For decades, ERP systems have been where business data goes to be recorded. Orders, invoices, inventory, payroll: it all lands in the ERP, and people do the work of interpreting it and acting on it.
AI agents in ERP change that. Instead of waiting for someone to run a report, spot a problem, and fix it, an agent can monitor the data, reason about it, and act within limits you define. Microsoft describes agents joining teams as "digital colleagues" that take on specific tasks, and points out that ERP systems are full of high-volume, rules-based work that suits automation. If you're new to how AI tools connect to business data, our guide to connecting Power BI to Claude AI using an MCP server explains the protocol behind many of these setups.
This guide covers how enterprises are combining AI agents and ERP systems, the integration patterns that work, what the major platforms offer, what it costs, where projects fail, and how to roll it out without becoming a statistic.
What are AI agents in ERP?
An AI agent in ERP is software that uses a large language model to understand a goal, query ERP data, decide on a next step, and carry it out through the ERP's tools or APIs. A human usually approves anything high-risk.
That sets it apart from tools you may already use:
| Traditional automation / RPA | AI copilot | AI agent | |
|---|---|---|---|
| Follows | Fixed scripted rules | User prompts | A goal, within guardrails |
| Handles exceptions | Poorly; it breaks | Suggests, human acts | Reasons and acts, escalates when unsure |
| Works across systems | Only as scripted | Rarely | Yes, through connected tools |
| Human role | Builds and maintains rules | Does the work, with help | Supervises and approves |
A useful framing from the pages I reviewed: the ERP remains the system of record, and agents act as an execution layer on top of it. They don't replace the ERP. They take over repetitive, low-judgment work that used to mean someone clicking through screens.
Why enterprises are doing this now
Four things came together.
1. ERP vendors shipped agent infrastructure. Oracle announced its SuiteAgent frameworks on October 7, 2025, alongside the AI Connector Service. SAP's Joule platform has expanded rapidly, though reported agent counts vary by source and by which SAP product is being described. Microsoft has released ERP agents such as a Financial Reconciliation Agent and a Sales Order Agent. One 2026 roundup notes that every major ERP vendor now ships some form of agentic AI, including SAP, Oracle Fusion, Microsoft Dynamics 365, NetSuite, and Odoo, at different depths.
2. Open standards lowered integration cost. The Model Context Protocol (MCP) gives AI assistants a standard way to connect to business tools instead of a custom integration for every pairing. Oracle's materials describe the NetSuite AI Connector Service as a secure way for AI assistants to talk to NetSuite and run actions through MCP.
3. Analysts expect agents everywhere. Gartner forecast that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from fewer than 5% the year before.
4. Pressure on hard-to-connect systems. One ERP analyst argues agents put pressure on platforms that are difficult to connect to or slow to change, pushing large enterprises toward modular, API-first architectures rather than full-suite replacement.
Three ways enterprises combine AI agents and ERP
Most deployments follow one of three patterns. Choosing early saves months.
| Pattern | How it works | Best for | Main trade-off |
|---|---|---|---|
| 1. Embedded agents | Use agents the ERP vendor builds and runs inside the platform (such as SAP Joule agents or NetSuite's native AI features) | Standard processes, fast time to value | Less flexibility; tied to the vendor's roadmap |
| 2. External agents via MCP or API | Connect your own AI assistant or agent platform to the ERP through a governed interface | Custom workflows and multi-system processes | You own design, security, and monitoring |
| 3. Orchestration layer | A workflow or integration tool (such as n8n, MuleSoft, or Celigo) coordinates agents, ERP, CRM, and other systems | Cross-system processes like order-to-cash | An extra layer to maintain |
Mature setups often combine all three: embedded agents for standard finance tasks, external agents for custom needs, and an orchestration layer tying the systems together.
How the major ERP platforms support AI agents
NetSuite
NetSuite offers two routes. The AI Connector Service lets customers connect external assistants such as ChatGPT and Claude while keeping NetSuite's permissions, controls, and governance in place. SuiteAgents are built with the SuiteCloud Development Framework and run inside NetSuite. The agentic workflow experiences in NetSuite Next let a human watch progress, review results, and step in.
On the tooling side, the MCP Standard Tools SuiteApp exposes common actions, such as loading records, running saved searches, and SuiteQL queries, as tools an agent can call. Announced NetSuite Next capabilities include operational monitoring, anomaly detection, root cause analysis, financial workflow automation, and continuous accounting close.
One practical point from a setup guide: connect external AI clients using a scoped, non-administrator role, never an admin login.
SAP
SAP's approach is platform-driven. Prebuilt Joule Agents suit standard processes in finance, HR, or procurement where you want fast time to value. SAP's own example is Mota-Engil, which eliminated manual goods receipt entry and sped up invoice clearance using a procure-to-pay agent. For custom needs, Joule Studio lets enterprises build agents around their own industry, rules, and approval workflows.
Microsoft Dynamics 365
Microsoft emphasizes low-code agent building and tight integration with its wider data and AI stack. Copilot Studio offers no-code and low-code tooling for custom ERP agents, aimed at finance, supply chain, and project operations.
Mid-market and open platforms (Odoo and others)
Smaller and open-source ERPs may not ship a full agent catalog, so the external pattern, with agents connected through APIs and workflow tools, is the most common route. The principles are the same: governed access, scoped permissions, and human approval for risky actions.
Six use cases where AI agents deliver value in ERP
Financial reconciliation and bank matching. Agents match transactions, flag exceptions, and prepare entries for review.
Accounts payable. Reading invoices, matching them to purchase orders and receipts, and routing discrepancies.
Month-end close support. Watching open items and anomalies so close doesn't depend on a last-minute scramble.
Procurement intake and supplier analysis. Capturing requests, routing approvals, and comparing supplier bids.
Order management and customer service. Checking stock, order status, and delivery information from plain-language requests.
Conversational reporting. Letting a finance lead ask a question and get an answer drawn from ERP data, with the source shown.
These share the same traits: high volume, clear rules, and a human who can check the result. Start there.
What does it cost to add AI agents to an ERP?
Prices vary widely, and the license is rarely the biggest line. Budget for:
Integration work. Guides on enterprise AI agents say this is frequently underestimated: connecting agents to ERP, CRM, and databases, and building or adapting APIs.
Data cleanup. Agents amplify bad data, so master data work often comes first.
Usage-based AI charges. Some vendors bundle a baseline and meter the rest, and overage pricing can be hard to predict, so ask for clear terms.
Build vs. buy. Prebuilt embedded agents cost less to deploy. Custom agents cost more but fit your processes.
Ongoing monitoring. Agents need review, tuning, and support after go-live.
To judge return, measure a baseline before launch, such as hours spent on reconciliation, invoice cycle time, or exception rates, and compare after the pilot.
The risks: why many agent projects stall
Hype needs a counterweight. Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. Forrester's 2026 research, as summarized by one analyst, found that roughly 75% of enterprise leaders report adopting agentic AI while only a small minority run it in meaningful production.
Gartner also flagged "agent washing," where vendors rebrand chatbots or RPA as agentic AI. It believes only around 130 of thousands of vendors offer genuine agentic capabilities. When evaluating vendors, ask to see an agent complete a real multi-step task on real data.
Most failures are organizational, not technical. Projects stumble when companies give systems access and authority before defining governance, ownership, and rollback controls. Data and integration friction is another blocker: agents are only as useful as what they can actually reach, and fragmented data and brittle integrations limit them quickly.
A governance checklist before any agent touches your ERP
Least-privilege access. Give each agent its own role, scoped to the minimum. With MCP-style setups, AI access can mirror existing ERP role-based controls, so a finance clerk's agent can't read HR payroll.
Human approval for high-impact actions. Payments, journal postings, vendor bank changes, and price changes should need sign-off.
Full audit trail. Log every query, decision, and action with the agent's identity.
Data quality first. Clean customer, vendor, and item records before automating.
Regulatory fit. Confirm the setup meets the data protection and audit requirements that apply to your business.
Named ownership. One person is accountable for each agent's behavior.
Rollback and kill switch. You must be able to pause an agent and reverse its actions.
A phased rollout plan
Phase 1: Pick the process (weeks 1–2). Choose one high-volume, rules-based workflow with measurable pain, such as bank reconciliation or invoice matching.
Phase 2: Fix the foundation (weeks 2–4). Clean data, define roles, and decide which of the three integration patterns fits.
Phase 3: Pilot with approvals (weeks 4–8). The agent recommends and prepares; humans approve. Test on real data.
Phase 4: Expand autonomy gradually. Let the agent act alone on low-risk, high-confidence cases. Keep approvals for the rest.
Phase 5: Monitor and improve. Review exceptions, accuracy, and cost monthly. Retire what doesn't pay off.
This staged approach guards against the pattern behind the cancellations Gartner describes: starting big, with no measurable value and no controls.
What it means for NetSuite customers
If you run NetSuite, you start from a strong position. The platform offers a governed MCP gateway, a framework for in-platform agents, and a role-and-permission model that agents can inherit. The practical questions are which processes to automate first, whether to use native features or external agents, and how to connect the surrounding systems.
The enterprises getting value from AI agents in ERP aren't the ones with the most agents. They're the ones with clean data, tight permissions, a narrow first use case, and a clear way to measure results. Start small, keep humans in control of high-impact decisions, and expand as trust is earned.
Ready to plan your first agent? Explore Versich's AI services to see how we build and integrate AI into the ERP and business systems you already run.
