Company Identity for AI Agents: Email, Accounts and Audit Badges
An AI agent needs its own company email, account and audit identity so actions stay attributable without sharing a person's login. Learn how to set it up.

An AI agent gets its own company identity when it receives a dedicated email address, a named service account in each tool it uses, and scoped credentials that belong to the agent rather than to a person. With that identity in place, every read, draft and write appears in the logs under the agent's name, so an operations leader can answer who did what without guessing or reusing someone's password. The rest of this article explains how to set it up and which decisions come first.
Key takeaways
- Give each agent its own named identity: a company mailbox, accounts in the tools it uses, and a role record that a person owns.
- Never run an agent under an employee's password, session cookie or personal token.
- Attribution needs three records: which identity acted, which grant allowed the action, and which approval covered any write.
- Begin with read access under the agent's identity, then add write permissions one tool at a time.
- Review, rotate and retire agent identities on a fixed schedule, the same way you handle staff access.
Why sharing a person's login breaks accountability
When an agent sends an email, updates a record or downloads a report using an employee's account, the system records that employee as the actor. The audit trail then says a person did something the person never reviewed. Three problems follow. First, nobody can separate human actions from agent actions when investigating an error or a customer complaint. Second, passwords, multi-factor prompts and session limits were designed for a person at a keyboard, so the automation breaks whenever those controls change, and teams often respond by weakening the controls. Third, when the employee changes roles or leaves, the agent either loses access without warning or keeps access that nobody remembers granting.
Shared identities across several agents have the same flaw in a larger form. If one account serves three agents, the log cannot tell them apart, and a misbehaving prompt in one role can reach permissions that were meant for another.
The fix is simple in principle. Every agent that touches a business system should act as itself: its own email address where it needs one, its own account or service identity in each system, and credentials that can be revoked without affecting any person.
What an agent identity includes
A usable agent identity is more than a username. It is a small set of records that together answer who the agent is, who owns it and what it may do.
- A company email address on your domain, named for the function, such as a billing assistant address rather than a human name.
- An account or service identity in each system the agent uses, kept separate from staff accounts.
- A role record with the agent's purpose, its owner, the people who approve its writes and the date of its last review.
- Scoped credentials for each tool, stored in your secrets manager and never placed in prompts, notes or shared documents.
- An entry in your access register, so the identity appears beside staff and vendor accounts when someone audits access.
Naming matters more than it seems. Names such as billing-assistant or support-triage make logs readable for people who were not present when the agent was configured. Avoid names that imitate colleagues, because a log entry that looks like a staff member invites the wrong assumptions during an investigation.
How to give an agent its own email and accounts
The sequence below is designed so that each step produces a record the next step depends on.
- Step 1. Name the role and its accountable owner before creating anything. The owner answers for the agent's behavior and approves changes to its scope.
- Step 2. Create a dedicated mailbox on your company domain. Do not attach it to a person's profile, signature or phone number.
- Step 3. Create a separate service identity or app registration in each system the role touches, following the vendor's guidance for non-human accounts.
- Step 4. Grant the smallest role in each system, starting with read access. Add write permissions only after the read path has been tested.
- Step 5. Store every credential in your secrets manager and record where it lives, without recording its value in the register.
- Step 6. Add the identity, owner, scopes and review date to your access register.
- Step 7. Run a test read, then confirm the audit log shows the agent's identity and not a person's name. Do not enable writes until this check passes.
Illustrative scenario: a small distribution company wants an agent to answer supplier questions about delivery dates. The team first creates an agent mailbox and a read-only identity in the order system. Only after a review period of reading reply quality do they consider a draft permission, and any reply that quotes a date goes to a named approver before it is sent. The sequence required no new tool, only clear identities and a clear order of permissions.
Service accounts or delegated grants: which to use
Two patterns cover most cases. A service account is an identity the organization owns outright, suited to actions the company performs in its own name, such as posting to an internal channel or reading a shared folder. A delegated grant is an authorization in which a user consents to an application acting on that user's data within defined scopes. The OAuth 2.0 framework used by Google and many other providers describes this model, with tokens limited to the scopes the user approved.
