Skip to content
Dry Run Cyber
Free guideUpdated November 17, 2026About 35 minutes

Securing AI for your clients: the MSP field guide

How an IT provider secures AI for small and mid-size clients, start to finish: your own house first, then the policy, the admin consoles in Copilot, ChatGPT, Claude, Gemini and Grok, the data controls, shadow AI, agents and the people. Click paths included. Vendor-neutral.

Chapter 1

Why this is your problem now

Your clients already use AI. The question is whether it runs through accounts you manage, or through personal logins that nobody can see, audit or switch off.

Every MSP has had some version of the same ticket: “Can we turn on Copilot?” or “Is it OK if I use ChatGPT for this?” Behind it is a messier reality. Staff paste client emails, contracts, spreadsheets and patient notes into whichever assistant is open in the next tab, often on a free personal account tied to a work email address. That data leaves the tenant you're paid to protect, and you have no log of it.

Three things have changed in 2026 that make this an MSP problem rather than a client preference:

  • The AI is inside the tools you already manage. Copilot reads whatever a user can reach in Microsoft 365, and Gemini does the same in Workspace. Years of loose SharePoint and Drive sharing used to be invisible; now an assistant can surface it in one prompt.
  • Agents act, not just answer. Assistants now connect to mailboxes, drives, CRMs and third-party tools, and can take actions. Each connector is a new path for data and a new set of credentials.
  • Regulators and insurers have started asking. In the EU, the AI Act's staff AI-literacy duty has applied since February 2025, and supervision and enforcement started on August 2, 2026.1 There's no standalone fine, but missing training is treated as an aggravating factor when something else goes wrong.2 US healthcare clients face a simpler test: if patient data goes into an AI tool without a business associate agreement, that's a HIPAA problem regardless of how careful the user was.

Chapter 2

The six layers

No single product secures AI use. You stack six layers, and most SMB tenants today have one or two of them.

  1. 1PolicyWhat's allowed, with which data, in which tools, and who decides.
  2. 2Identity and accountsWork AI runs on managed work accounts with SSO, never on personal logins.
  3. 3App admin controlsEach platform's own console: retention, connectors, web access, sharing, audit.
  4. 4Data controlsFix oversharing before the assistant finds it; DLP on prompts and uploads.
  5. 5Visibility and enforcementSee and steer AI use in the browser and on the endpoint, including tools you didn't approve.
  6. 6PeopleTraining that changes what someone does when the prompt box is open.

The order matters. A DLP rule doesn't help if staff use a personal account the rule never sees. Admin controls don't help if nobody agreed what the business allows. Training doesn't stick if the approved tool is worse than the free one. Work top to bottom, and expect to go round more than once: the platforms change their consoles every month.

Chapter 3

Start with your own house

Before you advise a client, make sure your own technicians aren't the leak. An MSP holds the keys to dozens of tenants; that makes your AI use the highest-risk use in the room.

Think about what your team pastes into an assistant on a normal day:

  • Error logs and scripts containing tenant IDs, hostnames, internal IPs or API keys.
  • Ticket text with client staff names, email threads and screenshots.
  • Firewall configs, backup reports and audit exports “just to summarize”.
  • Proposals and QBR decks with client pricing and contract terms.

If that happens on personal accounts, every one of your clients has an exposure they never agreed to. Fix it at home first, then use the same playbook with clients. It also makes the sales conversation honest: “here's exactly what we did for ourselves.”

Least privilege for AI administration

Don't manage AI settings from a Global Administrator account. In Microsoft Entra there is an AI Administrator role for managing Copilot settings and approving agents, and Microsoft's own guidance is that AI Administrator, Agent ID Administrator and Security Administrator together cover AI management without Global Administrator.3 Oversharing work that changes what Copilot can ground on needs SharePoint Administrator.3

Your own house

0/5

Chapter 4

Write the AI usage policy

The policy is the contract between the business and its staff. Short, specific and enforceable beats long and aspirational.

