AI governance has to happen during the work

Most financial advisory firms have started with AI in a safe corner of the business. An adviser uses it to improve an email. A paraplanner asks it to summarise a meeting transcript. Someone drafts a social post.

That is useful, but it isn’t where the real operational value sits.

The larger opportunity is in work that touches client records, portfolio data, advice documents, KYC files, and compliance evidence. This is where an AI agent can reduce the manual effort behind meeting preparation, file notes, onboarding packs, and review documentation.

It is also where a generic AI policy stops being enough.

An annual policy review may say that staff must protect client data and check AI outputs. That doesn’t answer the practical question that matters at 10:17 on a Tuesday morning. Can the agent retrieve this client’s portfolio holdings? Can it read the latest CRM notes? Can it prepare an ROA draft? Can it send anything outside the firm? Does a compliance reviewer need to approve the next action?

These are runtime decisions. The controls must apply while the agent is working, not only before it is switched on.

That is why regulated industries are leading the move toward runtime governance. As the VentureBeat reporting on runtime AI governance makes clear, organisations are moving from broad AI principles to active controls around what agents can access, do, and record.

For an advisory firm, the basic standard should be simple. Any AI agent action involving client or portfolio data needs defined permissions, a visible activity log, an approval point where risk warrants it, and compliance review built into the workflow.

The manual work hiding behind a client review

A client review meeting looks simple from the outside. The adviser meets the client, discusses progress, identifies changes, and agrees on next steps.

The internal work is far less simple.

Before the meeting, someone gathers recent communications, checks portfolio performance, looks for outstanding tasks, reviews stated goals, finds the last advice document, and identifies material changes. In firms without a disciplined process, this can mean opening six systems and working from memory.

After the meeting, the adviser or paraplanner has to turn conversation into evidence. What was discussed? What did the client tell the firm? Was risk tolerance reconfirmed? Were recommendations made? What disclosures are required? Which tasks belong to the adviser, administration team, or client?

It is common to see advisers lose 5 to 10 hours each week across meeting preparation and post-meeting documentation. Some of that work is necessary. Much of it is repeatable retrieval, sorting, formatting, and chasing.

Advice documentation is where the cost grows. SOAs, ROAs, and detailed file notes can absorb $3,000 to $8,000 in paraplanner cost per advice document, depending on complexity, rework, and the review process. Cycle times can stretch into weeks when transcripts, source documents, client changes, and compliance feedback move back and forth across email.

Then there is onboarding. New clients may wait 30 to 60 days while the firm collects documents, completes fact-finding, follows up on KYC gaps, and prepares the adviser for the first substantive discussion. Every extra handoff creates a chance for the prospect to lose momentum.

The annual leakage across these workflows often sits in the $70,000 to $200,000 range for a firm in the $1 million to $25 million revenue band. That isn’t necessarily a line item anyone can see in the P&L. It appears as adviser capacity, slow turnaround, paraplanner rework, missed follow-up, and a client experience that depends too much on who happens to be available.

AI agents can address this work. But only if the firm puts proper controls around the agent before it gets access to real client information.

What runtime control means in an advice business

Runtime control is not a policy PDF that staff acknowledge once a year. It is a set of checks that run when an agent tries to take an action.

Think of it like the permission model in your custody platform or CRM. A staff member does not get unrestricted access because they work at the firm. Their role determines what they can see and change. Some actions need a second person to approve them. The system records what happened.

AI agents need the same treatment.

A controlled agent workflow should answer five questions for every meaningful action.

What data can this agent access?

The agent should receive the minimum information required for the task. A Meeting Prep Agent may need the current portfolio summary, recent client communications, stated objectives, previous meeting notes, and open service items.

It doesn’t need unrestricted access to every household, all client documents, or a full database export.

Permissions should be role-based and client-specific. If the adviser is assigned to a household, the agent can prepare work for that household. If a staff member has no servicing relationship, the agent should not retrieve the records on their behalf.

This is especially important when firms use multiple systems. An agent may pull from a CRM, portfolio reporting tool, document management platform, email archive, and workflow system. The fact that the connections exist does not mean every agent should use every connection.

What is the agent allowed to do?

There is a major difference between reading, drafting, updating, and sending.

