Your Firm Needs an AI Use Policy, and It Can Be Two Pages
Updated: Nov 1
November 2026 | TAAUS Secure Technologies

Someone at your firm used an AI tool this week. Maybe they summarized a contract or a fund document, drafted a client email, or cleaned up a spreadsheet. The question isn't whether AI is in your firm. The question is whether your firm has said anything about how it should be used.
For many smaller regulated firms, the answer is still no. Usually that isn't neglect. "AI policy" sounds like a six-month governance project with a steering committee, and for some firms it should be. But if AI use is limited to standalone tools, without integrations into firm systems, a clear, enforceable policy fits on two pages. You can have one in place this quarter.
Why regulators, clients and allocators now expect one
AI governance has moved from a nice-to-have to an exam topic. The SEC's fiscal year 2027 examination priorities, released October 23, 2026, call out emerging and advanced technologies alongside cybersecurity and the protection of customer information (SEC). FINRA's 2026 oversight report gave generative AI its own section and is explicit that existing rules on supervision, communications and recordkeeping still apply (Debevoise).
Other regulated sectors are on the same path. For law firms, ABA Formal Opinion 512 (July 2024) ties generative AI use to existing duties of competence, confidentiality, client communication and supervision (Katten). For healthcare, HIPAA has no AI-specific rules, but an AI vendor that receives PHI is a business associate and needs a signed BAA first (Holland & Hart).
There is no single AI compliance rulebook that applies to every regulated firm. Instead, regulators evaluate AI use against existing obligations for supervision, confidentiality, recordkeeping, security and client protection. Four risks come up again and again:
Data leakage. Client PII, PHI, positions or privileged material pasted into a consumer chatbot leaves your control, and may be used to train the vendor's model.
Recordkeeping. If AI drafts client communications or meeting notes, those outputs may be books and records, part of a client file or part of a medical record.
Accuracy. AI output that reaches a client, investor or patient unreviewed is your firm's statement, errors included.
Disclosure. Telling clients or allocators you "use AI" invites the question of how it's governed.
The same questions now appear in client security questionnaires, outside counsel guidelines and operational due diligence questionnaires. "Do you have a policy governing the use of generative AI?" is a yes-or-no box. Firms that check "no" should expect a follow-up call.
When two pages is enough
Policy depth should match how deeply AI is woven into your environment. Most small firms are at the light end: staff use Copilot or ChatGPT in a browser to draft, summarize and research, and nothing connects to firm systems. At that level, an enterprise framework built for a risk committee and a model inventory is overkill, and a 25-person fund, law firm or medical practice won't maintain it.
For light use, a short policy wins on the things that matter:
People actually read it. Staff can absorb two pages at onboarding and remember the rules.
You can enforce it. Every rule maps to a control you already own, so nothing is aspirational.
It fits your governance set. It sits beside your WISP and Acceptable Use Policy, built on the same structure, so examiners see one coherent program.
It can grow. Version 1.0 covers today's tools; you revise it when you adopt AI in investment or client-facing workflows.
The goal of version one isn't to anticipate every use case. It's to make sure every employee knows which tools are allowed, what data never goes in, and who to ask.
When you need a full AI governance policy
The two-page approach stops working once AI is integrated into your environment. Signs you've crossed that line:
AI tools are grounded in firm data, such as Copilot with access to SharePoint, mailboxes and Teams
AI agents or automations take actions in your systems
Staff or vendors build applications on AI APIs
AI informs investment, legal or clinical decisions, or produces client-facing output at scale
At that point you need a full governance policy. It covers an inventory of AI systems, risk assessment before deployment, data classification and access controls for what AI can reach, vendor due diligence, testing and monitoring, and named accountability. It's a longer document because there's more to govern.
What goes on the two pages
Six sections cover what a small regulated firm needs. Each can be a short paragraph or a few bullets.
