Meet Caddi in personADVISE AIOct 20–22AI for Mid-Sized LawNov 5Legal InnovatorsNov 17–18
BlogAI for RIA operations

The Wealth Advisory COO's Guide to AI

What AI is good for in an RIA back office today, why reliability is the question that matters, what your CCO will ask, who should build it, and a 90-day plan you can take to your CEO.

In 2026, AI is genuinely useful in RIA operations for one kind of work: high-volume, rule-bound tasks that move data between systems, such as filling account opening paperwork from the CRM, catching NIGO errors before a form goes out, keeping CRM and portfolio accounting in sync, collecting onboarding documents and filing records. It is not a replacement for judgment, client relationships or compliance sign-off. The firms getting results use AI to build automations that run the same way every time, keep a person on the exceptions, and log every run so the CCO can review it.

This guide is for the COO, VP or Director of Operations who has been asked "what are we doing about AI?" and needs an answer that survives contact with the CCO, the partners and the team doing the work. If you want the broader picture of the role itself, start with what an RIA director of operations owns.

The four kinds of AI hitting your firm, and which one is yours

Most "AI for wealth management" conversations blur four different categories of product. They have different buyers, different risks and different owners. We cover the individual tools in AI for financial advisors; here is how they land on an operations leader's desk.

CategoryWhat it doesWho usually buys itThe COO's role
Meeting assistantsRecord or listen to client meetings, summarize, draft follow-ups, write notes and tasks to the CRMAdvisors and the CEOGovernance: where transcripts live, retention, CRM data quality, what happens to the tasks it creates
Planning analysisRead tax returns, statements and plans to produce planning inputsPlanning team and advisorsVendor due diligence and data flows; rarely day-to-day ownership
General assistantsCopilot or ChatGPT style drafting, summarizing and researchIT and the whole firmAcceptable use, what client data may be pasted in, retention
Back-office automation and agentsDo the operations work itself: paperwork, data entry across systems, reconciliation exceptions, document collection, filingOperationsOwner. This is your budget, your process and your risk.
The four categories of AI an RIA is likely to adopt, and where operations sits in each.

The first three categories create work for operations (a meeting assistant produces tasks that someone has to act on) but they are mostly other people's decisions. The fourth is the one that changes how your team spends its week, and the one the rest of this guide is about.

Where AI helps in the RIA back office

The pattern to look for is work that is frequent, follows rules you could write down, and involves copying information from one system into another. That describes a large share of what an RIA operations team does. The table below covers the workflows that come up most often; for step-by-step versions, see RIA back-office automation workflows.

WorkflowWhat the work isWhy it repeatsWhat a good automation doesWhat a person still owns
Account opening paperworkFilling custodian new account applications, advisory agreements and disclosures for each householdEvery new account, every account type, same fields retyped from the CRMFills the PDF forms from the CRM household record, sends them for e-signature, logs the signed copy back to the CRMChoosing account types and registration; submitting in the custodian portal
NIGO preventionCatching missing signatures, mismatched names, absent trusted contacts before submissionThe same few errors cause most rejectionsChecks each packet against a written rule list before it goes out and flags what is missingFixing the flagged item with the client or advisor
CRM data hygieneKeeping households, contacts, account numbers and service tiers consistentEvery system change or meeting creates driftCompares the CRM with portfolio accounting and reports or corrects mismatches on a scheduleDeciding which system is right when they disagree
Address and beneficiary changesUpdating client details across CRM, forms and recordsLife events are constant across a book of clientsPrepares the change form from the request, routes it for signature, updates the CRMVerifying the request is genuine; identity checks
Billing checksConfirming fee schedules, billable AUM and household groupings before billing runsQuarterly or monthly, with the same exceptions each timePulls the billing file, checks it against the fee schedule on record, lists exceptionsApproving the bill and resolving disputes
Reconciliation exceptionsWorking the exception list from portfolio accountingDaily, with recurring patternsSorts exceptions by known cause and drafts the fix or the note for eachAnything outside the known patterns
Reporting prepAssembling quarterly packets, pulling data from reporting tools into templatesEvery quarter for every householdGathers the reports and documents per household and stages them for reviewReview and release to clients
Onboarding document collectionChasing IDs, statements, trust documents and signed formsEvery new client, with multiple remindersSends the checklist, tracks what has arrived, files it and nudges on a scheduleThe client relationship and exceptions
Books-and-records filingSaving signed documents and correspondence to the right place with the right nameEvery document, every dayNames and files each document to the document store and links it in the CRMRetention policy and access rules
M&A repaperingRe-executing advisory agreements and account forms after an acquisition or custodian moveHundreds or thousands of households on the same form setGenerates each household's packet from the CRM, tracks signatures, reports statusThe consent strategy (ask counsel) and client communication
Common RIA operations workflows and how the work splits between an automation and a person.