Reading portfolio performance to create a meeting brief is one level of risk. Drafting a file note from an approved transcript is another. Updating a CRM task can be appropriate in some workflows. Sending advice, changing client details, submitting orders, or updating risk profiles should be locked down or prohibited unless the firm’s process specifically supports it.

Start with an action inventory. Write down each action an agent could take and put it in one of four categories:

• Read-only retrieval
• Draft creation
• Internal system update
• External or regulated action

The first two categories are where most advisory firms should begin. Internal updates can follow once accuracy is proven and a human review process is clear. External and regulated actions need the highest approval threshold, and many should remain human-only.

This is the operating logic behind Omni ops. The aim is not to hand an AI model the keys to the business. It is to build useful workflows with boundaries that staff can understand and manage.

Who approves the output?

Human approval should sit at points where an output becomes advice, a client record, a compliance record, or client communication.

For example, the Meeting Prep Agent can assemble a one-page review brief without an approval delay. The adviser still reads it before the meeting. If the brief identifies a material portfolio drift or a pending strategy issue, it should flag the source information rather than make a recommendation.

An Advice Document Agent can draft a structured ROA or file note from the meeting transcript and the firm’s approved template. It should not mark the document as final, issue it to the client, or treat a draft as compliance-approved. The adviser reviews the factual content. A designated compliance reviewer handles the required review under the firm’s procedures.

Approval doesn’t mean printing a draft and signing it manually. It means a defined workflow state. Drafted. Adviser checked. Compliance reviewed. Approved for issue. Issued. Archived.

Those states create clarity for the team and evidence for the file.

Can the firm see what happened?

Every agent action involving client or portfolio information needs a traceable record.

At a minimum, capture:

• The staff member who initiated the workflow
• The client or household record involved
• The source systems and documents accessed
• The information retrieved
• The prompt or instruction used, where appropriate
• The output created
• Any system updates attempted or completed
• The people who reviewed and approved the result
• The date and time of each step
• Exceptions, failures, and policy blocks

This doesn’t mean staff need to inspect a long stream of technical logs. They need a practical audit trail tied to the client record and workflow. If a compliance officer asks how a file note was produced, the firm should be able to show the transcript source, the template used, the draft, the review history, and the final stored version.

A vague statement that “AI assisted with this” is not enough.

What happens when the agent hits a boundary?

A good control system does not assume the agent will always have a clean task.

A client may mention a health issue, divorce, inheritance, complaint, or change in employment. The transcript may be incomplete. The KYC document may be expired. A portfolio record may be missing a valuation date. The agent needs a defined response to these conditions.

Usually, that response is one of three actions. Stop and ask a staff member for direction. Escalate to the relevant compliance or risk queue. Complete the safe parts of the task and flag what remains unresolved.

The agent should never quietly fill gaps with assumptions, create a recommendation from incomplete data, or continue past an access restriction because a user asked it to.

A controlled Meeting Prep Agent in practice

The Meeting Prep Agent is a sensible first build because its value is immediate and its permissions can be tightly defined.

Here is what a controlled end-to-end workflow looks like.

An adviser marks a client review as upcoming in the CRM. The agent receives the client ID, meeting date, adviser ID, and approved purpose of the meeting. It checks that the adviser is assigned to the household before accessing any information.

It then retrieves only the permitted data:

• Current portfolio summary and recent performance figures
• Stated client goals and planning priorities
• Recent meeting notes and completed action items
• Recent inbound and outbound communications
• Outstanding service tasks
• Documents that require renewal or review

The agent creates a one-page brief with three sections. What has changed since the last meeting. What needs the adviser’s attention. What questions should be clarified with the client.

It cites the underlying source records. If it detects conflicting information, such as two different retirement dates in the CRM and fact-find, it highlights the discrepancy. It does not decide which version is correct.

The adviser reviews the brief before the meeting. The system logs the data sources accessed and stores the completed brief within the approved client workspace. The agent cannot email the brief to the client, edit the client’s plan, or generate investment instructions.

That is runtime governance in practical terms. The control is part of the work, not an afterthought.

If your firm wants to identify the right first workflow and the boundaries it needs, Book a 60-min Omni Audit. We will work through the workflow, its data access, its human checkpoints, and the likely operational return.