Section | What it says | Example language |
Scope and owner | Who it applies to and who owns it | Applies to all employees, contractors and interns; owned by the COO or managing partner |
Approved tools | The short list of AI tools staff may use, on firm-managed accounts only | Microsoft 365 Copilot Chat and ChatGPT Business via firm-issued logins |
Data handling | What data may go into which tools, and what never goes into any | Credentials and secrets never; client data, PHI and privileged material only in tools approved for that data |
Acceptable use | What AI is allowed to help with | Drafting, summarizing public material, internal research, formatting |
Human review | Who checks AI output before it leaves the firm | No AI-generated content goes to a client, investor, patient or court without review |
Requests and violations | How to request a new tool and what happens if rules are broken | New tools need COO and IT approval before use |
The data-handling section does the most work, so be specific. For example: "Confidential client information, PII, PHI, privileged material, financial account data and proprietary business information must not be entered into AI tools unless the specific tool and use case have been approved for that classification of data. Credentials, authentication tokens and secrets are prohibited in all AI tools." A vague phrase like "sensitive information" means nothing to a busy analyst.
Human review is one of the most important safeguards. It's what keeps a hallucinated figure out of an investor letter, or a fabricated citation out of a court filing.
Picking approved tools: the account matters more than the brand
The most important distinction isn't Copilot versus ChatGPT. It's a firm-managed business account versus a personal consumer account. The same chatbot can be acceptable or a data-leak risk depending on how someone signs in.
Business and enterprise tiers typically give you four things consumer accounts don't:
Contractual terms that exclude your data from model training by default
Sign-in through your identity provider, with MFA and conditional access
Admin visibility into who is using the tool
The ability to cut off access the day someone leaves
These capabilities vary by vendor, license and configuration. ChatGPT Business, for example, supports SSO but not automated SCIM provisioning, so offboarding stays a manual step (Stitchflow). A business subscription alone doesn't guarantee that SSO, logging, retention or admin restrictions are configured correctly.
Start with what you already pay for. Most of our clients run Microsoft 365, so Copilot Chat under their existing licensing is the natural first approved tool. If staff are heavy ChatGPT users, approving ChatGPT Business is better than pretending they'll stop. One recent client approved both for exactly that reason.
Healthcare organizations have one extra test. If a tool will ever touch PHI, the vendor must sign a BAA covering it, and many self-serve business tiers won't. Until one is signed, PHI goes on the prohibited list for that tool.
Whatever you approve, read the vendor's data-handling terms before you sign. Confirm training, retention and where data is stored in writing.
Shadow tools are already in your environment
If you think staff will wait for permission, consider what we found at one client this year. Our EDR flagged a persistence mechanism on an employee workstation. It wasn't malware. A staff member had built a small background application that connected a cloud-hosted internal tool to a licensed market-data terminal on their desktop.
It was clever, it worked, and nobody in IT or compliance knew it existed. It ran under one person's login, and had no review, no documentation and no owner if that person left.
This wasn't necessarily an AI incident. It was a shadow IT problem. But AI-assisted development makes this kind of activity much easier: employees who once needed a developer to build an integration can now create one themselves, often without understanding the security implications.
A policy doesn't stop that creativity. It gives people a path: tell us what you want to build, and we'll help you build it safely.
It also helps to have monitoring that catches what the policy misses. Without endpoint monitoring, the application might have kept running without anyone in IT or compliance knowing it existed.
Getting started this quarter
A policy that says employees can't use unapproved AI tools is useful. A policy backed by identity, endpoint and web controls is enforceable.
You can have version 1.0 in place in a few weeks:
Ask staff which AI tools they already use. Make it amnesty, not an audit.
Pick one or two approved tools, starting with what your existing licensing covers.
Write the prohibited-data list in plain, specific terms.
Draft the two pages on the same template as your WISP and AUP.
Have every employee acknowledge it, and add it to onboarding.
Restrict unapproved AI services with DNS security, web filtering and application controls. DNS alone can't tell a personal account from a business one, so for approved services use identity policies, tenant restrictions and data loss prevention to control access and data sharing.
Review it annually and whenever a new AI tool or material use case is introduced. Move to a full governance policy when AI connects to your firm's data or systems.
TAAUS Secure Technologies builds both: lightweight AI use policies for firms with light AI use, and full AI governance policies for organizations with deep AI integration. Either one fits into a complete governance set alongside your WISP, Acceptable Use Policy and Incident Response Plan. If your next DDQ, client questionnaire or audit asks about AI and you don't have an answer yet, contact us.



