How to Audit AI Workflows: A Practical Checklist for RIAs

To audit an AI workflow, review one defined workflow from start to finish and compare what happens in practice with what your firm's policies say should happen. Check the tools and data it uses, who has access, which safeguards apply, how people review outputs, and what records exist. An audit produces evidence and actions, not proof of compliance.

This article is educational and is not legal or compliance advice. Your compliance team determines which rules and standards apply to your firm and how to document your review.

Scope the Audit to One Workflow

Auditing "AI at the firm" is too broad to finish. Pick a single workflow, such as preparing client meeting materials, drafting a follow-up email, or summarizing onboarding documents, and write down its boundaries before you start.

• Name the workflow and the single person accountable for it.
• State the purpose of the audit: a first review, a periodic check, or a response to a change.
• List the teams and roles that use the workflow.
• Record the period you are reviewing, such as the last quarter.
• Decide who will perform the review and who will receive the findings.
• Note what is out of scope, so a gap is not mistaken for coverage.

If you are still deciding which workflows to bring into the firm, start with our guide to AI in wealth management. If you need to design a workflow before reviewing it, see software automation in wealth management.

Inventory the Tools and Data Paths

You cannot review what you have not listed. Trace the workflow from the request to the final output and write down every place information travels.

• Which AI models or tools does the workflow use, and has your firm approved each one?
• Which systems supply information, such as your CRM, custodian, or planning tools?
• Which specific data fields reach a model, and are all of them needed for the task?
• Where do prompts, outputs, and logs get stored?
• Are there informal workarounds, such as staff pasting client details into a tool outside the workflow?
• Which vendors or third parties handle the information along the way?

The last two questions often reveal the biggest gaps. A documented workflow can be sound while an undocumented side channel is not, so ask people how the work really gets done.

Review Who Has Access

Access review checks that each person can reach only what their role and the task require.

• Who is permitted to use the workflow, and who approved that access?
• Does the system check each user's authorization before data is made available, or does it rely on one blanket permission?
• Do the permissions in the AI workflow match the permissions in the source systems?
• Has anyone changed roles or left the firm since the last review, and was their access updated?
• Who can change the workflow's settings, models, or connected data sources?
• Are administrator and service accounts identified and reviewed?

Sample a few real users and compare what the workflow lets them see with what their role should allow. A written policy is useful, but a sample test is stronger evidence.

Document the Safeguards

Next, list the safeguards that are meant to apply to the workflow and confirm each one is actually configured.

• Data handling: what sensitive information is redacted or withheld before a request reaches a model, and who decided that?
• Model approval: is every request routed only to models your firm has approved?
• Written rules: do your policies describe acceptable use, prohibited data, and exceptions?
• Vendor review: has your firm completed its own due diligence on each provider, and is it documented?
• Change control: how are updates to prompts, models, or connections reviewed before they go live?
• Known weaknesses: have you considered attempts to manipulate a model into exposing information or taking unintended actions, and what limits exist?

For each safeguard, record the control, who owns it, and the evidence that it is in place, such as a settings screenshot, a policy section, or a test result.

Check Output Oversight

Human review is part of the workflow, not an afterthought. Confirm that it happens and that it is meaningful.

• Is a named reviewer assigned to the outputs, and does that person see enough context to judge them?
• Does your firm treat AI-generated material as a draft that needs review before use, particularly for anything that reaches a client?
• How are errors, invented details, or inappropriate content reported and corrected?
• Is there a record of who reviewed what, and when?
• Which outputs are not allowed to leave the firm without additional approval?

Pull a handful of recent outputs and trace each one back to its inputs and its reviewer. If you cannot reconstruct that path, note it as a finding.

Test Logging and Retention

Records are what let you answer questions after the fact. Test them rather than assuming they exist.

• Does the workflow keep a record of requests, the user who made each one, and the time?
• Can your team search those records and export them when needed?
• How long are records kept, and does that match your firm's own retention policy?
• Who can read, change, or delete the records?
• Do the records themselves expose sensitive information that should be limited?
• Can you reconstruct one specific past output from the logs alone?

Run a simple test. Choose a recent output, ask a colleague to find its record, and time how long it takes. Which records your firm must keep, and for how long, is a question for your compliance advisers and the rules that apply to your firm.

The Audit Checklist

Use this condensed list as a working template for each workflow review.

• Scope: one workflow, one accountable owner, a defined review period.
• Inventory: every model, data source, storage location, and vendor is listed and approved.
• Data paths: only the fields needed for the task reach a model, and side channels are identified.
• Access: permissions are role-based, checked before data is used, and current after role changes.
• Safeguards: redaction, model approval, and change control are configured and evidenced.
• Output review: a named reviewer checks outputs, and the review is recorded.
• Logging: requests and users are recorded, searchable, and retrievable on request.
• Retention: records are kept according to the firm's own policy and are protected from unauthorized changes.
• Findings: each gap has an owner, an action, and a due date.
• Follow-up: the workflow is reviewed again on a set schedule and after any significant change.

Turn Findings Into Actions

An audit that ends with a document and no follow-through has limited value. Rank findings by how much sensitive information or client-facing work they touch, assign each an owner and a date, and note what you will retest. Keep the evidence together so you can show later what you reviewed and what you changed. Share the results with your compliance team, which can decide whether they raise any obligations for your firm.

How Ariath Fits Into an Audit

Ariath's product connects your CRM, custodian, and planning tools to the approved AI models your firm already uses. It validates role-based access before data reaches a model, and it configures access, redaction, audit, and retention controls around each workflow. These are the kinds of controls an audit examines, and they are designed to give your team something concrete to review against its own policies. Ariath describes how it documents these controls on its AI compliance page.

Having these controls does not mean a workflow has been audited or meets any regulatory requirement. Your firm still needs to test its own configuration and decide what its obligations are. If you want to walk through a specific workflow and the evidence you would collect, Book a Demo.