exposure to attach; the old obligations simply find a faster-moving object. Regulation S-P and the disclosure regime place the cost in the approval step rather than in the model, which means a firm that cannot show which human reviewed which recommendation has a supervision problem a better prompt will not fix.

It is a timing problem as much as a legal one. Firms that spent the last two years treating AI as a productivity experiment now face the same duties they always had, applied to output produced at machine speed, and the approval step becomes the choke point where a fast tool meets a slow control, a thousand drafts a week arriving at a desk built to sign off on twenty. The record of who approved what is the thing that either holds or does not.

Two product announcements this week read as direct answers to that problem, and neither of them is about model quality. Hamachi launched a patent-pending masking layer that keeps client data out of large language models, a design that assumes the model will see something and decides, up front, what; Conquest dropped "Planning" from its name and placed its bet on a white-label advice engine whose compliance case rests on the SAM audit trail. One decides what the model is allowed to know; the other decides what the firm can prove it reviewed — both data-custody arguments wearing technology labels.

That is a differentiator a firm can underwrite before it signs, which is more than most of the AI sales deck offers; an audit trail is the export clause of the AI era — the thing a practice discovers it needs when it tries to change platforms, or when an examiner asks for the file. Masking is the same wager from the other direction, settling what the model is permitted to see before anyone asks who is answerable for the answer, and both cost money on the way in and retire a category of problem on the way out.

Scale is what turns those from features into prerequisites: a 700-seat deployment generates review obligations at a volume no manual process survives, which is why masking and an audit trail stop being nice-to-haves and start being the floor. A practice running one assistant lightly can supervise it with a partner and a checklist; the trouble sits in the middle of the range, where firms adopt at a pace their compliance function was never sized to match. That is the profile the hiring survey describes, and it is the profile most exposed.

At the small end of the market the arithmetic sharpens, because XYPN bought AI-search access for its members, handing them the machinery of testimonial-driven visibility, though the credibility behind an AI answer accrues over years and the payoff sits past the length of most marketing contracts. Salesforce's Agentic Advisor suite turns meeting output into prioritized action rather than a transcript — a genuine move from note-taking toward follow-through — with adoption at small firms the part still unproven. These are the firms with the thinnest supervision benches, and the tools reaching them first are the ones that generate the most reviewable output.

Recruiting intent points the same way: the 73% chasing junior advisors are, in many cases, also building the succession pipeline the profession needs, and Cerulli counts 35% of advisors retiring within a decade, which turns the junior hire into a double bet on capacity and continuity. What the survey does not show is the compliance hire keeping pace; 15% is a thin control layer to lay under a production line that three-quarters of firms are trying to staff, and practices that spend on the drafting end first tend to meet the review cost on the back end — in an exam, a repapering, or the week a client asks who wrote the recommendation.

The 6% will not stay at 6% forever. Returns on a new tool tend to arrive after the tool stops being new, and the firms positioned to collect them are the ones whose masking, audit trail, and human sign-off were designed in from the first deployment rather than bolted on the week a regulator asks for the file. Mariner's $175 million answers the capacity question and leaves the control question open, and the number that resolves it will sit in the compliance headcount, whenever the 73% decide they need one.