AI meets an ERP in three ways: the AI your ERP vendor builds into the product, a general assistant such as Claude, ChatGPT or Copilot working on exports or through a connector, and a custom workflow built on the ERP’s API that reads and writes with approvals. Built-in AI is the easiest start for tasks the vendor has designed for. An assistant is good for questions and one-off analysis. A workflow is what you need when the work repeats, has exceptions and has to change records. In all three, the ERP stays the source of truth.
Product facts are as of November 2026. Vendors ship AI features monthly, so check their documentation before you plan around any of them.
The three ways AI meets your ERP
| Approach | What it is | Good for | Where it stops |
|---|---|---|---|
| Built-in AI | Assistants and agents the ERP vendor ships inside the product | Standard tasks the vendor designed for, inside one system | Your own rules, exceptions and the systems around the ERP |
| General assistant | Claude, ChatGPT or Copilot reading an export or connected through a connector | Questions, analysis, drafting, one-off clean-ups | Repeated work with a queue, approvals and an audit trail |
| Custom workflow | Software built on the ERP’s API for one job, with rules and approval steps | High-volume work that has to read, decide and write back | Needs the process written down first, and someone to keep it running |
1. The AI built into your ERP
Every major ERP vendor now ships its own assistant. The names change often, but the pattern is the same: a chat panel that answers questions about your data, plus a growing number of agents for specific jobs.
- SAP. Joule is SAP’s AI copilot across its cloud applications. SAP also offers Joule Agents for multistep tasks in areas such as finance and procurement, and Joule Studio in SAP Build for companies that want to build their own. SAP Business One, the version many smaller companies run, added “Ask AI” and AI-generated user-defined queries in August 2026, so users can ask for a list like “top five customers by outstanding payments” in plain language.
- Oracle NetSuite. Ask Oracle is a natural-language assistant that NetSuite began rolling out to US and Canadian customers with NetSuite Next. Bill Capture reads vendor bills sent by email or upload into NetSuite vendor bills; Oracle’s documentation says it is available only in the United States.
- Microsoft Dynamics 365 Business Central. Copilot covers chat, analysis of lists, autofill, bank reconciliation suggestions and sales line suggestions. Two agents work from a shared mailbox: the Sales Order Agent drafts quotes from customer emails for you to review, and the Payables Agent turns emailed PDF invoices into draft purchase invoices matched to purchase orders.
- Sage. Sage Copilot is embedded in Sage products. In Sage Intacct it includes a close assistant, ledger reconciliation and variance analysis. Sage says availability differs by product and region.
The strength of built-in AI is that it already knows your ERP’s data model and respects its permissions. The Payables Agent is a good example of a sensible design: Microsoft’s documentation says the agent doesn’t post invoices, a person reviews every draft, and any vendor it creates from an invoice starts blocked until someone approves it. That is the level of control to look for in any AI that touches money.
The limit is that built-in AI follows the vendor’s idea of the process, not yours. If your team checks a margin rule in a spreadsheet before confirming an order, the built-in feature won’t know. And it stops at the edge of the ERP: it won’t see the email thread or the shared Excel file where half the decisions are made.
2. General assistants working with your ERP
The second route is a general assistant your team may already use: Claude, ChatGPT or Microsoft Copilot. There are two ways to give it ERP data.
Exports
The simplest is to export a report (open purchase orders, aged debtors) and upload it to the assistant. This works for a one-off question: “which suppliers delivered late more than twice this quarter?” The assistant works on a copy, so nothing it does changes the ERP. That is the safety and the limit: the answer is only as fresh as the export, and someone still has to key changes back in.
Connectors and MCP
The newer way is a connector. Most of them use the Model Context Protocol (MCP), an open standard that lets an assistant call another system’s tools: search records, run a report, and in some cases create or update data. Some ERP vendors now publish their own MCP servers:
- NetSuite AI Connector Service connects external AI assistants to NetSuite. With the MCP Standard Tools SuiteApp, the assistant can run reports, saved searches and SuiteQL queries, and create or update records. Oracle says it uses the same roles and permissions as the NetSuite interface, and recommends a separate role with only the permissions the AI should have.
- Business Central MCP server lets agents built in Copilot Studio (and other MCP clients) work with Business Central. Microsoft’s default configuration is read-only; creating, changing or deleting data has to be configured explicitly.
On the assistant side, Claude supports custom connectors to remote MCP servers; on Team and Enterprise plans only owners can add them, and owners can limit which actions a connector may take, for example allowing reads but blocking writes. ChatGPT supports custom MCP servers with write actions on its Business and Enterprise plans, where admins publish the app and ChatGPT asks for confirmation before risky actions. Check each vendor’s current plan page for what your plan includes.
A connector makes the assistant far more useful for questions, but it doesn’t turn a chat into a process. There is no queue of cases, no guarantee the same check runs every time, and the record of who approved what lives in a chat history, not in the ERP. Give it write access only for tasks you would let a new colleague do unsupervised.
3. Custom workflows on the ERP’s API
The third route is software built for one job, using the ERP’s API: an order desk that reads customer emails and creates draft sales orders, or an invoice check that matches lines and routes only the exceptions. AI handles the parts that need reading and judgement (pulling fields from a PDF, classifying an email, matching a product description to an item code). Fixed rules handle the rest. A named person approves anything above a threshold, and every write to the ERP is logged with who approved it.
