Client confidentiality
Sometimes a firm may approve a specific AI service for a specific workflow. That is not the same as approving staff to paste client data into any chatbot.
Published by PrivateStride · Last updated August 11, 2026
Important context
The answer depends on the facts, including the data, purpose, product, settings, contracts, professional duties, and applicable law. Have your legal, privacy, tax, and security advisers review the use your firm is actually considering.
The short answer
Begin with a written approval process for one defined service, one defined data class, and one defined workflow.
Turning off model training is useful, but it answers only one question. It does not by itself answer retention, access, disclosure, confidentiality, security, contractual, or professional-review questions.
Consumer ChatGPT accounts and business offerings do not use the same defaults. OpenAI says content from individual services may be used to improve models unless the user opts out. It says its business offerings and API do not use inputs and outputs for training by default. Your firm should still review the current terms and controls for the exact product before approval.
See how PrivateStride defines the data boundaryFive questions
A useful review is specific enough that a staff member can tell what is permitted without guessing.
01
A personal consumer account, a business workspace, an API, and a private deployment can have different terms, data controls, retention settings, and administrative features. The model name alone does not answer the risk question.
02
Classify the proposed inputs. Tax return information, personally identifiable information, financial records, credentials, workpapers, and de-identified examples do not carry the same risk. Include prompts, uploaded files, retrieved documents, outputs, logs, and support records in the data map.
03
Document where processing, storage, backups, monitoring, and support can occur. Ask who can access each copy, how long it remains, how deletion works, and whether any subprocessors are involved.
04
Summarizing a document for the same engagement is different from using client information for marketing, product improvement, benchmarking, or an unrelated service. Your advisers need the actual workflow, not the label “AI.”
05
AI can produce plausible but incorrect work. Define the professional who checks facts, calculations, citations, tax positions, and tone before an output reaches a file or a client.
A practical boundary
The examples below are a starting point for policy design, not a legal classification for your firm.
| Lane | Example | Practical rule | Required review |
|---|---|---|---|
| Open | Public research or generic writing with no firm or client facts | Only in an approved service and account | Normal professional review |
| Controlled | Internal templates, policies, and nonpublic firm material | Only in the environment approved for that data class | Owner, security, and records review |
| Restricted | Tax return information, client records, credentials, or regulated data | No use until the exact workflow is expressly approved | Legal, privacy, tax, security, and engagement review as applicable |
An approval path
A policy that only says “do not use AI” leaves the work pressure in place. Staff may create their own workaround. Give them a place to bring use cases, a short response time, and a useful approved alternative.
Start with the acceptable-use policySources and claims
AI services change. Recheck material settings, contracts, and data flows when the vendor or workflow changes.
Primary sources: OpenAI’s official explanation of how data may be used, the FTC’s Safeguards Rule guide, the IRS Section 7216 Information Center, and the NIST AI Risk Management Framework. Review the evidence standards behind PrivateStride’s own claims in our methodology.
Keep reading
These guides cover the decisions that sit next to this one.
Find unmanaged use and replace it with a governed path.
Read the guideFrame the use and disclosure questions around tax return information.
Read the guideSee how public, hosted, and on-premises options change the control model.
Read the guideCapacity & AI risk assessment
In 30 minutes, we identify your highest-value workflows, likely shadow-AI exposure, and the controls a private AI program would need. The findings are yours to keep.