Advice documents need stronger gates

The Advice Document Agent can create more value than a meeting brief, but it needs tighter controls because its output may become part of the formal client file.

A well-designed agent receives a meeting transcript, client record references, approved compliance template, and specific instructions about the document type. It extracts factual statements, maps them to the relevant template sections, and produces a draft SOA, ROA, or file note.

The key word is draft.

The agent should identify missing inputs rather than invent them. If the transcript says a client “wants to take less risk” but no updated risk profile has been completed, it should flag that issue. It should not translate the statement into a new risk classification.

It should also separate facts from interpretations. A phrase like “client is comfortable with the changes” might be included in a file note as a sourced statement. It should not be treated as proof that suitability requirements have been met.

The review workflow should include:

1. Adviser review for factual accuracy and context.
2. Paraplanner or documentation review for completeness.
3. Compliance review where the firm’s policy requires it.
4. Final approval before the document is issued or marked complete.
5. Archiving of the source transcript, draft versions, review comments, approval record, and final document.

This approach can reduce drafting and rework time without weakening compliance. It also gives the firm a clearer documentation process than the usual mix of handwritten notes, Word files, inbox searches, and memory.

For more examples of where operational AI belongs in a controlled environment, review the Omni platform and the practical material in our AI guides.

Onboarding agents need permission limits too

The Client Onboarding Agent is another workflow where firms can remove friction without allowing the agent to make decisions it should not make.

The agent can run a guided fact-find, request required KYC documents, track missing items, and prepare a clean onboarding pack for the adviser. It can remind a prospect that proof of address is still outstanding. It can extract fields from an uploaded document for staff review. It can identify when the fact-find contains an incomplete response.

It should not independently verify identity, classify a client as suitable, accept a risk profile, or decide that KYC is complete. Those actions have regulatory and operational consequences. They need human ownership.

Permission design matters here. A prospective client may provide sensitive documents before they are fully onboarded. Access should be limited to the staff and workflow roles responsible for onboarding. The agent should not make those records available to every adviser or administrator who can access the general CRM.

A good onboarding workflow also makes exceptions visible. If a client has not supplied documents after three reminders, the agent should place the file in an exception queue. If the fact-find reveals a complex trust structure or overseas tax issue, it should route the case to the appropriate person. It should not keep pushing the prospect through a standard sequence as though the issue does not exist.

Start with a workflow, not an AI tool

Many firms buy an AI tool first, then try to find a use for it. That creates scattered experimentation and an impossible governance task.

Start with one workflow where the team already feels the cost. Meeting preparation is often a good candidate. Advice document drafting can be next. Onboarding follows once the firm has learnt how to manage permissions, reviews, exceptions, and audit evidence.

For each workflow, document:

• The business outcome you want
• The current manual steps and handoffs
• The client and portfolio data required
• The specific actions the agent may take
• The actions it must never take
• The approval points
• The compliance evidence you need to retain
• The exception paths
• The owner responsible for ongoing control

This is not bureaucracy for its own sake. It is how you turn AI capacity into an operating process your advisers and compliance team can trust.

The AI audit for financial advisory firms is built around this question. We identify where work is leaking, which workflows are suitable for automation, and what controls need to exist before an agent touches live data.

The commercial case is capacity and control

For an owner or partner, the return is not simply fewer hours spent writing. It is more adviser capacity directed toward client conversations, new relationships, and decisions that need judgement.

If an adviser recovers even part of the 5 to 10 hours a week currently spent on preparation and notes, the firm gains usable capacity without adding another senior salary. If a paraplanner spends less time rebuilding documents after incomplete briefs or inconsistent file notes, cycle times become easier to manage. If onboarding moves faster, new clients are less likely to stall before the relationship begins.

The firms that benefit most will not be the ones that allow an AI agent to do anything. They will be the ones that define exactly what the agent can do, require people to review what matters, and retain evidence of every important action.

That is the point of runtime control. It makes AI usable in a regulated business.

You can see Omni for financial advisory firms to understand how we assess the workflow opportunities, control requirements, and likely leakage in your firm. Or, if you are ready to map a specific process with us, Book my Omni Audit.

It is 60 minutes. You leave with three outputs: a view of the highest-value workflow, the runtime controls it requires, and a practical next-step plan. No deck, no generic AI strategy session.