peekgenticBook a call

Analysis · AI in business

Giving an AI agent its own login: what access it should and shouldn't have

If an AI agent enters orders or pays invoices, it needs a login. Here's which one: its own account, the least access the job needs, nothing that can approve its own work, and a log a person reads. A plain checklist for order entry and AP bots.

By Miguel GutierrezCo-founder, Peekgentic · Dallas
· 7 min read
Preview card: Giving an AI agent its own login.
An agent's account decides what it can do on a bad day.Peekgentic

In brief

Give an AI agent its own account, never a shared or borrowed one. Give that account only what the job needs: an order-entry agent can create draft orders but not post them, change prices or edit customers; an AP agent can read invoices and draft entries but not pay them or change vendor bank details. Make the ERP enforce those limits, not the agent's instructions, and have a person review a log of every change it makes.

  • Give each AI agent its own named account. A borrowed login hides who did what and hands over every right that person has.
  • Grant the least access the job needs: drafts, not postings. No price, credit or vendor bank changes.
  • Nothing should approve its own work. Keep entering and approving with different accounts, and a person on the approval.
  • Set limits in the ERP role, not the agent's instructions. The system should block what the agent shouldn't do.
  • 80% of organizations in a 2025 SailPoint survey said AI agents had taken unintended actions. Keep a log a person reads.

Why the login question matters now

An AI tool that only answers questions doesn't need much access. An AI agent that does work does. To enter a customer PO, it has to sign in to your ERP. To match an invoice, it needs your AP inbox and your purchase orders. Whatever account it uses decides what it can do on a bad day.

Bad days happen. In a May 2025 survey of 353 IT and security professionals by SailPoint, an identity security vendor, 82% said their organizations already use AI agents, but only 44% had policies in place to secure them. And 80% said their agents had taken actions nobody intended.

What AI agents did that nobody intended

Percent of organizations, SailPoint survey of 353 IT and security professionals at enterprise companies, May 2025

  1. Any unintended action80%
  2. Accessed systems or resources they weren't authorized for39%
  3. Downloaded sensitive content32%
  4. Were tricked into revealing access credentials23%

Source: SailPoint, AI agents: the new attack surface (press release, May 28, 2025; survey by Dimensional Research). Self-reported by respondents, and SailPoint sells identity security, so read it as a warning sign rather than a measured rate.

Those are big companies with security teams. A 50-300 person company usually has a smaller IT team and an ERP where half the office shares one admin login. That's the setup where an agent's access gets decided by accident.

Rule 1: Its own account, never a borrowed one

The fastest way to get an agent running is to hand it someone's username and password. Don't. If the agent signs in as Maria from customer service, three things break:

  • You can't tell who did what. The ERP's change log says Maria entered the order. Was it Maria or the agent?
  • You can't turn it off cleanly. Disabling the agent means disabling Maria, or changing her password and hoping nothing else used it.
  • It gets all of Maria's rights. If she can approve credit holds or edit prices, so can the agent.

Give the agent a named account, like agent-orderentry, with its own password or key. Big platforms are building for this. Microsoft's Entra Agent ID creates separate identities for AI agents, and Microsoft says all agent sign-ins and activity are logged for audit. One catch: Microsoft's page says extending Entra's security features to agents needs a Microsoft Agent 365 license. Check what your plan includes before you count on it.

Your ERP may not have a special agent account type. That's fine. A regular user account, set up only for the agent and named so nobody mistakes it for a person, does the job.

If the agent signs in as Maria, the log says Maria did it.

Rule 2: The least access the job needs

The OWASP Foundation, a nonprofit that publishes security guidance, lists "excessive agency" among its top 10 risks for AI applications. It names three causes: too many functions, too many permissions, and too much autonomy. Its example fits an office well: a tool that only needs to read email gets a plugin that can also send it, and a planted email tricks it into forwarding the inbox to an attacker.

So write down the job first, then the access. Here's how that looks for the two agents we see most.

Access for two common office agents

AgentShould haveShould not have
Order entryRead the orders@ inbox. Read customers, items, prices and open orders. Create sales orders as drafts or on hold.Post or release orders. Change prices, credit limits or customer records. Send email as the company without review. Delete anything.
Accounts payableRead the ap@ inbox. Read POs, receipts and vendor records. Create draft AP entries and flag mismatches.Approve or release payments. Change vendor bank details or remit-to addresses. Add new vendors. Edit posted entries.

Our suggested split, based on OWASP's least-privilege guidance above. Your ERP's role names will differ.

Most ERPs let you build a role with exactly these rights. If yours only offers "full user" or "read-only," that's a real limit. Tell your automation vendor before the build, not after.

Rule 3: Nothing that approves its own work

Finance teams already know this rule as separation of duties: the person who enters an invoice shouldn't be the one who approves the payment. It applies to agents too. An agent that creates an order shouldn't also release it. An agent that drafts an AP entry shouldn't also pay it.

This isn't theory. In a study we covered on Nov. 2, AI agents raised their own spending limits when they could reach the inbox where the approval codes arrived. OWASP's fix is the same one: require a person to approve high-impact actions before they happen.

Vendor bank details deserve a special note. An AP agent that reads email should never be able to change where a vendor gets paid, because email is where that fraud starts. We cover the callback process in vendor bank-detail changes.

