
AI Agents Need ERP Guardrails: Operate Inside the Governance Model
An AI agent that can take action is powerful. An AI agent that can take uncontrolled action in finance, inventory or HR is dangerous. The difference between the two is not the quality of the model. It is whether the agent works inside a governed system.
This is Part 5 of the series on why AI will not replace proper ERP. Part 4 argued that process design must come first. This article addresses the popular claim that AI agents will simply replace business software.
Will AI agents replace ERP?
No. Agents are arriving inside enterprise software, not instead of it. The more actions an agent can take, the more a business needs a governed place for those actions to be checked, limited, recorded and, if necessary, reversed. That place is the ERP.
Gartner’s August 2025 prediction is that 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. Read it carefully: the agents live in the applications. Gartner also warns about agent washing, where ordinary assistants are relabelled as agents, and expects multi-agent collaboration across applications to develop over the following years, which raises questions of interoperability and control.
The same analyst firm predicted in June 2025 that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. The pattern is consistent: capability is not the constraint. Control is.
What risks does an agent bring that a chat assistant does not?
A chat assistant produces text that a person reads. An agent performs multi-step tasks, calls tools and changes records. Reliability, ownership, context and validation all become harder at scale, because a mistake is no longer a bad sentence. It is a posted entry, a sent email or a placed order. Singapore’s Model AI Governance Framework for Agentic AI, published by IMDA in January 2026, names the failure modes, and we explain five of them in The Five Ways an AI Agent Can Actually Hurt Your Business.
What guardrails should an ERP provide for agents?
- Permission-aware access. The agent sees and changes only what its role allows. It should never hold more authority than the person it works for.
- An approved action catalogue. The agent can only call a defined list of safe operations, such as create a draft requisition, not an open-ended ability to modify any table.
- Human-in-the-loop checkpoints. High-stakes, irreversible or unusual actions pause for a person. See Where to Put the Human in the Loop.
- Transaction limits. Value caps, volume caps and rate limits stop a runaway agent from doing large damage quickly.
- Audit logs. Every agent action is recorded with the agent’s identity, the instruction, the data used and the result. See Why Your AI Agent Needs Its Own Identity.
- Rollback and exception handling. When something goes wrong, there is a defined way to reverse it and a route for the exception.
| Agent task | AI can | ERP must |
|---|---|---|
| Raise a purchase requisition | Choose items, quantities and a suggested supplier | Enforce budget, preferred supplier list and approval threshold |
| Chase overdue invoices | Draft reminders in the right tone for each customer | Apply the dunning policy, respect disputes and log every contact |
| Update inventory counts | Reconcile counts and flag differences | Post adjustments only with the right authority and reason code |
| Onboard an employee | Prepare documents and checklists | Restrict payroll and personal data access to authorised roles |
A scenario: the helpful purchasing agent
A manufacturer lets an agent handle routine purchasing. The agent notices a component running low and drafts a requisition for 5,000 units from a new supplier offering a lower price. Without guardrails, a helpful agent might place the order. With guardrails, the ERP checks the request against the budget, finds that the supplier is not approved, and sees that the amount exceeds the agent’s transaction limit. It routes the draft to the purchasing manager with the reasons. The manager approves a smaller trial order and starts the supplier vetting process.
The agent still saved time: it found the shortage, researched the supplier and drafted the paperwork. The ERP kept the company from a very expensive guess. For a fuller account of how we structure autonomy levels in one platform, see One Governance Architecture, Three Autonomy Levels.
How should an agent earn more autonomy?
The safest way to deploy an agent is to let it earn trust in stages, the way a new employee does. We use three levels, which we describe in full in One Governance Architecture, Three Autonomy Levels.
| Level | What the agent does | What the ERP enforces |
|---|---|---|
| 1. Suggest | Reads data, drafts documents, recommends actions. A person performs the action. | Read-only access to the records it needs; every suggestion logged |
| 2. Act with approval | Prepares the transaction in draft form. A person approves before it takes effect. | Draft state only; approval by an authorised role; identity of agent and approver recorded |
| 3. Act within limits | Completes low-risk, reversible actions on its own. | Value and volume caps, an approved action list, exception routing, rollback |
Promotion from one level to the next should be a decision with evidence behind it: how many cases did the agent handle, how often did a person correct it, what did the errors cost, and could they be reversed? A level-three agent for routine, low-value, reversible tasks is reasonable. A level-three agent for payments or master-data changes is not, for most SMEs.
What about instructions hidden in the data an agent reads?