A good SMB AI policy fits on three pages and answers seven questions:

QuestionWhat a good answer looks like
Which tools are approved?Named products and plans (for example “ChatGPT Business, company workspace only”), with an owner who approves new ones.
Which accounts?Company-managed accounts signed in with work SSO. Personal accounts are never used for company data.
Which data can go in?A short data-class table: public, internal, confidential, regulated, and which tool tier each may enter.
What's never allowed?Regulated data in unapproved tools, credentials and secrets, other people's personal data without a reason, and decisions made by AI alone about people.
Agents and connectors?Nobody connects an assistant to mail, files or business systems without IT approval and a named owner.
Human review?AI output is checked by a person before it reaches a customer, a regulator or a contract.
What if something goes wrong?Who to tell, how fast, and that reporting a mistake is never punished.

Make it real

  • Approve before you ban. Banning AI without giving people a sanctioned tool sends them to their phones. Pick one approved assistant per client and make it good.
  • Tie it to the data, not the brand. “Patient data only in tools with a signed BAA” survives the next product launch; “don't use ChatGPT” doesn't.
  • Get an acknowledgment. Each employee signs it, the date is recorded, and it's re-signed yearly. That record is your evidence.
  • Review quarterly. The platforms change faster than any other policy you maintain.

Chapter 5

Get admin access the right way

The client should own the AI workspace and the billing relationship. You get a named, least-privilege admin seat, written into your agreement.

AI platforms other than Microsoft have no partner-delegation model like GDAP. That changes how you ask for access. The principle is the same everywhere:

  • The client is the owner. The primary owner or billing owner is a client executive, on a client email domain. If you leave, the workspace stays with them.
  • You get a named admin account, on the client's domain or invited by email, never a shared login. MFA is enforced through the client's identity provider when the plan supports SSO.
  • It's in writing. Your MSA or SOW says which AI platforms you administer, what you're allowed to change, and that you'll hand back access on exit.
PlatformHow you get inAsk for
Microsoft 365GDAP relationship from Partner Center; the client approves it. Relationships run from 1 to 730 days.5AI Administrator for Copilot settings and agents; SharePoint Administrator for oversharing; Purview roles for DLP. Use Microsoft's least-privileged-role-by-task list.3,4
Google WorkspaceThe client's super admin creates an admin account for you, or grants a custom admin role.A custom role with the Gemini Settings administrator privilege, which is what the Gemini settings in the Admin console require.35
ChatGPT Business / EnterpriseThe workspace owner invites your account and assigns a role.Admin (not Owner). Enterprise adds a global admin console and SCIM.21
Claude Team / EnterpriseThe owner adds you in the admin console at claude.ai/admin-settings.28Admin, or a custom role scoped to Identity & Access on Enterprise.26
Grok Business / EnterpriseThe organization admin invites you in the xAI console at console.x.ai.39Team admin. Enterprise controls (SSO, SCIM, retention) go through xAI sales.

Chapter 6

Platform playbooks

What each plan actually gives an admin, what to switch on first, and where it hides. Pick the platform your client uses.

One rule applies to every vendor: the consumer and entry tiers aren't the business product. Data-handling terms, admin controls, audit logs and BAAs live on the business and enterprise plans. Your first job is getting the client onto the right tier.

SSOSCIMAudit / compliance APICustom retentionBAA available
Microsoft 365 CopilotEntra IDEntra IDPurview auditPurview retentionYes, under Microsoft's DPA20
ChatGPT BusinessYesNoNoNoNo20
ChatGPT EnterpriseYesYesCompliance APIYes (90-day minimum)21Sales-managed only20
Claude TeamYesNoNoNoNo
Claude EnterpriseYesYesAudit logs + Compliance APIYesYes, once enabled30
Gemini in WorkspaceWorkspace sign-inWorkspace directoryWorkspace audit3, 18 or 36 months, or off31Yes, after the BAA is accepted36
Grok BusinessEnterprise onlyEnterprise onlyEnterpriseEnterprise, via sales39Not stated for chat
Business-tier rows are the cheapest plan with real admin controls. Check current pricing and terms with each vendor.