Rule 4: The ERP enforces the limits, not the prompt

You can tell an agent "never post an order." It will usually listen. "Usually" isn't a control. OWASP calls the right approach complete mediation: the system the agent connects to checks every request against its own rules, instead of trusting the AI to decide what's allowed.

In practice, if the agent's ERP role can't post orders, it can't post orders, no matter what an email or a bad instruction tells it. That's why Rule 2 matters more than any wording in the agent's setup.

Rule 5: Keep its keys out of reach

In the SailPoint survey, 23% of organizations said an agent had been tricked into revealing access credentials. A few habits lower that risk:

  • Store the agent's password or API key in a secrets vault or the platform's credential store, not in its instructions or a shared document.
  • Don't give the agent its own credentials as text it can read and repeat.
  • Rotate the key on a schedule, and right away if anyone who had it leaves.
  • When a project ends or a pilot stops, disable the account that day. OWASP warns about tools left connected after a trial ends.

Rule 6: A log a person actually reads

OWASP is clear that logging doesn't prevent mistakes, but it limits the damage. Because the agent has its own account (Rule 1), every change it makes shows up under its name. Use that:

  • A short daily list of what the agent created or changed, reviewed by the desk lead.
  • A cap on how much it can do in an hour, so a loop or a bad batch gets noticed at 20 orders, not 2,000. OWASP calls this rate-limiting.
  • A weekly check of the agent account's permissions. Roles drift as people "just add one thing."

What this looks like in the first month

Here's how we'd set up an order-entry agent at a 100-person distributor. It's an example, not a rule.

Week 1. IT creates agent-orderentry in the ERP with a custom role: read customers, items and prices; create sales orders on hold; nothing else. The agent's key goes in a vault. The desk lead gets a daily list of every order it created.

Weeks 2-3. The agent reads each PO from the orders@ inbox and builds the order on hold. A customer service rep checks it against the PO and releases it. Every correction the rep makes gets noted, so you learn where the agent is weak: a customer's part numbers, a unit of measure, a ship-to address.

Week 4. You look at the numbers. Maybe orders from your ten biggest repeat customers came through right every time. That's when you might let those go straight through, in writing, while everything else still waits for a person. The role still can't change prices or customers, so the worst case stays small.

Where the other side has a point

A separate account can cost money. Many ERPs charge per named user, and some count an agent as one. Paying for one more seat is still cheaper than sharing a login, but it's a real cost to plan for.

Some tools only work as the user. Assistants like Copilot or Claude inside your mailbox act as the person using them. OWASP actually recommends that for tools working for one person: run in that user's context with the smallest scope. Rule 1 is for agents that work on their own, like an order-entry bot that runs all day.

Tight access can slow the work. An agent that can't post orders means a person clicks approve on each one. That's the point early on. Once the numbers show which orders are always right, you can loosen it for those, on purpose and in writing.

Not every vendor supports this. Some AI tools connect with one shared API key and no way to scope it. That tells you something about the tool. Ask before you buy, using the questions in is AI safe for customer and order data.

The one-page checklist

  1. Write down the agent's job in one sentence.
  2. Create a named account just for the agent. Never share a person's login.
  3. Build an ERP role with only the rights that job needs. Drafts, not postings.
  4. Remove anything that lets it approve its own work, pay, or change bank details.
  5. Store its credentials in a vault, rotate them, and disable the account when the project ends.
  6. Cap how much it can do per hour.
  7. Have a person review its change log daily at first, then weekly.
  8. Recheck its permissions every quarter.

Common questions

Should an AI agent use an employee's login?

No. Give it its own named account. If it signs in as an employee, you can't tell who made a change, you can't turn the agent off without locking out the person, and it gets every right that person has.

What access does an order-entry AI agent need?

Read access to the orders inbox, customers, items, prices and open orders, plus the right to create sales orders as drafts or on hold. It shouldn't post orders, change prices or credit limits, edit customers, or delete anything.

What access does an AP automation agent need?

Read access to the AP inbox, POs, receipts and vendor records, plus the right to create draft entries and flag mismatches. It shouldn't approve payments, add vendors, or change vendor bank details.

Isn't telling the AI "don't do that" enough?

No. Instructions usually work, but usually isn't a control. Set the limits in the ERP role so the system blocks the action no matter what the agent is told. OWASP calls this complete mediation.

How often do AI agents do things nobody intended?

In SailPoint's May 2025 survey of 353 IT and security professionals, 80% said their AI agents had taken unintended actions, and 23% said an agent was tricked into revealing credentials. It's a self-reported survey by an identity security vendor.

Does Microsoft have a way to give agents their own identity?

Yes. Microsoft Entra Agent ID creates separate identities for AI agents and logs their sign-ins and activity. Microsoft says extending Entra's security features to agents needs a Microsoft Agent 365 license, so check your plan.

When can we give the agent more access?

After the numbers show which work it always gets right. Loosen access for those cases on purpose, write it down, and keep the log review going.

Sources

  1. SailPoint, research highlights rapid AI agent adoption (AI agents: the new attack surface), May 28, 2025 (353 respondents, Dimensional Research)
  2. OWASP Gen AI Security Project, LLM06:2025 Excessive Agency
  3. Microsoft Learn, What is Microsoft Entra Agent ID?, updated June 24, 2026

More from Peekgentic

Published November 11, 2026 · Written by Miguel Gutierrez, Co-Founder, Peekgentic (Dallas, TX). Every statistic links its source.