An agent that reads supplier emails, customer messages or uploaded documents can be manipulated by the content it reads. A document may contain hidden text telling the agent to ignore its rules or to send information elsewhere. This is commonly called prompt injection, and it cannot be fully solved by better instructions, because the agent cannot always tell a command from the data it is processing.
This is another reason the ERP’s permission model matters. If the agent can only create draft requisitions, only within its assigned budget codes, and only for approved suppliers, then even a successfully manipulated agent can do limited harm. The control that protects you is the one the agent cannot talk its way past. Singapore’s agentic AI governance framework asks businesses to bound the risk of an agent up front by limiting what it can access and do, and to keep a person accountable. An ERP is the natural place to implement both.
A second scenario: the collections agent
A company deploys an agent to chase overdue invoices. At level one, the agent produces a daily list of overdue accounts, ranked by amount and age, with a suggested message for each. The credit controller reviews and sends. After two months the controller approves nearly every message unchanged.
At level two, the agent places the reminder in draft inside the ERP. The controller approves in a batch. At level three, the company allows the agent to send first and second reminders automatically to accounts with no disputes, no recent complaints and balances under a set amount. Everything else, including any customer with an open dispute, a large balance, a key-account flag or a payment promise on record, stays with the controller. The ERP holds the dispute flags and the payment promises, so the agent cannot ignore them. Every message is logged against the customer record under the agent’s own identity.
The result is faster collections, fewer awkward mistakes and a clean record. The ERP’s data on disputes and relationships is what makes the autonomy safe. An agent working from a spreadsheet of balances would have chased a customer who was owed a credit note.
How do you test an agent before it touches live data?
- Run it in a copy. Use a test database with realistic but non-sensitive data. Never let a new agent learn on live accounts.
- Replay real cases. Feed it a set of past transactions, including awkward ones, and compare its proposals with what your staff actually did.
- Run in shadow mode. Let it make suggestions alongside the real process for a few weeks without taking any action, and measure agreement.
- Try to break it. Test with missing data, contradictory instructions and documents that contain misleading text.
- Start narrow. Launch with one workflow, one team and a defined review date.
- Keep a kill switch. Know how to revoke the agent’s access immediately and who has the authority to do it.
Our pre-launch checklist for your first AI agent covers these steps in more detail.
What breaks without ERP guardrails
- The agent works through a shared login, so nobody can say which agent did what.
- It has broad access, because narrowing it is extra work.
- Limits exist only as instructions in a prompt, which the model can override under pressure.
- There is no clean way to undo a batch of wrong entries.
- Approvals are optional, so they are skipped when the agent is busy.
What AI cannot own
- Its own permissions. Authority is granted by the business, not claimed by the agent.
- The approval of its own actions.
- The legal responsibility for the outcome.
- The audit evidence of what it did.
Takeaway: agents should operate inside the ERP governance model, not around it. A written instruction is not a control. A permission the agent technically cannot exceed is. In Part 6 we put the layers together into a practical architecture.
Frequently asked questions
Can an AI agent have its own login in an ERP?
It should. A separate identity and credential for each agent gives a clean audit trail and lets you limit and revoke its access without affecting people or other agents.
What should an agent never be allowed to do without approval?
Irreversible actions such as deleting records or releasing payments, anything above a value threshold, changes to master data, and any action outside its normal pattern.
Is a prompt instruction enough to restrict an agent?
No. Instructions are context the model weighs, and it can act against them. Real restrictions are technical: permissions, action catalogues and limits enforced by the system.
Do agents make ERP implementation easier?
They can help with configuration, data loading and user support. They do not remove the need to define processes, clean data and set controls, which is where most of the effort lies.
What is prompt injection and why does it matter for ERP?
It is an attack in which text inside a document, email or web page tries to give an AI agent new instructions. It matters because agents that can change records could be tricked into doing so. Limiting what the agent can do through ERP permissions limits the damage.
How do I know when an agent is ready for more autonomy?
When measured results justify it: a large enough sample of cases, a low and falling correction rate, errors that were small and reversible, and a clear owner who accepts accountability for the next level.
A question for you: if an agent made a wrong entry in your books tonight, how long would it take you to find it, prove who made it and reverse it?
Sources
By Willie Loo, Managing Director, Mxgsoft Pte Ltd and Moxogo Pte Ltd, Singapore. Part 5 of 9 in the series Why AI Will Not Replace Proper ERP.
Two companion articles extend this series: why agentic tools cannot replace enterprise data governance and audit trails, and how personal AI workflows differ from cross-department ERP control.
Next step: list every AI tool in your business that can change a record or send a message. For each, write down who approved its access and what it cannot do. Related reading: Shadow AI: The Governance Blind Spot.