Naming first, because it confuses users. In August 2026 Microsoft renamed the Microsoft 365 Copilot app to the Microsoft Copilot app.6 The license your clients buy is still Microsoft 365 Copilot (or Copilot Business for smaller tenants). What matters is the account people sign in with: an Entra work account gets the enterprise experience; a personal Microsoft account gets the consumer one.

Copilot answers using the data each user already has permission to see.8 So the biggest Copilot risk isn't Copilot; it's ten years of “Everyone except external users” on SharePoint sites. Fix that before you assign licenses.

Switch these on first

  1. Run the oversharing assessment. SharePoint Advanced Management comes with Microsoft 365 Copilot; one assigned Copilot license turns it on for admins.9 Start at SharePoint admin center›Advanced management with the Content Management Assessment, then read the data access governance reports.9
  2. Lock down what it finds. Use site access reviews so site owners confirm who should have access, and restrict or exclude high-risk sites before Copilot indexes them.8
  3. Decide on web search. The Allow web search in Copilot policy lives in the Cloud Policy service for Microsoft 365 and also appears on the Copilot Control System settings page in the admin center. It covers both Copilot and Copilot Chat.10 If you don't configure it, web search is on by default.10
  4. Add Purview DLP for Copilot. DLP can stop Copilot from processing sensitive items and prompts, and can now stop sensitive content in a prompt from being sent to external web search while still answering from internal data.11
  5. Turn on DSPM for AI in Purview to see AI interactions that touched sensitive data, including Copilot, Copilot Studio agents and third-party AI sites.12
  6. Govern agents in the Agent 365 registry in the Microsoft 365 admin center. Chapter 9 covers it.

Chapter 7

Data controls: DLP and oversharing

Admin consoles control the tool. Data controls follow the data: what an assistant can read, and what a person can paste into one.

Inside Microsoft 365

For Microsoft clients, Purview is the data layer, and the AI pieces stack like this:

  • Sensitivity labels and oversharing fixes decide what Copilot can ground on (chapter 6).
  • DLP for Microsoft 365 Copilot blocks sensitive prompts and keeps sensitive content out of web search.11
  • DSPM for AI gives one view of AI activity that touched sensitive data, across Copilot, enterprise AI apps such as ChatGPT Enterprise and Claude Enterprise, and other AI sites seen in the browser.12

Outside Microsoft 365: other AI sites

On Windows devices onboarded to Purview, Endpoint DLP can warn or block users who paste sensitive information into third-party AI sites in the browser. For example, it can stop someone pasting credit card numbers into ChatGPT, or show a warning they can override.12

That leaves real gaps for SMB clients: unmanaged and personal devices, Macs outside Purview, browsers without the extension, and AI tools that aren't websites at all. Those gaps are why the next chapter exists.

Data controls

0/5

Chapter 8

Shadow AI and the browser

Whatever you approve, people will try the next tool. Most AI use happens in a browser tab, so that's where you get visibility and control.

The Microsoft-native option

If the client has Defender for Endpoint and Defender for Cloud Apps, you can find and block generative AI apps without another agent:

  1. Go to Microsoft Defender portal›Cloud apps›Cloud app catalog and filter Category to Generative AI.15
  2. Tag the apps your policy doesn't allow as Unsanctioned. On devices onboarded to Defender for Endpoint they're blocked, or set to warn and educate instead.15
  3. Create a discovery policy that alerts when a new generative AI app shows up in the logs.

