@Tim Mrock , full disclosure, I run Prodos and we build this stuff for RIAs, so grain of salt. A few things I'd want to know first: - Where do requests come in today (email, Teams, CRM tasks)? That's usually where AI should plug in. - How many systems does each process touch? One tool is quick, three is a real project. - Does everyone need training, or just the 1-2 people who'd own it? - Re licensing, which tier gives you the audit trail and vendor oversight Reg S-P expects? - Have you decided which AI outputs count as books and records? I'll shoot you a DM with more detail.

A few thoughts on this since we've worked on adjacent parts before. - Tier by risk, not just amount. First-party transfers to bank accounts already on file should be nearly frictionless. Third-party payees, new instructions, and anything following a recent contact-info change get the full treatment. That's how you stay client-friendly: most requests are the easy kind. - Verify out-of-band, and never through email. If the request came by email, call the number of record. Never use contact info in the request itself. SMS OTP (DTS Verify is fine for this) works as the low tier, not for third-party wires. - Static challenge questions are mostly theater. "Mother's maiden name" and "first car" are in a breach dump somewhere. Use things only the two of you would know from actual calls or meetings, not from email threads, since a compromised inbox gives the attacker your whole history. - Plant a false flag on higher-risk requests. Reference an account or detail that doesn't exist and see if they go along with it. Fraudsters play along. Real clients correct you. This holds up against voice deepfakes, which are now good enough that "sounded like them" means nothing. - Log every verification: who, when, channel, what was confirmed (examiners will ask!) Happy to share more if useful.

Depends what you're trying to solve for. If you want accurate numbers, don't ask AI to count/summarize. Generally, AI is bad at math (unless you use coding models), which is why the built-in assistant gives you different answers on different days. Better path: get the data out of Wealthbox on a schedule into a spreadsheet or database, and let the AI write the lookup while something else does the actual counting. Same answer every time. Also, "keep the data in our own database" doesn't mean it stays private. Most setups have the AI read the results back so it can explain them to you, and the moment it does, those records go to OpenAI or Anthropic like anything else you type in. If that matters, it has to be built so the results go to the screen without the AI seeing them. There's also a middle option: strip the names, emails and account numbers out before anything gets sent, and swap them back in after. The AI works off the numbers and never sees who the client is. Tools exist for this, but you need someone technical to set it up and maintain it. Two things I'd check before spending money on this: 1/ Does the API actually have what you want? Contacts, custom fields, birthdays, workflows come out fine. Anything feeding in from other integrations, like balances, has been hit or miss for me. 2/ Read the terms of the plan you're actually on. Neither company trains on business data by default, but "we don't keep it at all" is a separate thing and not standard. And once data leaves Wealthbox & relates to planning & advice, theb retention and supervision under 204-2 are yours to manage. I build software for RIA back office ops and work with the Wealthbox API regularly, so I live in this tradeoff. Happy to talk it through if useful.

@Jill M Yaconiello CyberSecure has great reviews in the space. Happy to connect you with them. For the billing side - is the current workflow spreadsheet exports, transforming with your own rules & logic, then plugging it into a custodian to raise cash? If it's something along those lines, happy to chat more.

I would suggest Hadrius; they are awesome and used them at my last company https://www.hadrius.com/

@Bethany Leiker yes we were on GR too - was astonishing how big a UI/UX gap there was.