This is the most work to set up, and it only pays off on jobs with volume. It also depends on something most companies don’t have: a written description of how the job is really done, including the exceptions. The documented process and the one your team follows on a Tuesday when a supplier is late are rarely the same, and the difference usually sits in a layer of habits around the ERP: a spreadsheet of agreed prices, an email to the warehouse, a person who knows which customers always order by phone. Our lesson on documenting the exceptions shows how to capture that layer before you build anything.
The risks to manage, whichever route you take
- Write access. Reading is low risk; writing is where mistakes cost money. Start read-only and add writes one record type at a time.
- Permissions. Give the AI its own role, scoped to what the task needs, never an administrator’s.
- Audit trail. For every record the AI creates or changes, you should be able to answer: what triggered it, what data it used, who approved it and when. If the answer lives only in a chat window, it isn’t an audit trail.
- Fallback. Decide what happens when the AI is unsure, wrong or down. The answer should be a named person and a queue, not “someone will notice”.
AI can read from the ERP, propose changes and write approved changes back. It should never become a second place where orders, prices or balances live. If a figure in an AI answer and a figure in the ERP disagree, the ERP wins, and the disagreement is a bug to fix.
Which approach fits which ERP task
| Task | Built-in AI | General assistant | Custom workflow |
|---|---|---|---|
| Order entry from customer emails | Good if your ERP has an order agent and orders are standard | Can draft an order from one email; you key it in | Best fit at volume, with your price rules and customer exceptions |
| Three-way match | Good start: several ERPs now match invoices to orders and receipts | Useful to investigate one mismatch, not to run the process | Best fit when tolerances, partial deliveries and routing are specific to you |
| Supplier onboarding | Can prefill a vendor card from an invoice | Good for checking a supplier list for duplicates and gaps | Fits when onboarding spans forms, bank checks and several approvers |
| Stock queries | Good: ask in plain language inside the ERP | Good through a read-only connector | Rarely needed |
| Month-end close | Good for checklists, reconciliation suggestions and variance notes | Good for explaining variances from an export and drafting commentary | Fits for recurring checks that pull from several systems |
A practical way to choose: if the built-in feature handles the task the way your team actually does it, use it. If you need an answer once, use an assistant. If the task repeats every day, has exceptions and must change records, scope a workflow. The lesson from process map to automation has a scoring method for deciding which step to start with.
Before you build: map the job, not the screen
Whatever you pick, start by following one job from start to finish, not by opening the ERP. Interview the team leads first for breadth, then the people doing the work for depth; short voice conversations work better than workshops. Write down each step, who does it, which system it touches and what happens when it goes wrong. Keep one evidence rule: nothing goes into the description unless someone actually said it. When two teams describe the same hand-off differently, you have found a step nobody owns, and that is exactly where an AI will fail if you automate it blind.
Data and privacy
- What not to paste. Bank details, payroll, personal data about customers or staff, and anything under a confidentiality agreement should not go into a consumer AI account. Prefer connectors with scoped permissions over pasting exports.
- Business vs consumer plans. Anthropic and OpenAI both say they don’t train their models on content from their business plans by default. On consumer plans, data can be used for training depending on your settings. Use business plans for company data and check the current terms.
- Admin controls. Business Central lets administrators deactivate individual Copilot features and control which users can use them. Claude and ChatGPT let workspace owners decide which connectors and actions are allowed. Use these controls before you roll anything out.
- GDPR (EU and UK readers). Supplier and customer records contain personal data, so the AI vendor acts as a processor and you need a data processing agreement. Check where data is processed: Business Central, for example, shows an “Allow data movement” setting when some Copilot features run in a different region from your environment. This is not legal advice; check with your data protection lead.
A one-page brief to scope an AI use case on your ERP
Fill this in before you buy a feature or start a build. If you can’t fill a line, that is the work to do first. A written procedure for the task (see how to write an SOP) makes most of it quick.
Task: one job, start to finish (for example: “turn customer order emails into draft sales orders”). Trigger: ___ Volume per week: ___
Owner: the person accountable for the result: ___ Who approves: ___
Data read: which ERP records (customers, items, price lists, open orders) and which other sources (mailbox, spreadsheet, portal): ___
Data written: which records the AI may create or change, and in what state (draft only / ready for approval / posted): ___
Approvals: what a person must confirm before anything is written, and thresholds (for example: any order line more than 2% off the price list): ___
Exceptions: the known cases that need a person (missing data, numbers that don’t match, urgent customer, new supplier): ___
Permissions: the role the AI runs under and what it can’t touch (bank details, payments, payroll): ___
Audit: where each action is logged, with trigger, data used, approver and time: ___
Fallback: what happens when the AI is unsure, wrong or unavailable, and who picks up the queue: ___
Approach: built-in AI / general assistant / custom workflow. Why: ___
Measure: the number today and the target (time per order, error rate, exceptions by cause): ___ Review date: ___
If you want to see where AI could help your own operations first, the free Bonofici grader reads your company’s public pages and shows where to start.