Purpose-built AI security for MSPs

Dedicated tools fill the gaps the native stack leaves: in-prompt redaction instead of a blunt block, coverage for non-Microsoft clients, and one console across all your tenants. Three worth knowing, for different jobs:

What it isWhere it worksBest fit
DefensXBrowser security with AI governance: blocks or restricts unapproved LLMs and filters sensitive data, such as card or social security numbers, before it's pasted.42 Added an AI Cloud Connector in July 2026 to isolate browser AI traffic.40The browserMSPs who want a channel-first, multi-client browser layer for SMBs, with a monitor-only start.
Prompt Security (SentinelOne)A GenAI security platform that SentinelOne acquired in 2025. A browser extension watches prompts, file uploads and responses, and SentinelOne says it governs 15,000+ AI tools.43,44 It also uses a lightweight endpoint agent.44Browser, endpoint, desktop AI appsMSPs already on SentinelOne; it plugs into SentinelOne's multi-tenant console.45
ZenitySecurity and governance for AI agents: inventory, posture management and runtime detection across Microsoft 365 Copilot, Copilot Studio, ChatGPT Enterprise and Agentforce.46 Enforces policy inline as agents run in Copilot Studio and Foundry.47Inside agent platformsClients building their own agents in Copilot Studio or similar. More upmarket than the other two.

Chapter 9

Agents, connectors and MCP

An assistant that can read your mail is a data risk. An agent that can send mail, move files and call other tools is an identity with permissions, and you should manage it like one.

The OWASP Top 10 for LLM Applications (2025) is the shared vocabulary here. For an MSP, four entries matter most: prompt injection (instructions hidden in a document or webpage take over the assistant), sensitive information disclosure, excessive agency (the agent can do more than its job needs), and system prompt leakage.48

Prompt injection is why connectors are dangerous. If an agent can read an inbound email and also send email, a hostile message can try to make it forward data. The fix isn't a smarter model; it's fewer permissions and a human approval step for anything that sends, deletes or pays.

In Microsoft 365: Agent 365

Agent 365 reached general availability on May 1, 2026, as the control plane for agents in the Microsoft 365 admin center.17 The registry inventories agents and lets admins block or delete them.19 A Tools view lists the MCP servers, plugins and connectors agents rely on, lets you block risky ones, and lets you approve or reject new tool requests. Intune policies can block unsanctioned local agents on devices.18,17

Agents and connectors

0/5

Chapter 10

Train the people

Every control above fails open if someone pastes the customer list into a personal chatbot on their phone. Training is the layer that travels with the person.

AI changes security awareness training in two directions at once:

  • Using AI safely: what data goes where, which account to use, why a confident answer can still be wrong, and what to report. This is also what the EU's AI-literacy rule asks for, and training must fit the systems each person actually uses.1
  • Recognizing AI-powered attacks: phishing with perfect grammar, cloned voices asking for wire transfers, deepfake video calls, and lures that ride real events within hours. When Microsoft 365 had an outage on November 12, 2026, phishing emails using it as a lure started arriving within hours.49

Slide decks and once-a-year videos don't change behavior under pressure, and pressure is exactly what these attacks create. People need to practice the decision: the urgent request, the convincing voice, the “just this once” paste, in a setting where getting it wrong costs nothing.

From the people who wrote this guide

Dry Run Cyber is security awareness training people play.

One first-person night shift with real-world attacks, built from lures circulating now, with every decision recorded against NIST, CIS, HIPAA and CMMC controls. It's built for MSPs to resell, with a partner portal across all your clients.

Chapter 11

A 30/60/90 plan for each client

Use this as the statement of work. Each phase ends with something the client can see.

Days 1–30: see it and agree it

0/4

Days 31–60: configure it

0/5

Days 61–90: enforce and train

0/4

Sources

Checked November 17, 2026. Admin consoles change monthly; when a click path has moved, the source page is the tiebreaker. We aren't affiliated with any vendor named here.

