Claude for Financial Advisors: the model is the commodity, the skills and connectors are the product
Anthropic launched Claude for Financial Advisors on September 14, and the model is the least interesting line in the announcement. What ships is a set of skill bundles (advisor onboarding, compliance and AI-policy review, portfolio rebalancing, pre- and post-meeting prep) plus connectors into the systems an advisory firm actually lives in: Schwab, BlackRock, Addepar, Envestnet, iCapital, Orion, Wealthbox and more. Around those sit the generic enterprise pipes, Microsoft 365, Salesforce, DocuSign, Box, and the data vendors FactSet and Morningstar. Analytics and risk tooling comes from BlackRock, Vanguard and others. It arrived a few days after OpenAI's finance version of ChatGPT, so both labs are now selling the same shape of thing inside one week.
I read this as a pricing statement more than a product launch. Nobody in the announcement argues that Claude is smarter than the alternatives at portfolio math. The pitch is that the work is already packaged.
A bundle is a job description
Look at the four skill bundles again. Onboarding a new advisor, reviewing material against a firm's AI policy, rebalancing a book, preparing for a client call and writing it up afterwards. None of these is a capability in the benchmark sense. Each one is a job, with inputs from three or four systems, a house style, a compliance checklist and a person who signs off at the end.
That is what a skill is when it grows up. A folder of instructions, reference material and small scripts that tells the model how this particular firm does this particular thing. If you have written one for your own repo (I covered the mechanics in Claude Code skills) you know the model is maybe a fifth of the effort. The rest is deciding what good looks like and writing it down in a way that survives contact with a messy Tuesday.
Now multiply that by a vertical. Somebody had to sit with advisors and compliance officers and decide what a sane pre-meeting brief contains. That knowledge is the product, and it does not depend on which model runs underneath. You could swap the model next quarter and the bundle would mostly keep working. Try saying that about a fine-tune.
Connectors are the unglamorous moat
The connector list reads like a sales deck, but it does real work. Schwab holds custody data. Addepar and Orion hold the portfolio and performance view. Envestnet sits in the middle of a lot of managed-account workflows. Wealthbox is where the client notes live. An advisor assistant that cannot read those is a chatbot with a finance vocabulary.
Each connector is also a negotiation. Somebody has to get an integration approved, agree on scopes, deal with the vendor's security review and keep it working when an API changes. This is slow, dull and very hard to copy quickly, which is exactly why it makes a better moat than a model score that gets overtaken every six weeks. I wrote about the protocol side of this in MCP explained: the wire format is the easy part, the permissions and the trust are not.
It also explains the BlackRock and Vanguard angle. If the analytics come from the firms advisors already trust for risk numbers, Claude is not asking anyone to believe in a new calculation. It is a new way to reach the old one.
If you sell AI to a vertical
This is the part I care about professionally. The pattern is repeatable, and it is not reserved for labs with a bank of integration lawyers. Pick a narrow profession. Find the four recurring jobs that eat their week. Write the skill for each, wire the two or three systems that hold the data, and put a human approval step where the regulator would expect one. Then sell the result per seat, not per token.
The awkward consequence is for anyone whose offer is a thin wrapper around a model. If a lab ships the vertical bundle itself, your wrapper competes with the platform's own default, and the platform has the model and the distribution. What survives is what the lab will not bother to build: the odd in-house system, the firm-specific policy, the training that gets forty advisors to actually use it on Monday.
I should be honest about what I do not know. I am working from the announcement and the coverage, not from a hands-on trial, so I cannot tell you how good any of the bundles are, how they are priced, or how much of the compliance review is automated versus merely drafted. In a regulated business that difference is most of the story. A rebalance suggestion a human checks is a productivity tool. A rebalance the system executes is a liability, and the announcement does not make clear which of the two this leans toward.
When every lab can rent you the same frontier model, the thing worth paying for is the packaged job and the plumbing it runs through.
So the question I would ask any vendor this autumn is short. If you replaced your model with a competitor's tomorrow, what would your customer actually lose? If the answer is nothing, you are selling the commodity. If the answer is forty workflows and nine integrations, you are selling the product.