Delegated grants carry a trap. If an employee connects an agent to their own mailbox, the agent inherits that person's reach, and the grant persists after the employee changes roles. When the agent acts for a team rather than one person, prefer a service identity. When a delegated grant is unavoidable, issue it to the agent's own identity, record who consented, limit the scopes to the minimum and place the grant on the access review calendar. The trade-offs are covered in Service Accounts or OAuth Tokens for AI Agents: Which to Use.
How to make every action attributable
Identity alone does not create accountability. The audit record must also capture enough context to reconstruct a decision. For each action, the record should contain the agent identity, its role, the tool and operation, a reference to the input, the timestamp, the result, and any approval that authorized a write. Observability practice treats logs, metrics and traces as complementary signals. The same discipline applies here: the log says what happened, and a shared request reference links that event to the step that caused it.
OWASP's authorization guidance recommends denying access by default and checking permissions on each request rather than once at setup. For agents, this means the tool layer checks the grant and the approval status at the moment of a write, not when the agent was configured months earlier.
A practical way to decide write permissions is a short decision guide:
- Read: the agent views the data it needs to answer a question. Every read is logged under its identity.
- Draft: the agent prepares an action, such as a reply or a record update, and a named person sends or posts it.
- Write with approval: the agent performs the action only after a named approver accepts the specific payload.
- Never: actions such as moving money, changing bank details or deleting records stay with people, whatever the agent's track record.
Illustrative scenario: a support agent drafts a refund note. The log shows the agent identity, the order reference, the drafted text and the approver who accepted it. When a customer later disputes the amount, the team can see that the agent prepared the note, that a named person approved it and what the system returned. The permission design behind this pattern is described in Least privilege for AI agents: start with the smallest access.
Review, rotation and retirement
Agent identities need the same lifecycle as staff accounts, and they are easier to forget because nobody sees them each day. Put every identity on a review calendar with a named owner. At each review, confirm that the role still exists, that the scopes still match the work, that credentials were rotated on schedule and that no tool quietly added a permission nobody approved.
Retirement deserves its own checklist. When a role ends, disable the agent mailbox, revoke delegated grants, rotate or delete service credentials, remove the agent from scheduled jobs and archive its audit logs according to your retention policy. If a connection breaks during this process, a clean recovery matters more than a fast one. What should an AI agent do when a connection expires? explains how to keep a broken credential from becoming a silent failure.
NIST's AI Risk Management Framework treats governance, mapping, measurement and management of AI risk as continuing activities rather than one-time approvals. Identity review is one concrete place where that continuing work becomes visible to the business.
Where to start with Be Human
Be Human designs and operates personalized AI infrastructure in which agents keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement starts with one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Identity is part of that definition: the agent receives its own mailbox and scoped accounts, the owner and approvers are named, and the audit fields are agreed before any write is enabled. Managed operations can then cover monitoring of the audit trail, periodic permission reviews, incident response when a credential is misused and controlled changes when a vendor alters how its identities work. The service is delivered remotely, in English or Portuguese.
Choosing which role deserves an identity first is often the hardest decision. The free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Can an agent share a company inbox with staff instead of having its own address? A shared inbox mixes agent and human messages, so the log cannot show which actor sent what. If you must start that way, give the agent a distinct sender name and limit it to drafts that a person sends. Move to a dedicated mailbox as soon as the agent needs to send messages on its own.
Does every agent need its own email address? Not always. An agent that only reads a database may need a service identity and no mailbox at all. Give it an address when it must receive or send messages in a channel people rely on, since replies and notifications need a destination the team monitors.
What happens to an agent's access when an employee leaves? If the agent runs on the departing employee's account or grant, its access can stop without warning or continue without anyone noticing. This is why delegated grants should be issued to the agent's own identity and reviewed during offboarding. Add agent identities to the offboarding checklist so that ownership passes to a named successor.
Do I need an identity for an agent that only reads data? Yes, although the identity can be lighter. A read-only identity still appears in logs, so reads remain attributable and you can see which data the agent reached. Read-only access is the easiest stage at which to get identity right, before any write is considered.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