If an acquisition is on your calendar, the RIA acquisition integration checklist and the breakaway advisor transition checklist walk through where repapering and data migration sit in the plan. For the paperwork mechanics, see how to auto-fill custodian forms.

Notice what is not in the table: deciding on an investment recommendation, handling an upset client, choosing how to register a trust account, or signing off on a compliance exception. Those stay with people. A good automation hands them a clean, complete packet and gets out of the way.

Reliability: the question that decides everything

Operations work has a different standard from drafting an email. A draft that is 90% right is useful. A transfer form that is 90% right is a NIGO rejection, a delayed account and an unhappy client. So the question to ask of any AI in your back office is not "how smart is it?" but "will it do the same thing, correctly, on run number 500?"

This is where a whole category of product deserves scrutiny: agents that reason through every step of a process from scratch on every run. Each step involves a model making a fresh decision, and small error rates compound. If each step is right 95% of the time and errors are independent, the chance that a run with many steps is right from start to finish is 0.95 multiplied by itself for every step:

Steps in the process95% per step99% per step
5 steps0.95^5 = 77%0.99^5 = 95%
10 steps0.95^10 = 60%0.99^10 = 90%
20 steps0.95^20 = 36%0.99^20 = 82%
Probability that an entire run is correct when every step is decided fresh, assuming independent errors. Calculated: per-step accuracy raised to the number of steps.

A 20-step process (open the household, read the registration, pick the form, fill 40 fields, check signatures, route, file, log) is routine for account opening. At 95% per step, roughly two out of three runs would have something wrong somewhere. Even at 99% per step, nearly one in five would. Real errors are not perfectly independent, so treat the numbers as an illustration of the shape, not a forecast. The shape is the point: fresh reasoning on every run gets less reliable as the process gets longer, which is exactly the kind of work operations owns.

What to require instead:

  • Build with AI, run as code. Use AI to turn your process into an automation, then run that automation as fixed, verified steps. The same input produces the same output.
  • Scope any model step. Some steps genuinely need a model, such as reading a scanned trust document. Those steps should be narrow, with a defined input and output, and checked before the result is used.
  • Log everything. Every run, every input, every output, every model step, kept so compliance can review it.
  • Route exceptions to a person. When the automation hits something outside its rules, it should stop and ask, not guess.
The line to hold vendors to: AI builds it, verified code runs it, and any model step inside a run is scoped, logged, and reviewable.

Governance and compliance: what your CCO will ask

Your CCO will not ask whether the AI is impressive. They will ask how it fits obligations the firm already has. None of this is legal advice; the judgment calls belong to your CCO and counsel. But you should arrive with answers to these:

  • Books and records. Advisers Act Rule 204-2 requires keeping, among other things, written communications relating to advice and transactions, preserved "in an easily accessible place" for at least five years from the end of the fiscal year of the last entry. If an automation sends client emails or files signed forms, those outputs are likely records. Where do they land, and can you produce them?
  • Compliance program. Rule 206(4)-7 requires written policies and procedures reasonably designed to prevent violations, reviewed at least annually. An automation that changes how work is done may need to be reflected in those procedures and in the annual review.
  • Customer information. Regulation S-P governs how you protect client nonpublic personal information. The SEC's 2024 amendments added a written incident response program and a requirement to notify affected individuals as soon as practicable, and no later than 30 days after becoming aware of unauthorized access. Every AI vendor that touches client data is part of that picture.
  • Supervision of AI. The SEC Division of Examinations' fiscal 2026 priorities say the Division will assess whether firms have policies and procedures "to monitor and/or supervise their use of AI technologies," naming back-office operations among the tasks, and will review representations about AI capabilities for accuracy. The same document lists oversight of third-party vendors as a focus for Regulation S-P and S-ID exams.
  • Vendor due diligence. Security attestations, subprocessors, where data is stored, what the vendor's model providers retain, and what happens to your data at contract end.
  • Access and permissions. Does the automation run under a named person's existing permissions, or does it need a super-user service account that sees everything?
  • Retention of logs. How long are run logs kept, and can you export them for an exam?

For how Caddi supports compliance teams specifically, see Caddi for RIA compliance.

