Managed servicesGovernance · DLP · ZDR
Use AI at work without losing client data, IP, or audit trails
Staff are already trying ChatGPT, Claude, and Gemini — shadow AI and pasted matter files create leakage paths long before leadership approves a rollout; governance, DLP, and enterprise AI tenants put guardrails where work actually happens
AI is changing how work gets done — and employees are already using consumer chat tools to draft emails, summarize documents, and speed up research. Without governance, that convenience becomes shadow AI: client PII, tax data, and firm IP pasted into accounts the business does not control, often with default settings that may train on prompts. Manage IT NY maps where AI already sits, blocks what should not leave the tenant, and deploys approved enterprise paths — Microsoft Copilot, Azure OpenAI, or governed alternatives — with DLP, single sign-on (SSO), and logging partners can ask about.
Technology partners
Approved AI vs shadow AI — why blocking everything fails and free ChatGPT at work is risky
Blocking every AI site sounds safe until a partner pastes a brief into a personal account anyway — browsers, phones, and browser extensions do not respect a memo. Shadow AI is any generative AI tool staff use without IT or legal review: free ChatGPT, Claude, Gemini, or plugins that read mail and files. The risk is not the model itself — it is uncontrolled data leaving your Microsoft 365 or Google tenant, with no acceptable use policy (AUP), no data loss prevention (DLP), and no audit trail when a client asks what happened to their matter file.
The workable path is not prohibition — it is approved enterprise tenants with zero data retention (ZDR) where available, AI-aware DLP that redacts or blocks sensitive pastes, granular identity and access management (IAM) for who may use which tool, and training so staff know the difference between a governed copilot and a consumer tab. Manage IT NY helps firms get there without pretending staff will stop experimenting.
How governed AI should flow — inputs, guardrails, then an approved tenant
Think of three layers. First, what staff try to send: prompts, uploaded PDFs, mail snippets, and exports from line-of-business systems. Second, enterprise guardrails — DLP policies, web and SaaS controls, and AUP enforcement that catch client numbers, matter IDs, and export-controlled text before they leave. Third, a secure isolated AI tenant: enterprise Copilot, Azure OpenAI, or a vendor contract with ZDR (zero data retention — prompts not used to train public models), SSO so accounts are firm-owned, and audit logs for admin and high-risk actions.
Corporate inputs → guardrails → approved AI tenant
Staff prompts & file uploads
│
▼
Enterprise guardrails (DLP · web filter · AUP)
│
▼
Approved AI tenant (ZDR · no training · SSO · audit log)Read each box with leadership — the middle layer is where most firms are missing controls today, not the existence of AI itself.
Consumer AI tabs vs enterprise governed AI
Four dimensions partners and compliance officers actually ask about
Flip any row for plain-English detail. The left column is what we find when staff solve problems in personal browser sessions; the right is what a governed program targets — with evidence, not a blanket ban.
Convenience without ownership
Tools staff can use with boundaries
Timelines depend on tenant size, how many shadow tools are in use, and whether Microsoft 365 Copilot or Azure OpenAI is already licensed. Most firms phase discovery and DLP first, then enterprise deployment and training — without stopping filing season to rewrite every workflow.
Four controls that stop the usual AI failure modes
Shadow blocking, ZDR enclaves, AI-aware DLP, and granular IAM — as one program
Manage IT NY ties each control to a failure mode business leaders recognize — not a feature list for a single vendor SKU.
Shadow AI blocking
Stops this failure mode: staff paste client work into personal ChatGPT because IT never offered an alternative
Discover unapproved AI domains, browser extensions, and OAuth connectors — then block or coach at the web filter, endpoint, and identity layers while approved enterprise paths remain usable.
ZDR enclaves
Stops this failure mode: prompts retained and used to train public models
Zero data retention (ZDR) and contractual no-training terms in enterprise Copilot, Azure OpenAI, or governed vendor tenants — isolated from consumer accounts with different legal terms.
AI-aware DLP
Stops this failure mode: SSNs, matter numbers, and CUI text copied into any AI upload box
Data loss prevention (DLP) policies tuned for generative AI endpoints — redact, warn, or block based on sensitivity labels, regex patterns, and destination (browser paste, cloud API, or copilot scope).
Granular IAM
Stops this failure mode: every user inherits full copilot scope across all shares
Single sign-on (SSO), group-based access, and least privilege for who may enable plugins, connect line-of-business systems, or use high-risk AI features — aligned with Zero Trust identity hygiene.
Shadow AI
Tools IT never approved
Any generative AI use outside your governed tenant — personal ChatGPT, Claude, Gemini, or extensions that summarize mail. Define once for staff: if it is not on the approved list, it is shadow AI.
ZDR
Zero data retention
Contractual promise that prompts and outputs are not kept for model training — common in enterprise AI agreements. ZDR is not automatic on free consumer tiers; verify in writing.
DLP
Redaction before paste
Data loss prevention watches copy/paste, uploads, and API calls. AI-aware rules catch tax IDs, health data, and export-controlled strings — block or mask before they reach an external model.
Approved vs consumer
Enterprise tenant vs free tab
Approved tools run under firm SSO, logging, and DLP. Consumer tabs use personal terms, personal retention, and no offboarding hook — fine for lunch plans, not for client work.
What good looks like
A short buyer checklist before you trust the AI program — not a hype slide, a readiness scan you can walk through with partners and risk committees.
AUP published?
Staff have a one-page acceptable use policy — what they may paste, which tools are approved, and who to call when unsure — not a 40-page PDF nobody opened.
Enterprise tenant live?
