An AI governance framework is the set of decisions, owners, and records that define how a firm uses AI. For an RIA, it covers an inventory of AI tools and workflows, who approves them, who can access which data, how sensitive data is handled, what gets logged, and how the firm reviews it over time.
Why an RIA Needs Its Own AI Governance Framework
Most published guidance on AI governance is written for large enterprises with dedicated risk teams. An RIA has a different shape. Client records sit in a CRM such as Salesforce, positions and balances come from a custodian such as Schwab, and reporting often runs through a platform such as Orion or Black Diamond. A framework for your firm has to account for how information moves between those systems and what an AI tool can reach along the way.
If you are still deciding where AI belongs in your firm, start with our guide to AI in wealth management, then return here to formalize the decisions.
The Parts of an AI Governance Framework
A workable framework for an RIA can be organized into six parts. Each part should have a named owner.
- Inventory of AI tools and workflows
- Ownership and approvals
- Access scoped to the user and workflow
- Data handling and redaction
- Logging and retention
- Ongoing review
Treat the sections below as prompts for your own policy, not as a standard your firm must adopt.
Part 1: Inventory of AI Tools and Workflows
You cannot govern what you have not listed. Start with a register of every AI tool and workflow, including informal ones. Staff may already use AI for drafting or summarizing, and a register of approved tools alone will miss that.
For each entry, record:
- The workflow, such as meeting preparation, onboarding, or drafting a follow-up email
- The tool and model involved, and who approved them
- The systems that supply data, such as your CRM, custodian, or planning tools
- The roles that use it, and where prompts, outputs, and records are stored
Operations typically owns the register because it sees how work really gets done. Compliance decides which entries need a closer look. To design a workflow before you add it, see software automation in wealth management.
Part 2: Ownership and Approvals
Every workflow needs one accountable person, and the framework needs a clear path for approving new ones. Decide these points in writing:
- Who is accountable for each workflow, and who is their backup
- Who approves a new tool, a new model, or a new data connection
- Who reviews AI-assisted output before it is used, and which outputs need additional approval before they reach a client
- How a change to a model, prompt, or connection is reviewed before it goes live
Many firms assign approval to a small group that includes operations, compliance, and a principal. What matters is that an approval is a recorded decision rather than an informal agreement.
Part 3: Access Scoped to the User and Workflow
A general permission to use an AI tool is not the same as permission to see every connected record. The framework should state that access depends on who the user is and which workflow they are running.
Decide the following:
- Which roles can use each workflow
- Which client data each role may reach inside that workflow, and whether it matches their access in the source system
- How access changes when someone changes roles or leaves the firm
- Who can change workflow settings, models, or connected data sources
A workflow connected to Salesforce, Schwab, Orion, or Black Diamond can widen what a user sees if it does not honor the permissions of those systems. Ask any vendor how it checks authorization before data is made available, and test the answer with real users.
Part 4: Data Handling and Redaction
This part defines what information a model may receive. Start from the task: if a workflow does not need a field, the field should not reach a model.
Decide the following:
- Which data categories are permitted, restricted, or prohibited in each workflow
- Which fields are masked or withheld before a request reaches a model, and who signs off on that list
- Which models are approved, and where your data is processed
- Whether your firm's prompts and workflows are kept separate from a vendor's other customers
- What staff may paste into tools outside an approved workflow
Redaction rules are rarely identical across workflows. Setting handling rules per workflow, rather than one blanket rule, can reduce both over-sharing and broken workflows, though it adds configuration your team has to maintain.
Part 5: Logging and Retention
Records let you answer questions after the fact. The framework should say what is recorded, who can read it, and how long it is kept.
Decide the following:
- Which events are recorded, such as the request, the user, the workflow, and the time
- Whether records can be searched and exported
- How long records are kept, and how that matches your firm's own retention policy
- Who can read, change, or delete records
Keeping more records supports traceability, while keeping fewer limits how much sensitive information accumulates. Which records your firm must keep, and for how long, is a question for your compliance advisers and the rules that apply to you.
Part 6: Ongoing Review
A framework that is written once and never revisited drifts out of date. Set a review rhythm and a set of triggers for an earlier review.
- A scheduled review of each workflow, such as quarterly for higher-risk workflows
- Triggers for an unscheduled review, such as a new model, a new data connection, a change of vendor, or an error found in output
- A sample test of access, redaction, and logging against what the policy says
- A record of findings, owners, and due dates
Our checklist on how to audit AI workflows walks through a full review of one workflow, including the evidence to collect.
Generative AI Governance Framework Considerations
A generative AI governance framework follows the same six parts, with additions because generative tools produce new text rather than returning stored records.
- Treat output as a draft for human review, not a finished client communication
- Consider how the workflow limits attempts to manipulate a model into exposing information
- Decide who reviews errors and invented details, and how they are recorded
A reviewer still decides whether an output is accurate and appropriate for its audience.
AI Governance Framework Template: Fields for Each Workflow
If you want a working template, a single page per workflow is often enough. Use these fields as a starting point and adapt them to your firm's policies.
- Workflow name and purpose
- Accountable owner and backup
- Approved tool and model, with approval date and approver
- Data sources and the specific fields used
- Roles permitted to use it, and how access is checked
- Redaction and data-handling rules
- Output reviewer and any additional approvals
- Records kept, retention period, and who can read them
- Review schedule, last review date, and open findings
Keep these pages together so a vendor review or a compliance request can be answered from one place. Whether a document satisfies a regulator is for your compliance team to decide.
Common Questions About AI Governance Frameworks
What an AI governance framework is
It is a written set of owners, controls, and review steps for how a firm approves and uses AI, usually organized around workflows and the data they touch.
Who should own an AI governance framework
Ownership is usually shared. Operations tends to own the inventory and workflows, compliance tends to own policy and review, and a senior leader approves the approach. Each workflow still needs one accountable person.
How a framework differs from an AI policy
A policy states rules, such as acceptable use. A framework includes the policy and also assigns owners, approvals, controls, and reviews.
Whether a framework is required
Requirements depend on your firm, your regulators, and your clients. Ask your compliance lead or counsel which obligations apply to the way you use AI.
How Ariath Fits Into Your Framework
Ariath's product is designed to support several parts of a framework. It checks role-based access for the specific user before connected data reaches a model, supports configurable redaction and data-handling controls per workflow, configures audit and retention settings around each workflow, and keeps workflows and prompts in a tenant-specific deployment. Its AI compliance page maps each control to questions your team can ask in a vendor review.
These controls are operational. They do not certify that a firm or workflow meets any regulation, and what they capture, mask, or keep depends on how your firm configures them. Your firm still owns its policies, its review of AI output, and its determination of what applies to it.
If you want to walk through your own workflows and the decisions in each part, Book a Demo.