A request is reaching more IT teams at small and mid-size firms. A managing partner, owner, or COO has seen an AI agent that can sort email, draft replies, manage a calendar, and pull documents on request, and they want it connected to their account. IT administrators have been discussing how to handle it in professional forums, and it tends to come from the people with the most demanding schedules and the most sensitive inboxes.
It’s a reasonable thing to want. These tools can save a senior person real time. The question for the firm is what has to be decided before the agent gets access, so the time savings don’t come with an unmanaged path into client data.
What Makes an Agent Different
A chatbot answers questions about whatever someone types into it. An agent signs in and takes actions. It reads the mailbox, opens files, and in many tools can send, move, or share on the user’s behalf, with whatever permissions it was given when it was connected.
That matters because of whose account is usually involved. An executive’s mailbox tends to hold the firm’s most sensitive material: client financials, matter files, HR issues, banking and wire communications. Depending on what’s granted, an agent connected to that account can reach much of it.
Microsoft 365 provides several ways for applications and AI agents to connect with business systems, and what’s possible depends on how the organization is configured. Some applications can request access directly from the user, while others require administrator approval, according to the tenant’s consent policies. Separately, beginning November 4, 2026, Microsoft started rolling out a change that allows default users to create agent blueprints in Microsoft Entra unless administrators restrict that capability through the “Users can create agent blueprints” setting (Message Center MC1506733). Creating an agent blueprint doesn’t by itself grant access to company email or documents. Both settings are worth reviewing before broader AI integrations are switched on, because together they determine how much of this can happen without IT involved.
Five Decisions to Make Before Connecting an AI Agent
1. Which tool, and on what terms
Evaluate the tool by its contract and controls, not its features. A business or enterprise plan should state how long data is retained, whether it’s used to train models, and where it’s processed. Whatever the plan is called, firm data should only go into a tool whose contract covers confidentiality, privacy, retention, and security in a way the firm has reviewed and approved. Consumer and free tiers rarely offer those terms, which is why they usually fail this test. Microsoft 365 Copilot, for comparison, documents that it only surfaces data a user already has permission to view and that prompts and responses aren’t used to train its foundation models. Every agent the firm considers should be able to answer the same questions in writing.
2. Whose identity it uses
How much an agent can reach depends on the integration model and the permissions granted. In some setups, the agent acts on the executive’s behalf and can see much of what that account can see. Where the tool supports it, the agent should have its own scoped identity, and consent for new apps and agents should route through an administrator. In Microsoft Entra, that means reviewing the “Users can create agent blueprints” setting and the user consent settings for applications, and turning on the admin consent workflow so a request becomes a review instead of a click.
3. What it can reach, and what it never touches
Start narrow. Where the tool supports granular permissions, read-only access to a specific mailbox or folder is enough to prove whether it’s useful. Some material should stay out of scope from the beginning: client trust and banking communications, privileged matter files, patient health information, and HR records. If the tool can’t be limited that precisely, that’s important information about the tool.
4. What it can do without a person
Drafting is a sensible default. Sending email, deleting or forwarding messages, sharing files outside the firm, and anything involving payments should require a human to approve each action. An agent that drafts a reply saves most of the time. An agent that sends one on its own is a different level of trust.
5. Who approves and who reviews
Name one person who approves new agent connections. Make sure activity is logged somewhere the firm can see it, schedule a review at 30 days, and know exactly how to shut the agent off. Microsoft Entra records agent identity creation in its audit logs, which gives a firm on Microsoft 365 a starting point.
The Professional Obligations Layer
For many Greater Sacramento firms, these decisions sit on top of existing professional duties. CPA firms should account for the IRS Office of Professional Responsibility’s June 2026 AI guidance, which points practitioners toward safeguards for client data, vetting of AI vendors, written firm procedures, and human review of AI-assisted work.
Law firms carry confidentiality and supervision duties, and California’s SB 574 adds a statutory rule on what confidential information can go into AI systems starting January 1, 2027, covered in our SB 574 post.
Healthcare practices need a business associate agreement and an updated HIPAA risk analysis before an agent touches patient information; our post on patient data in AI tools explains why.
These are summaries, not legal or professional advice. Confirm your obligations with your own counsel or professional advisors.
A 30-Day Pilot
The practical path is a controlled yes. Start with one executive, one approved tool, and one scoped mailbox or folder. Limit the agent to reading and drafting. Review its activity log weekly with whoever handles IT, and at the end of 30 days decide whether to expand access, change tools, or stop. A pilot like this gives the executive the time savings they asked for and gives the firm a record of what the agent actually did, which is the information needed to approve the next one.
It also gives the IT lead something better to offer than “no.” Our guide to what employees can and can’t share with AI covers the broader data question, and our look at Black October covers how quickly attackers are moving when access is left loose.
How to Verify Your AI Controls
A policy only helps if the firm can see whether it’s being followed. Three separate questions are worth answering. The first is identity and permissions: which applications or agents have been authorized to reach business systems, and with what scope. The second is visibility: which employees, devices, and AI platforms are in use, and what categories of sensitive information are passing through them. The third is protection: which of those interactions the firm can restrict or block under its own rules.
How well each can be answered depends on the tools involved and how they connect. An employee using an AI tool in a browser and an agent reaching Microsoft 365 through an API on a server leave different traces, and they usually need different controls. A firm that has only reviewed one of those paths has a partial picture.
Someone at your firm asking to connect an AI agent?
Vision Quest Cyber helps firms answer all three questions: reviewing which agents and apps have access to Microsoft 365, monitoring AI usage across employees and devices for sensitive data, and applying policy-based controls on supported AI platforms. Learn more about our AI governance services.
Talk With Us About AI GovernanceSources
- Microsoft 365 Message Center, MC1506733, “Review who can create agents in Microsoft Entra,” Oct. 13, 2026 (archived): mc.merill.net
- Microsoft Learn, “Configure how users consent to applications”: learn.microsoft.com
- Microsoft Learn, “How are agent identities created?”: learn.microsoft.com
- Microsoft Learn, “Data, Privacy, and Security for Microsoft Copilot”: learn.microsoft.com


