Default no for unmanaged ChatGPT. Consumer chat is not a books-and-records system. A licensed model API behind a firm-owned context layer, with redaction, retention, and human approval, is the workable pattern.
Can financial advisors use ChatGPT with client data? The default answer is no if you mean unmanaged ChatGPT: the consumer product, a personal Plus login, or any chat window that is not under a firm contract with logging you control. That pattern creates an NPI, retention, and exam problem you cannot replay. The workable pattern is a licensed model API behind a firm-owned context layer, with redaction, retention, and a human approval step.
Do not paste households, IPS files, tax documents, or CRM exports into ChatGPT.
A licensed API, with a contract that restricts training on your data, can sit behind retrieval and policy you own. That is not "using ChatGPT with client data." That is borrowing model capability the same way you borrow any vendor processor: under a written control set your CCO can defend.
Public chat products are not your books-and-records system.
Three failures show up in an exam.
You cannot prove what left the building. If an advisor pasted a household file into a personal account, you do not have a complete record of the prompt, the output, or the destination. Retention is not optional because it is convenient.
NPI is now in a system you do not control. Names, account numbers, tax identifiers, and holdings in a consumer chat are a privacy event. "We told people not to" is not a control if the tool is still on their laptop.
The output is not an official record. A chat transcript is not a filed advertising review, a signed onboarding checklist, or a billing exception memo. If the work matters, it has to land in the system your examiner already knows.
Schwab's public AI guidance puts data governance and security as the condition for moving from experiment to strategy. Pasting client files into a generic chat product is the experiment you do not want in the file.
A contracted, licensed channel is a different object than a browser tab.
Enterprise or API access can be part of a workable stack when:
That is still not enough on its own. You also need the context layer: retrieval over CRM, IPS documents, meeting notes, and portfolio snapshots, plus the policy shell that redacts NPI before a model sees it. Keep embeddings and indexes inside your security boundary. Run models through contracted APIs.
Forbes's 2026 buy-versus-build piece is the right frame: you compose intelligence. You buy or borrow the model. You own the layer where client data and compliance controls live. That composition is spelled out in Build vs buy AI for RIAs.
Write the rules before anyone gets a login.
If a partner cannot answer where client files live after a prompt runs, keep looking. That question is in the partner screen in Who helps RIAs implement AI.
Licensed model access behind a firm-owned context layer.
The model does not browse your book. Retrieval pulls the fields the workflow needs from systems you already keep. Policy redacts NPI. The model drafts. A person who already owns the process approves. The artifact lands in CRM, the compliance archive, or the billing system.
That is how you implement AI at an RIA without turning ChatGPT into a second, invisible filing cabinet.
howtheF designs and operates that internal layer for US RIAs between $1B and $10B AUM. We do not sell a client-facing chatbot. First workflows are onboarding, compliance, billing, and trading ops.
Yes, if the firm allows it and the input is not client, prospect, or employee NPI. Marketing copy with no household data, internal process drafts, and generic research are a different risk class than CRM exports. Put that distinction in the written policy so people are not guessing.
Not by itself. A firm contract and better logging are necessary. They are not a context layer, a redaction policy, or an approval queue. Treat Enterprise as a licensed channel you still have to wrap.
Treat it as an incident under your existing privacy and vendor procedures. Preserve what you can. Stop the practice. Do not "fix" it by creating a second unofficial archive of the chats. Your CCO and counsel own the next step, not a vendor.
Ban unmanaged use with firm or client data. You do not have to ban licensed model access behind controls you own. The useful split is the same as the rest of the stack: commodity tools under contract, context layer under the firm.
Custom internal AI on your CRM, custodian, and compliance archive. Licensed models, firm-owned retrieval and policy, human approval, output written to a system your examiner already knows.
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.