Drafts arrive faster, but nobody can explain which approved system produced them.
Unmanaged AI use
The managing partner’s guide to shadow AI.
Shadow AI is work performed with an AI service the firm has not approved, configured, or monitored. Treat it as a workflow problem as well as a security problem.
Published by PrivateStride · Last updated August 11, 2026
Why it appears
Your people are solving a real capacity problem.
If the approved way is slow and the unapproved way is useful, policy alone will not hold the line.
Busy professionals use AI to draft emails, summarize files, research questions, and get through repetitive work. The risk is not limited to model training. It can include unknown retention, weak access control, unclear data location, unreviewed outputs, missing records, and terms the firm never accepted.
The goal is not a hunt for rule breakers. It is a reliable view of where work and client information are moving, followed by a governed option people will actually choose.
Review the security control modelWhat to look for
Signals are not proof. They are a reason to ask.
Use interviews, anonymous workflow questions, and existing security records. Follow your employment, privacy, and monitoring policies.
Staff forward work to personal email or move between a firm computer and a phone.
Browser histories, expense reports, or single sign-on records show unsanctioned AI services.
A team shares prompts in chat while the firm has no approved AI workflow.
Documents contain invented citations, inconsistent terminology, or phrasing nobody can trace.
People avoid the official tool because it is slower or less useful than a public alternative.
A 30-day response
Find, contain, replace, and measure.
Move quickly enough to reduce exposure without making promises before you understand the work.
01
Days 1–5 · Find
Ask about workflows, not confessions.
Survey where AI is helping, which files are involved, and why the current process is painful. Include partners and managers. Inventory approved and discovered services, accounts, browser extensions, meeting bots, document tools, and embedded AI in existing software.
02
Days 6–10 · Contain
Set an interim boundary people can understand.
Pause unreviewed use with restricted data. State what is allowed for public or synthetic information. Give staff one place to ask questions and a response deadline. Preserve relevant records under the firm’s established process.
03
Days 11–20 · Replace
Build one approved workflow around the real need.
Select a frequent, reviewable task. Define the data boundary, access roles, logging, retention, output review, and owner. Train champions on real examples with synthetic or approved information before expanding.
04
Days 21–30 · Measure
Watch adoption and risk separately.
Track approved-tool usage, policy questions, unapproved-service detections, output corrections, time returned, and exceptions. A high usage number is not proof of good control, and a low incident count is not proof that shadow AI disappeared.
Make approval usable
A strong control has an owner and a fast path.
Your acceptable-use policy should name approved tools and data classes, prohibited uses, review duties, and the person who can grant an exception. It should also explain how a professional proposes a new workflow without waiting through an entire busy season.
Use the policy starter and checklist- One accountable owner for the AI inventory.
- One approved intake path for new use cases.
- Data classifications staff can apply without a lawyer.
- Technical controls that support the written policy.
- Visible professional review before client use.
- Quarterly review after model, vendor, or workflow changes.
A defensible program
Document what you know and what you still need to test.
PrivateStride separates measured results from assumptions. Your governance records should do the same.
Primary sources: NIST’s voluntary AI Risk Management Framework organizes work around govern, map, measure, and manage. NIST’s Cybersecurity Framework provides a broader structure for cybersecurity risk. Review our claims methodology for the way PrivateStride labels baselines, observations, and projections.
Keep reading
Turn the policy into a working system.
These guides cover the decisions that sit next to this one.
ChatGPT and client data
Use five questions to review a specific service and workflow.
Read the guideSafeguards Rule and AI
Bring AI systems into the risk assessment and WISP process.
Read the guideCompare deployment options
Choose a data boundary your firm can govern and support.
Read the guideCapacity & AI risk assessment
Map the risk before you choose the tool.
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.
Book your audit 30 min · No preparation · Confidential