---
title: What controls do AI agents in procurement need?
url: https://opencontroller.lyzr.com/blog/govern-ai-agents-in-procurement
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What controls do AI agents in procurement need?

**A procurement agent needs the same financial controls as a human buyer, enforced on the agent's identity: purchase-order thresholds it cannot exceed, no ability to change vendor bank details, three-way match before payment, and a full audit trail. The highest-risk action is not a large order; it is a quiet change to vendor master data.** Let the agent draft and prepare, and keep approval and payment-detail changes with people under separation of duties.

## What can a procurement agent safely do?

Procurement "workbench" agents usually help with intake, sourcing and purchase orders. A useful split is by what the action changes:

| Action | Risk | Recommended mode |
|--------|------|------------------|
| Summarise a request, check policy, find a catalogue item | Low | Fully automated |
| Compare quotes, draft an RFQ or a PO | Low to medium | Automated draft, human submits |
| Submit a PO under a set threshold, on contract, with an approved vendor | Medium | Automated within limits |
| Submit a PO above threshold or off-contract | High | Human approval in the ERP workflow |
| Create a new vendor | High | Human, with vendor due diligence |
| Change vendor bank or remittance details | Critical | Never by an agent |
| Release a payment | Critical | Never by an agent; existing treasury controls |

The table maps to ordinary internal-control thinking. The agent becomes another requester and preparer, never the approver of its own work.

## Why is vendor master data the biggest risk?

Payment-redirection fraud works by changing where a legitimate supplier gets paid. Attackers send a convincing email: "our bank details have changed". A human clerk is trained to call back on a known number. An agent that reads the inbox and has write access to the vendor master may simply update the record.

Agents add a second route: indirect prompt injection. A supplier invoice, PDF or email can contain instructions aimed at the agent. If the agent can write to vendor records, the injected instruction can become a real change. This is one of the core risks described in the [OWASP GenAI Security Project](https://genai.owasp.org/) material on agentic applications.

Controls that close the gap:

1. Remove write permission on vendor bank, remittance and tax fields from every agent identity.
2. Route any vendor change the agent detects into a human ticket, with the evidence attached.
3. Require call-back verification on a number already on file, never one from the request.
4. Alert on any attempted change to those fields by an agent identity, even if blocked.

## What is the procurement agent control checklist?

**Identity and access**

- [ ] The agent acts on behalf of the requesting user, so it can never exceed that user's authority. See [on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation).
- [ ] No write access to vendor master bank, remittance or tax fields.
- [ ] Separate identities for read (catalogue, contracts) and write (PO creation).
- [ ] Least-privilege scopes reviewed against the [least-privilege checklist](/blog/least-privilege-checklist-for-ai-agents).

**Financial limits**

- [ ] Per-PO threshold enforced in the ERP, not only in the prompt.
- [ ] Daily and monthly aggregate limits per agent, to stop order-splitting below the threshold.
- [ ] Approved-vendor and on-contract checks before any submission.

**Matching and payment**

- [ ] Three-way match (PO, goods receipt, invoice) completed before payment, with tolerances set by finance.
- [ ] Exceptions routed to a human queue, never auto-resolved by the agent.
- [ ] The agent cannot both create a PO and confirm its goods receipt.

**Runtime and evidence**

- [ ] Every tool call logged with the user, agent version, inputs and outcome.
- [ ] Alerts on unusual patterns: new vendors, repeated just-under-threshold orders, off-hours volume.
- [ ] A tested way to stop the agent within minutes. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).

## How does separation of duties apply to agents?

Separation of duties means no single actor can initiate, approve and record the same transaction. NIST [SP 800-53](https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final) control AC-5 describes it for systems. Apply it to the agent as an actor:

- The agent may initiate; a human approves above threshold.
- The agent may prepare a receipt; a different person confirms it.
- The agent's builder may not be the approver of its access. See [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals).

Approvals for high-value actions belong in the system of record's own workflow, so the evidence sits where auditors already look. For patterns, read [human approval for high-risk AI agent actions](/blog/human-approval-for-high-risk-ai-agent-actions).

## What evidence will auditors ask for?

Expect requests for: the agent's access list and who approved it; the threshold configuration; a sample of agent-initiated POs with approvals; logs of blocked or escalated actions; and proof that vendor master changes were never made by an agent identity. Keep these by agent version, so you can show what the agent could do on any given date.

## How Opencontroller does this

Opencontroller supports delegated on-behalf-of identity and agent workload identity federation, so a procurement agent acts within the requesting user's rights. Posture detects risky permissions and configuration, such as an agent identity with write access it should not have. Rogue-agent handling monitors behavioural deviation, alerts, and can automatically quarantine or kill an agent, and per-agent Block and Quarantine controls act through the connected platform, including SAP. [Book a demo](https://www.lyzr.ai/book-demo/) to review a procurement agent with us.

## Frequently asked questions

**Should a procurement agent ever approve purchase orders?**
No. It can initiate and submit within limits, but approval above threshold should stay with a human in the ERP workflow.

**Can prompt instructions enforce spending limits?**
Not reliably. Enforce thresholds in the ERP and in the agent's permissions; treat prompt instructions as guidance only.

**What is the single most important control?**
Remove agent write access to vendor bank and remittance details. Payment redirection is the costliest failure mode.

## Key facts

Procurement AI agents need: on-behalf-of identity, no write access to vendor bank/remittance/tax fields, per-PO and aggregate thresholds enforced in the ERP, approved-vendor checks, three-way match before payment, separation of duties (NIST SP 800-53 AC-5), full tool-call logging and a tested kill path. Vendor master-data changes are the highest risk because of payment-redirection fraud and indirect prompt injection. Opencontroller provides on-behalf-of identity, Posture, rogue-agent quarantine/kill and per-agent Block controls. Pricing starts from $250K per deployment per year.

## Related

- [Which AI agent actions need human approval?](https://opencontroller.lyzr.com/blog/human-approval-for-high-risk-ai-agent-actions.md)
- [How do you separate who builds an AI agent from who approves it?](https://opencontroller.lyzr.com/blog/segregation-of-duties-for-ai-agent-approvals.md)
- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