Questions to take to any AI vendor

  1. Runtime: On an ordinary run, which steps are fixed code and which are a model deciding? Can you show me the run log for both?
  2. Repeatability: If I run the same input twice, do I get the same output? How do you test that?
  3. Exceptions: What happens when a run hits something it has not seen? Does it stop and ask, or continue?
  4. Records: Where do outputs and logs live, for how long, and in what format can I export them?
  5. Data: What client data do you store, where, and what do your model providers retain? What is your security attestation?
  6. Permissions: Whose credentials does the automation use, and can I limit it to specific systems and records?
  7. Change: When a custodian revises a form or we change a procedure, who updates the automation, and how long does it take?
  8. Ownership: If we leave, what do we keep?

Who builds it: four models, honestly compared

After you know what to automate, the biggest decision is who builds and maintains it. There is no single right answer; there is a right answer for your team's capacity and how often your processes change.

ModelWho does the workWorks well whenWatch out for
IT projectInternal IT or developers build integrations and scriptsYou have engineers and a stable, high-value processQueue time, and every change to a form or procedure goes back into the queue
ConsultantsAn outside firm scopes and builds for a feeA one-off migration or a defined project, such as a custodian moveCost of each change after the engagement ends; knowledge leaves with them
Vendor expertsThe vendor's own team builds and maintains agents for youYour ops team has no spare capacity and wants it done for themYou depend on the vendor's queue for changes; confirm how changes are priced
Your ops team owns itThe people who do the work build and change the automations on a platformProcesses change often and you want control and speedNeeds a few hours of an ops person's time to start, and a clear owner
Four ways RIAs get back-office automation built, with the trade-offs operations leaders report most often.

The vendor-expert model is a real option and the right one for some firms. PitCrew, for example, pairs its own experts with the platform and publicly describes formal verification and human approval for its agents; see Caddi vs PitCrew for a direct comparison. Caddi is built for the fourth model: teams that want their own operations people to own the automations and change them as the work changes. Firm size shapes the choice too; see our picks for small, midsize and large RIAs.

A 90-day plan for AI in RIA operations

The plan below is built to produce evidence, not a demo. By day 90 you should have two workflows in production, numbers you trust and a business case your CEO can say yes or no to.

Weeks 1 to 2: find the work

  • Inventory: list every recurring task the ops team does, with frequency, minutes per instance and the systems involved. A discovery tool can do much of this from the tools you already use; a shared spreadsheet also works.
  • Baseline: for the top candidates, record current volume, time per item, NIGO or rework rate and time to open an account.
  • Deliverable: a ranked list of ten candidate workflows and a baseline sheet. Share it with the CCO now, not at the end.

Weeks 3 to 6: automate two workflows

  • Pick two: one high-volume paperwork workflow (account opening or transfers) and one data workflow (CRM hygiene or document filing). Clear rules, frequent, low client-facing risk.
  • Build with the person who does the work. The ops specialist who runs the process today should define the steps and the exceptions.
  • Run in parallel: for the first two weeks a person checks every output before it goes anywhere.
  • Deliverable: two workflows in production with a named owner, a written exception procedure and a run log.

Weeks 7 to 12: govern and measure

  • Move from checking every run to checking exceptions when the parallel period shows the outputs match.
  • Update procedures: work with the CCO to reflect the automations in your written policies and procedures and your records map.
  • Measure against the baseline every two weeks.
  • Deliverable: a one-page results summary, a governance note signed off by compliance, and a shortlist of the next five workflows.

Day 90: the business case to your CEO

Keep it to one page: what you automated, the before and after on the measures below, what compliance reviewed, what it cost, and what you propose next with expected hours. Frame the result as capacity (households per ops person, accounts opened per week without new hires) rather than headcount cuts. That is the argument that resonates with partners who are growing through recruiting or acquisition.

How to measure it

MetricHow to measureWhy partners care
Hours returnedRuns completed multiplied by baseline minutes per item, less review timeCapacity to grow without hiring
Runs completedCount from the run log, by workflowShows the automation is actually used
Exception rateRuns routed to a person, divided by total runsTells you whether the rules are right
NIGO rateCustodian rejections divided by submissions, before and afterFewer delays and client friction
Time to open an accountSigned agreement to account open, medianClient experience and faster asset transfer
Error escapesMistakes found after release, from the log reviewWhat the CCO will ask first
Operations metrics to baseline before automating and report monthly afterward.

Report the same six numbers every month, in the same format, to the same audience. Consistency builds trust faster than a large number does. For more on where ops time goes in the first place, see RIA operations challenges.