  1. 1.The EU AI Act's AI literacy requirement: key considerations · Travers Smith
  2. 2.The EU's quiet training mandate now faces enforcement · Mitratech
  3. 3.Configure privileged roles to manage AI · Microsoft Zero Trust Assessment
  4. 4.GDAP role guidance: least-privileged roles by task · Microsoft Learn
  5. 5.What is GDAP and what does it mean for you · CloudCockpit
  6. 6.Microsoft adds to its branding confusion: Microsoft 365 Copilot app renamed Microsoft Copilot · Windows Latest
  7. 7.Microsoft 365 Copilot security: top questions answered · A Guide to Cloud
  8. 8.Configure data security for Microsoft 365 Copilot · Microsoft Learn
  9. 9.SharePoint oversharing remediation before Copilot · EPC Group
  10. 10.Data, privacy, and security for web search in Microsoft Copilot and Copilot Chat · Microsoft Learn
  11. 11.Microsoft Purview: securing data and enabling apps and agents · Microsoft Tech Community
  12. 12.Microsoft Purview protections for generative AI apps · Microsoft Learn
  13. 13.Purview capabilities supported for third-party AI sites · Microsoft Learn
  14. 14.DLP policy: block sensitive info from AI sites (MVP answer) · Microsoft Tech Community
  15. 15.Discover, monitor and block generative AI apps · Microsoft Learn
  16. 16.Microsoft Defender for Cloud Apps in 2026: what it finds, what it can't stop · dope.security
  17. 17.What's new in Agent 365: May 2026 · Microsoft Tech Community
  18. 18.Microsoft Agent 365 adds centralized controls · Redmond
  19. 19.Best Microsoft 365 governance tools for Copilot and AI readiness (2026) · Orchestry
  20. 20.Is ChatGPT HIPAA compliant? Which plans come with a BAA · BD Emerson
  21. 21.ChatGPT Business vs Enterprise: what IT actually gets for connectors and audit · Composio
  22. 22.ChatGPT Enterprise admin quickstart · OpenAI Help Center
  23. 23.Getting started with identity and provisioning in ChatGPT Enterprise · OpenAI Help Center
  24. 24.Verifying your domain for OpenAI identity · OpenAI Help Center
  25. 25.Onboarding employees with existing ChatGPT accounts · OpenAI Help Center
  26. 26.Claude Enterprise security: what admins control · Composio
  27. 27.Claude Team vs Enterprise plan comparison · Claude Training
  28. 28.How to roll out Claude to your whole team · Beginners in AI
  29. 29.Claude Enterprise · Anthropic
  30. 30.Which AI vendors sign a BAA for therapy practices · Sagnik Bhattacharya
  31. 31.Turn access to gemini.google.com on or off · Google Workspace Admin Help
  32. 32.Manage access to Gemini features in Workspace services · Google Workspace
  33. 33.Control Workspace Intelligence for generative AI features · Google Workspace
  34. 34.Enterprise security controls for Google Workspace with Gemini · Google Workspace blog
  35. 35.Allow or block Gemini for Google Workspace steps · Google Workspace
  36. 36.HIPAA-compliant AI for a medical office: the BAA check · Sagnik Bhattacharya
  37. 37.Grok Business · xAI
  38. 38.xAI launches Grok Business and Enterprise plans · TestingCatalog
  39. 39.Grok hardening guide · How to Harden
  40. 40.DefensX unveils AI browser security features for MSPs and SMBs · Channel Insider
  41. 41.DefensX Shadow AI protection · DefensX
  42. 42.Interview with DefensX at ChannelCon 2026 · DefensX
  43. 43.Prompt Security review · AppSec Santa
  44. 44.A new chapter for AI and cybersecurity: SentinelOne acquires Prompt Security · SentinelOne
  45. 45.SentinelOne strengthens AI tool visibility for MSSPs with Prompt Security buy · MSSP Alert
  46. 46.Agentic SaaS security · Zenity
  47. 47.Zenity brings inline prevention to Microsoft Foundry and Copilot Studio · Zenity
  48. 48.OWASP Top 10 for LLM Applications 2025 · AI Governance Library
  49. 49.Attackers exploit the Microsoft 365 outage with fake held-messages lures · Mimecast