Governed Copilot, Azure OpenAI, or contracted alternative with SSO — not only consumer accounts reimbursed on expense reports.
DLP on AI paths?
Policies test against real matter files and tax exports — with alerts routed to someone who reviews, not an empty dashboard.
Shadow AI inventory?
You can name unapproved tools discovered in the last 90 days and what changed — block, allow, or replace — since the last audit.
Training on a calendar?
Onboarding and refreshers cover approved vs shadow AI, with dates suitable for a partner agenda or compliance file.
Four-step AI security roadmap
Shadow AI audit, AUP, secure enterprise deployment, and continuous training — phased so experimentation becomes a program leadership can explain. Manage IT NY documents rollout realism: discovery and DLP often lead; full copilot scope and plugin governance follow once the inventory is honest.
Step 1
Shadow AI audit
Inventory browser use, tenant copilot settings, OAuth connectors, and extensions — approved, shadow, and unknown. You get a written map of where data could leave before anyone signs an enterprise order form.
Book AI security assessmentStep 2
Acceptable use policy
Publish plain-language rules: what may be pasted, which tools are allowed, how to request an exception, and who approves plugins or API keys. Legal and IT align on one page staff can actually read.
See what good looks likeStep 3
Secure enterprise deployment
Roll out governed Copilot, Azure OpenAI, or selected vendors with ZDR terms, SSO, DLP, and shadow-AI blocking for consumer paths. Phase by department so filing, audit, or plant teams are not disrupted on day one.
Talk deployment phasingStep 4
Training and continuous auditing
Onboard staff on approved vs shadow AI, run periodic re-scans for new tools, review DLP incidents, and update the register when models or plugins change — AI governance is a calendar, not a one-time project.
Schedule follow-up auditHow AI governance maps to your industry
ABA competence and confidentiality, IRS and FTC safeguard programs, and CMMC controlled unclassified information (CUI) each ask what happens when client or export-controlled data enters a model. Here is how Manage IT NY translates AI security into language each vertical already uses — without promising that AI removes professional judgment.
Law firms — ABA Model Rules 1.1 and 1.6
Competence (Rule 1.1) expects lawyers to understand technology they use — including generative AI risks to privilege and confidentiality (Rule 1.6). Governance means knowing which tools staff use, whether prompts leave the firm, and what you tell clients when AI assists drafting or research. Shadow AI in personal browser tabs is the hardest story to tell a malpractice carrier.
Law firm cybersecurityAccounting — IRS Pub 4557 and FTC Safeguards
Taxpayer PII in prompts or uploaded PDFs triggers the same safeguard expectations as any other exit path from your environment. IRS Publication 4557 and the FTC Safeguards Rule imply access controls, monitoring, and vendor oversight — including AI vendors and copilots scoped to client files. DLP and enterprise tenants support the questions state boards and the IRS ask about WISP programs.
Accounting firm cybersecurityDefense & CMMC — CUI and export control
CUI and export-controlled technical data cannot casually enter public models. CMMC and NIST SP 800-171 expect media protection and access enforcement — AI paths included. Enclave-style deployment, blocked shadow AI, ZDR contracts, and audit logs support assessor questions about how generative tools interact with defense work — not a generic commercial ChatGPT tab.
CMMC enclave pathProfessional services and internal IP
Manufacturing, agencies, and multi-site ops firms worry about formulas, bids, and HR data in prompts — not only client files. IAM scoping, DLP for internal classification, and approved copilots reduce the chance that a helpful summary becomes an accidental leak to a vendor training set.
Operations cybersecurityFrequently asked questions
Straight answers on shadow AI, safe use, DLP, ZDR, blocking strategies, and enterprise Copilot — the questions partners and risk committees ask before approving wider rollout.
Shadow AI is any generative AI tool employees use without IT, legal, or compliance review — personal ChatGPT, Claude, Gemini, browser extensions, or mobile apps connected to work mail. The firm does not control accounts, retention, or training use. Discovery is the first step: most leaders underestimate how often staff already paste client or financial data into consumer tabs.
Yes — with boundaries, not blind trust. Safe use means approved enterprise tenants (for example Microsoft Copilot or Azure OpenAI under contract), DLP that catches sensitive pastes, SSO and offboarding on firm accounts, published AUP, and training. It does not mean every task belongs in a model; partners still decide what requires human judgment and client disclosure.
Combine technical and human controls. AI-aware DLP and browser policies block or warn on sensitive patterns; web filters can limit unapproved AI domains while enterprise paths stay open; sensitivity labels in Microsoft 365 or Google Drive reduce what copilots can read. Training explains why paste is risky. No single toggle replaces all three — consumer ChatGPT will always tempt a shortcut until a governed alternative is easier.
Zero data retention (ZDR) is a contractual commitment that your prompts and outputs are not stored for vendor model training — typical in enterprise AI agreements, not guaranteed on free consumer tiers. ZDR pairs with SSO and audit logging in governed tenants. Verify terms in writing; marketing language about “privacy” is not the same as ZDR in a BAA or enterprise order form.
Outright bans rarely stick — staff find workarounds on personal phones and home networks. A better default: block or warn on high-risk consumer AI destinations, deploy approved enterprise tools with DLP, and publish clear rules. Blocking without an alternative drives shadow AI underground, which is harder to audit than governed use.
Copilot and Azure OpenAI are common enterprise paths — they integrate with Microsoft 365 identity, logging, and DLP when configured deliberately. They are not automatic governance: scope (which files the copilot may read), plugin approvals, and ZDR or data-processing terms still require design. ${site.name} helps firms deploy them as part of a program — shadow audit, AUP, IAM, and continuous review — not as a checkbox purchase.