How to add AI agents to Made2Manage
Short answer
Made2Manage has no modern agent framework of its own, so you build the agent outside it: a small tool layer that reads the SQL Server database (or a replica), exposes narrow functions such as open-PO lookup and shortage list, and queues any write for a person to approve. The model can run on-prem so shop data never leaves the building.
Many Made2Manage sites are job shops and make-to-order manufacturers that have run the system for ten or twenty years. The knowledge of why a job is late lives in a planner's head and in a dozen screens: sales orders, job routings, purchase orders, inventory locations, and the MRP output. The system records everything but it does not reason about it.
An AI agent is a worker that follows a goal across those screens. Instead of a planner opening the shortage report, checking each part's open POs, and emailing the buyer, the agent does the lookup chain and hands over a drafted action. It does not replace the planner's judgement. It removes the clicking.
Because Made2Manage is often self-hosted on an aging Windows server, the practical design is a read-mostly agent beside the ERP, not inside it. You keep the ERP untouched, add a tool layer, and keep a person in the loop for anything that changes a record.
What teams use it for
MRP exception triage
After an MRP run the agent reads planned-order and shortage output, groups exceptions by part and customer ship date, and proposes which to expedite, reschedule or ignore, with the open-PO and stock evidence attached.
Late purchase order expediting
The agent finds POs past their promise date that feed jobs due soon, drafts a supplier email with the PO number, line and needed date, and queues it for the buyer to send.
Job status answers for customer service
Given a sales order number, the agent traces it to jobs, current operation, and material status and writes a plain-language status the CSR can paste to the customer.
Quote cost sanity check
Before a quote goes out, the agent compares routing hours and material cost against recent similar jobs' actuals and flags quotes that look under-costed.
Inventory discrepancy hunting
The agent scans for negative on-hand, stale reservations and receipts without matching PO lines, then lists the suspect records for a stockroom lead.
Draft NCR and rework paperwork
From a shop-floor note about a bad part, the agent gathers job, operation, lot and supplier history and drafts the nonconformance text for quality to review.
Made2Manage integration surface
- The database is the real integration surface. Current Made2Manage releases run on Microsoft SQL Server; very old installs used Visual FoxPro tables. Confirm which you have before designing anything.
- Table and column names are terse and vary by version (jobs, job routing, sales orders, purchase orders, inventory, item master). Map them from your own schema rather than trusting a blog post, including this one.
- Aptean has offered integration tooling and APIs over the years, and the exact options depend on your version and support agreement. Ask your Aptean contact or reseller what is licensed for you.
- Custom screens and reports often exist in your install (Crystal Reports, custom SQL views). Those views are a good source of already-validated business logic for the agent to reuse.
- Writes done outside the supported application path risk bypassing business rules such as costing and inventory transactions. Prefer queueing a request for a person or using a supported import path.
How it works
- 1
Inventory the data and pick a read source
Identify the SQL Server instance and set up a read-only login, ideally against a nightly or log-shipped replica so agent queries never slow the shop.
- 2
Define a small set of tools
Write functions such as get_job(job_no), open_po_for_part(part), shortages_for_job(job_no). Each has typed inputs, a fixed query, and row limits. The model never writes raw SQL against production.
- 3
Choose and host the model
Use an open-weight model on a local GPU server for private data, or a cloud model under your enterprise terms. The tool layer is the same either way.
- 4
Add an action queue
Anything that changes data or sends mail becomes a proposed action stored with its evidence. A planner approves in a simple web screen.
- 5
Execute approved actions through a supported path
Send the email, create the task, or hand the change to a person to key in. Keep the first release away from direct database writes.
- 6
Log and evaluate
Record every tool call and answer, replay 50 past exceptions, and compare the agent's proposals with what planners actually did before widening scope.
Example flows
Morning shortage brief
Trigger: Scheduled at 6:30 after the overnight MRP run.
- Pull jobs due in the next 10 working days with unmet material.
- For each short part, fetch on-hand, allocated quantity and open PO lines with promise dates.
- Classify as covered by PO, needs expedite, or no supply.
- Write a ranked brief and queue draft supplier emails.
Result: The planner opens one summary instead of walking the shortage report line by line.
Customer asks where order is
Trigger: CSR types a sales order number into an internal page.
- Look up the sales order lines and linked jobs.
- Read current operation and last labor posting per job.
- Check material availability for remaining operations.
- Draft a status message with an honest ship estimate range.
Result: The CSR replies in minutes and no one interrupts the floor supervisor.
Keeping data on-prem
Made2Manage is frequently on-prem on a Windows server, and the agent should match. Put the tool layer and a small GPU box on the same network, read from a replica, and serve an open-weight model with Ollama or vLLM so prompts, part numbers and customer names stay inside. For customers with export-controlled drawings or defense work, keep the agent in an isolated segment with no internet egress and update model files through a controlled transfer. The ERP server itself needs no inbound changes, which matters when it runs on legacy operating systems you do not want to touch.
Security checklist
- Use a dedicated read-only SQL login limited to the views the tools need, never the application's own account.
- Do not let the model generate free-form SQL; expose parameterised tools only.
- Require human approval for every write, email or record change, and store who approved.
- Log each tool call with inputs and row counts for audit and replay.
- Mask or exclude sensitive tables such as payroll and cost rates if the audience should not see them.
Ways to get it done
Build in-house
Effort: Roughly 6 to 12 weeks for a first read-only agent, longer with approvals UI.
You have a developer who knows your Made2Manage schema and want full control.
Use Aptean-provided or marketplace tooling
Effort: Days to weeks depending on licensing; confirm what exists for your release.
Your version has supported AI or integration features that cover your use case.
Work with a partner such as Netray
Effort: A scoped pilot is typically a few weeks, then iterate use case by use case.
You want an on-prem agent with approvals and do not have in-house AI experience.
Pitfalls
- !Pointing the model at production tables directly and discovering that a heavy query locks the shop during a shift.
- !Trusting column names without checking how your install uses them; custom fields are common.
- !Starting with write-back. Earn trust with read-only answers first.
- !Ignoring data quality: stale item masters and unreleased jobs produce confident but wrong summaries.
- !No evaluation set, so nobody can say whether the agent is improving.
Frequently asked questions
Does Made2Manage have built-in AI agents?
Not in the older on-prem versions most people still run. Aptean's newer products may offer AI features, but availability depends on your version and contract. For an existing install, the usual route is an external agent that reads the database and proposes actions for people to approve.
Can an agent write transactions into Made2Manage?
Technically possible, but risky. Direct database writes can bypass costing and inventory rules. Start read-only, queue proposals for humans, and only automate a write through a path your vendor supports and you have tested on a copy.
Which database does the agent read?
Most current installs use SQL Server, which is easy to query. Very old installs on FoxPro tables need an export or a different reader. Check your version first, and prefer a replica or nightly copy for the agent.
Do we need a GPU server?
Only if you want a private model. A modest workstation-class GPU can serve a mid-size open model for a small team. Otherwise a cloud model under enterprise terms works with the same tool layer, at the cost of data leaving your network.
How do we measure whether the agent is useful?
Replay a few dozen historical exceptions and compare its proposals to what your planners did. Track accepted versus edited versus rejected actions weekly. If planners keep rewriting the drafts, fix the tools or data before adding scope.
Want AI agents working on your Made2Manage?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment