The seven steps at a glance

An AI agent is software that uses an AI model to choose its next step, then acts through tools you give it. This is the order we build one in. Each step has its own section below.

  1. One job first. Pick work with a clear start and end, then decide if it needs an agent or a fixed workflow.
  2. The route in. Find the API your ERP opens to outside software, or the database if it has none.
  3. Its own login. The agent signs in as itself, with only the rights its one job needs.
  4. Narrow tools. Each tool is one small function, such as “save a draft sales order”.
  5. A person at every write. The agent saves drafts, and a person approves them.
  6. A sandbox test. Run the agent on a test copy of your ERP, then read its log.
  7. Limits and dates. Vendors cap how many calls an API takes at once, and retire old sign-in methods on set dates.

Step 1: Pick one job and decide its shape

Start with one job that has a clear start and a clear end. An emailed purchase order that ends as a draft sales order is one. “Use AI in finance” is a wish.

Then decide its shape. Anthropic, the company behind Claude, splits these systems in two. In a workflow, code fixes the steps. In an agent, the AI model picks its own. Anthropic's guide advises the simplest build that works. OpenAI's guide says to build an agent where fixed rules fall short, and that a fixed program may be enough otherwise. Our article on agentic AI vs AI agents compares the two shapes.

Most ERP paperwork fits a workflow, because the steps are known:

  1. Read the document.
  2. Match each line to the item master, the list of every product you sell.
  3. Apply the price list.
  4. Save a draft.

The model's part is the reading. OpenAI names work that leans on unstructured documents as one place where an agent beats fixed rules. Unstructured means no fixed layout, like most emails and PDFs.

Our team's ERP work covers RFQ automation, daily bookkeeping, and purchase and sales order generation, in NetSuite, SAP Business One, Odoo, ERPNext and custom ERPs. An RFQ is a request for a quote.

Step 2: Find the route into your ERP

Every ERP below opens at least one door to outside software, called an API. The table shows the door, what an agent can read and write through it, and where a person checks. We read each vendor's documents on 15 November 2026.

Scroll the table sideways to see every column.

ERPRoute for an outside agentWhat it reads and writesWhere a person checks
NetSuiteREST web services, NetSuite's web API, signed in with OAuth 2.0, a standard way to let an app in without sharing a password. Oracle says every new integration should use them from release 2026.1. Oracle's AI Connector Service is an MCP route, free from Oracle, for Claude Pro and ChatGPT.Create, read, update and delete records, and look data up with SuiteQL, NetSuite's query language. Connector tools run with the user's role permissions only.Approval routing: Oracle says entered transactions are not processed until someone approves them.
SAP Business OneThe Service Layer, a web API built on HTTP and the OData standard (version 4), for both SAP HANA and SQL Server. Release FP 2608 added an open-source MCP server for it.Orders, purchase orders and quotations, plus drafts and other documents. Since FP 2602, webhooks tell an agent when a record changes.With approval procedures on, an order sent through the Service Layer becomes a draft, and SAP returns error code -2028. The draft waits for an approver.
Odoo 20The JSON-2 API at /json/2, on Custom plans only. Odoo AI also offers an MCP server.Any record the agent's user may reach. Odoo checks each call against that user's access rights, record rules and field access. Over MCP, five read tools are on by default; the rest must be switched on.Studio approval rules sit on buttons. Odoo's page does not say whether an API call that skips the button is stopped, so test it in your database first.
ERPNextFrappe's REST API, which Frappe generates for every DocType (ERPNext's name for a record type), with a token login.Any DocType the API user's roles allow. Frappe logs each call against that user.A submittable document without a custom workflow starts as Draft, and a person submits it. Workflows can add several levels of approval.
Custom ERPThe database, on a role made for the agent, or an MCP server or small API built on top of it.Whatever the role allows. In PostgreSQL, row security policies can also limit which rows each user can read or change.A draft table or status that your team reviews before anything posts. No vendor ships this gate, so you build it.

Three of the five now offer an MCP route. MCP, the Model Context Protocol, is a shared standard that lets AI apps such as Claude find and call a system's tools. It is optional, and each ERP's own API works without it. Our NetSuite page covers Oracle's connector and its limits in full.

When the ERP has no API

On 14 November 2026 we asked four AI assistants who can build an AI agent on a custom ERP with no API. All five answers (the ChatGPT app ran twice) named screen automation, also called RPA. RPA software clicks through screens the way a person would. Four of the five answers also named a database connection.

We would try the database first. A database role is a login inside the database with its own rights. A role made for the agent can be granted only the tables its one job reads, plus one draft table it writes.

The built-in read roles are wider than that. SQL Server's db_datareader reads every user table and view. PostgreSQL's pg_read_all_data reads all data, though it still obeys row security policies. So give the agent its own narrow role instead.

Screens come second. OpenAI's guide names computer-use models as the way into old systems without APIs. These are AI models that see a screen and work the mouse and keyboard. Anthropic's documents for its computer use tool add a warning: Claude sometimes follows instructions it finds on screen over yours. Anthropic advises a separate virtual machine with few rights, and a person who confirms any step with real-world consequences. A virtual machine is a walled-off computer that runs inside another one.

