A five-step plan for US RIAs: pick one internal workflow, map it to books-and-records, run a licensed model behind a firm-owned context layer, require human approval, then measure hours before you expand.
How to implement AI at an RIA is a five-step ops problem, not a software purchase. Pick one internal workflow, map every field to a books-and-records system you already keep, run a licensed model behind a firm-owned context layer, put a human on the approval step, and measure hours on that process before you add a second. For US RIAs between $1B and $10B AUM, that sequence beats a firm-wide "AI transformation" and it beats pasting households into ChatGPT.
Start with one painful internal process: onboarding, advertising review, fee billing, or trade exceptions. Do not start with a client-facing chatbot.
Buy the note-taker if you need one. Hire for the context layer if the work crosses CRM, custodian, and compliance archive. Schwab's public guidance is that firms are moving from experiment to strategy. Strategy here means one governed workflow in production, not a second copilot login.
The first workflow has to be internal, repeatable, and already owned by a person who can approve the output.
Good first scopes:
Bad first scopes: a firm-wide knowledge bot, a client advice assistant, or "AI everywhere" on the CRM. Those fail the exam test and they fail the hours test because you cannot measure a blur.
Hartford Funds and WealthTech Today both treat operations and data discipline as the reason AI sticks. That is the right diagnosis. Pick the process your ops lead already hates.
Implementation is not a pilot login. It is a workflow with a source of truth, a retention rule, and a human approval step.
You can buy the note-taker in week one. You should not buy a second copilot until the first internal workflow has a replayable audit trail.
A longer treatment of who should run this with you is here: Who helps RIAs implement AI.
Three controls have to exist before the model runs.
Source of truth. CRM, custodian, document archive, or billing system. Not a shared drive of exports and not a chat history.
Retention. Prompts, outputs, and the human decision sit in a system your examiner already knows, for as long as the underlying record has to live.
Approval. A named person, already in the operating model, signs off. If you cannot name that person, the workflow is not ready for a model.
howtheF's refusal list is short and public: no client-facing advice bots, no NPI in plaintext prompts, no "set and forget" agents on regulated work.
After the first workflow runs in production. Not after a demo. Not after a strategy offsite.
Expansion means a second internal process with the same five steps, not a wider model and the same missing controls. Forbes's 2026 buy-versus-build piece is useful here: firms compose intelligence. You compose a second workflow the same way you composed the first.
If the first workflow does not run in production, do not expand the scope.
The composition choice (buy the commodity layer, own the context layer) is spelled out in Build vs buy AI for RIAs.
We start with one internal workflow, not a firm-wide program. The first deliverable is a production path: data stays put, the model is contracted, a person signs off, and the artifact lands in the system your examiner already knows.
We design and operate custom internal AI for wealth firms. We do not sell a client-facing chatbot. If you do not have a standing engineering team, that is the default case we work in.
One internal workflow with a named owner and a books-and-records destination. Onboarding, compliance review, billing exceptions, or trade breaks. Not a client-facing assistant and not a firm-wide knowledge bot.
No. You need a mapped source of truth for the fields that workflow uses. Warehouse programs are how first implementations die. Map the fields you will actually send through the context layer.
Not with client data. Unmanaged chat products are not your books-and-records system. Licensed model access behind a firm-owned context layer is the workable pattern.
Time-to-value should be measured in weeks on one workflow, not a multi-year platform program. If a partner cannot name the first production artifact, they are selling strategy, not implementation.
Custom internal AI for wealth firms. The product is the operating system around your existing stack: retrieval, policy, approval, and the workflow UI your team uses every day.
An AI CRM is not a native feature on a vendor roadmap. Keep Wealthbox, Redtail, or Salesforce. Buy the note-taker. Own a context layer across CRM, custodian, and the compliance archive, with human approval before anything writes to a household record.
Automate RIA client reporting as internal ops: retrieve verified performance data, draft commentary, require human review, and publish to the portal you already have. Not a client-facing advice bot and not a new reporting platform.
Buy commodity tools when the vendor owns the artifact. Build or hire a firm-owned context layer on CRM, custodian, and the archive, scoped to one internal workflow, when off-the-shelf cannot file the packet. For US RIAs between $1B and $10B AUM.
See how HowTheF.ai can help your firm implement AI that actually works.