Where Caddi fits

Caddi is an operations AI platform for teams that want to own their automations. It follows the same three stages as the plan above:

  • Discover finds the repetitive work in the tools your team already uses, which replaces most of the manual inventory in weeks 1 and 2.
  • Automate builds the automation: someone on the team screen-shares the process or describes it in chat, and they change it in plain English as the work changes. It connects to the RIA tools in your stack, including Salesforce, Wealthbox, Practifi, Orion, Black Diamond, DocuSign and Box, plus meeting assistants Jump and Zocks. For custodian paperwork, Caddi fills the PDF forms from the CRM household record, sends them for signature through DocuSign and logs them back to the CRM; work inside custodian portals stays with your team.
  • Govern gives firm-wide visibility into the work done by AI and by people, with every run logged.

AI builds it, verified code runs it, and any model step inside a run is scoped, logged, and reviewable. Caddi is SOC 2, data is encrypted in transit and at rest, and automations run under your team's existing permissions.

0%
of agent runs complete successfully
0
hours automated for every 1 hour spent teaching Caddi (average)
0
minutes on average to build the first agent
0
hours of repetitive work discovered per customer
Caddi platform averages. Hours discovered is a conservative figure from small first rollouts.

"We thought our processes were finely tuned, but every one still pointed back to a human. Caddi got the systems we already have to talk to each other, and gave my team back an hour a day."

Matt Mercer, COO, The Planning Center, an independent RIA

Caddi's customers include a Barron's Top 10 RIA. Discover and Automate are available on the free Individual plan; Govern starts on the paid team plan. See how the pieces map to a real stack in the RIA tech stack guides and on the Caddi for wealth management page.

When Caddi is not the right fit: if you want a vendor's team to build and run the automations for you, a vendor-expert model will suit you better. If your main problem is meeting notes, start with a meeting assistant such as Jump or Zocks (Caddi does not take meeting notes; it connects to them). And if the work lives entirely inside a custodian portal, Caddi will not reach it.

More for RIA operations leaders

Caddi connects to the systems an RIA back office runs on: Salesforce, Wealthbox, Practifi, Orion, Black Diamond, Tamarac, Addepar, DocuSign, Jump, Zocks. See how it runs client onboarding and the operational side of RIA compliance, or the RIA operations overview.

Caddi

See how Caddi AI Agents can give your operations team back hours every week and turn your back-office workflows into governed automations

Or sign up free

Frequently asked questions

What is AI actually good for in RIA operations?

AI is good at the repetitive back-office work that moves data between systems: filling account opening paperwork from the CRM household, catching missing fields before a form goes out, keeping CRM records in sync with portfolio accounting, collecting onboarding documents and filing records. It is not a substitute for judgment calls, client relationships or compliance sign-off. The best results come from automations built to run the same way every time, with a person owning the exceptions.

How should an RIA COO start with AI?

Start with an inventory, not a tool. Spend two weeks listing the repetitive work your team does, how often it happens and how long it takes. Pick two workflows with high volume and clear rules, such as account opening paperwork and onboarding document collection. Automate those, run them with a person reviewing exceptions, measure hours returned and error rates, then take that evidence to your CEO before expanding.

Why is an AI agent that reasons through every step risky for operations work?

Small error rates compound across steps. If each step of a 20-step process is right 95% of the time and errors are independent, the chance the whole run is right is 0.95 to the 20th power, about 36%. Operations work needs the same result every time. The safer pattern is to use AI to build the automation, run it as verified code, and keep any model step inside a run scoped, logged and reviewable.

What will my CCO ask about AI in the back office?

Expect questions about books and records under Advisers Act Rule 204-2, how the tool fits your compliance program under Rule 206(4)-7, how customer information is protected under Regulation S-P, what due diligence you did on the vendor, who reviews AI output before it reaches a client or a record, how long logs are kept and what permissions the tool runs under. The SEC's 2026 exam priorities name supervision of AI used in back-office operations.

Should an RIA build AI automation in-house or hire a vendor to build it?

It depends on who you want owning the automations a year from now. A vendor-expert model, where the vendor's team builds and maintains agents for you, suits firms without operations capacity to spare. A platform your ops team owns suits firms that want to change workflows themselves as custodian forms, systems and procedures change. Traditional IT projects and consultants tend to be slow and leave you dependent on them for every change.

How do you measure the ROI of AI in wealth management operations?

Measure what partners already care about: hours returned to the team, number of runs completed, exception rate, NIGO rate on custodian paperwork and time from signed agreement to open account. Take a baseline before you automate, using a two-week sample, then report the same numbers monthly. Translate hours into capacity, such as households each ops person can support, rather than headcount cuts.