Step 3: Give the agent its own login

The agent signs in as itself, with only the rights its one job needs. The OWASP Gen AI Security Project, which lists security risks in AI software, names too many permissions as one cause of what it calls excessive agency. That is an AI system able to take damaging actions when its output goes wrong. Each vendor has a rule that points the same way.

  • NetSuite will not run its AI Connector Service for a user signed in with the Administrator role, or any role with full permissions.
  • SAP Business One gives a Superuser full rights to every module and function, and SAP says those rights cannot be changed. An agent should never sign in as one.
  • Odoo recommends dedicated bot users for integrations. The bot gets the fewest rights, and Odoo's access log then shows the bot instead of a person.
  • Frappe, the framework under ERPNext, checks each token call against the roles of its user, and suggests creating a user just for API calls.
  • Your own database takes a role made for the agent. Grant it reads, plus writes to the one draft table it needs.

Step 4: Write narrow tools

A tool is one function the agent may call, such as find_customer or save_sales_order_draft. Small tools are easier to test and to limit.

Anthropic tells builders to put as much care into tools as into the prompts, the written instructions the model reads. It also says to shape each tool's inputs so mistakes are hard to make. In ERP terms, a tool that saves a draft order takes a customer and the order lines, and offers no way to post.

OpenAI's guide names three kinds of tools: data tools that read, action tools that write, and other agents used as tools. It suggests giving each tool a risk rating of low, medium or high, using four tests:

  1. Does it only read, or does it write?
  2. Can its effect be undone?
  3. What rights does it need?
  4. How much money is at stake?

Business rules belong in the tool's own code. Odoo says so for its AI server actions: the AI picks a tool and its inputs, and the tool must enforce the rules. OWASP gives the general version. The system the agent calls should decide what is allowed, and the AI model should never be the judge of that.

Step 5: Put a person at every write

Reading carries little risk. Writing is where the agent needs a person.

OpenAI's guide says actions that are sensitive, hard to undo or high stakes should go to a person. It names canceling orders and making payments as examples. The MCP specification says a person should always be able to deny a tool call. A human in the loop is the usual name for this.

Use the gate your ERP already has:

  • SAP Business One: approval procedures turn an agent's order into a draft for an approver.
  • NetSuite: approval routing holds a transaction until someone approves it.
  • ERPNext: the agent saves a Draft, and a person submits it.
  • Odoo needs two cautions. Odoo says an agent started by a trigger creates and changes records without asking a user. Its MCP server lets you mark a tool Readonly, but Odoo says that mark only tells the AI client the tool is safe to run unasked.

We build agents to save a draft and wait. A person approves each record, such as a quote, an order or a ledger entry, before it is sent or posted.

Step 6: Test in a sandbox, then read the log

A sandbox is a copy of your system kept for testing. Anthropic recommends extensive testing in one before an agent runs for real, because agents cost more and their errors can compound. Oracle says a NetSuite sandbox copies the live account's setup and data, customizations included, and nothing done in it touches production.

A test worth running, in this order:

  1. Take last month's real documents, such as the purchase orders your team keyed in.
  2. Run the agent on them in the sandbox.
  3. Compare its drafts line by line with what your team entered.
  4. Fix what differs, then run the same documents again.

After go-live, read what the agent changed. NetSuite's System Notes record when a change was made, who made it, from which interface, the field changed, and its old and new values. SAP's Service Layer returns change history through a method called GetChangeLog. In Odoo and ERPNext, the agent's own user name marks its calls in the logs.

Step 7: Check the limits and the dates

Each route has a ceiling or a deadline. These come from the vendor pages we read on 15 November 2026.

Scroll the table sideways to see every column.

SystemLimit or dateWhat it means for the build
NetSuiteA Standard-tier account on a contract from June 2020 or later runs 5 requests at once. One SuiteCloud Plus license raises that to 15. Oracle's connector shares the same limit with every other integration.Count the agent's calls before go-live. Oracle notes that one AI action can send several requests.
NetSuiteFrom release 2027.1, no new integration can use token-based authentication, NetSuite's older sign-in method. SOAP integrations, built on NetSuite's older web services, stop working at release 2028.2.Build on REST web services with OAuth 2.0.
OdooThe external API works on Custom plans only. API keys last three months at most. The older external APIs, XML-RPC and JSON-RPC, are due for removal in Odoo 22 (fall 2028) and Online 21.1 (winter 2027).Confirm the plan and rotate keys on a schedule. Build on JSON-2.
SAP Business OneOData version 3 has been deprecated in the Service Layer since FP 2405.Build on OData version 4.
Any MCP serverThe MCP specification makes authorization, the check on who may call a tool, optional.A server built for your own ERP needs a login added on purpose.

For cost, our article on what an AI agent costs to build sets six published price lists side by side, each with its author and date.

Sources

Every source behind a figure or a vendor fact in this article, with the day we read it.