Old ERP, New Tricks: Putting AI on Top of Software You Can't Replace

Yes, you can add AI to software you can't replace. AI for legacy ERP works by sitting in front of the old system instead of inside it. It reads your data through a read-only connection, answers plain-English questions, builds reports that used to need the vendor, and runs routine exports. The ERP stays your system of record, and nothing should write back to it until a person approves the change. The safe order is simple: read-only questions first, automation one task at a time, and a security review before anything connects.

That is the idea behind "Old ERP, New Tricks," a Builder Track session at the AI Business Retreat in Spooner, Wisconsin, on September 15, 2026. Here is what the session covered and how to apply the same approach to your own shop.

What the Old ERP session covered#

The retreat's page for Old ERP, New Tricks: Let AI Upgrade Your Line-of-Business Software lists it in the Builder Track, the retreat's track for hands-on builders, from 1:00 to 2:20 p.m. at The Island in Spooner. Tim Holum taught it. He has been Coulee Tech's software developer since 2008, and his work includes custom ERP software for small manufacturers and the servers underneath it.

It lists three parts:

Plain-English answers pulled live out of a "crusty demo ERP."
Automated exports that run without the software vendor's involvement.
The recipe for putting an AI layer in front of any legacy system.

Its stated takeaway is an ask-your-ERP recipe. The page describes the audience as the people who run a 20-year-old ERP or line-of-business system every day, plus the managers who keep hearing that a report "isn't possible" without a five-figure vendor quote. For the rest of the track, from the morning coding bootcamp to wiring AI into live business data, see our walk-through of the Builder Track.

Why put AI on top of old software instead of replacing it?#

An older ERP holds the things that are hardest to rebuild: routings, bills of material, years of job costing history and the habits of everyone who uses it. Replacing it means a data migration, retraining and a cutover that can stall shipping. For many small manufacturers and distributors, the system works. The problem is getting answers out of it.

The report you want needs the vendor, a consultant who knows the reporting tool, or the one employee who understands the tables. So the question waits, or someone rebuilds the numbers in Excel every Monday. An AI layer goes straight at that bottleneck while the ERP keeps doing its job. If you do replace the ERP someday, what you learned (which questions matter, what each field means, how to check an answer) carries over.

How AI for legacy ERP works, step by step#

The session page names the recipe without spelling out every step. In practice, putting AI on top of old software follows a pattern like this one.

1. Find the door into the data#

Your ERP's data usually lives somewhere an outside tool can read it. Often that is an ordinary database: ECI's system requirements for JobBOSS releases 2018 through 2021 (opens in a new tab), for example, list Microsoft SQL Server. Other systems offer an API or scheduled file exports, and any of the three can feed an AI layer. Our JobBOSS² IT support and Epicor ERP support both cover the SQL Server underneath the application, which is where this work starts. Pick a door your vendor supports that does not slow the shop floor, often a nightly reporting copy rather than the live database.

2. Write down what the data means#

Legacy field names were written for programmers, not owners. A one-page data dictionary fixes that: which table holds open jobs, which field is the promised ship date, what each status code means. AI writes better queries with that page in hand, and so do new employees.

3. Connect through a short list of tools#

The Model Context Protocol (MCP) (opens in a new tab) is an open-source standard for connecting AI applications to external systems such as databases and tools. A connector built on MCP, or on a plain API, can offer the AI a few named actions, such as "find late jobs" or "get job cost by job number," instead of the keys to the whole database.

4. Start with your real questions#

List 5 to 10 questions people ask every week that take too long to answer. Those become your first tests.

5. Check the answers before anyone relies on them#

Have the AI show the records or the query behind each answer, and compare its first answers with a report you already trust. Once they match for a few weeks, people can lean on it for new questions.

Read-only first: reports and questions over ERP data#

Read-only is where to start. It answers many day-to-day questions, and it cannot change a job, a price or an inventory count. Picture a 30-person job shop. Good first questions look like these:

Which open jobs are past their promised ship date, and which operation is each one waiting on?
On last month's closed jobs, how did quoted hours compare with actual labor hours?
Which customers ordered less this quarter than in the same quarter last year?
Which parts are below their reorder point with no open purchase order?
What is sitting in work in process right now, by work center?

Then pick the one report you were told would need custom work and make it the first scheduled output. Often the data for that report is already in the system. What is missing is an easy way to ask for it.

Where automation fits#

Once the answers check out, automation is the next step, one rung at a time.

Level	What the AI does	Example	Human check
Ask	Answers questions from read-only data	"Which jobs ship late this week?"	You verify early answers against known reports
Report	Builds and sends scheduled summaries or exports	A Monday late-jobs email, a month-end file for your accountant	The report's owner reviews the first runs
Draft	Prepares a change for a person to make	A draft purchase order for parts below reorder point	A buyer reviews it and enters it
Act	Writes to the ERP through one narrow, approved tool	Updating a job's status after a shipment is scanned	A person approves each action, and every action is logged

You may never need the fourth row. If you do add it, add it last and keep it narrow. The OWASP Top 10 for Large Language Model Applications lists Excessive Agency (opens in a new tab) as a risk and recommends limiting an AI tool's permissions to the minimum necessary and requiring a human to approve high-impact actions. When a connector or export needs to be built rather than bought, that is custom software and automation work.

Security review before connecting AI to your ERP#

Run through this list before any AI tool touches ERP data. It is general guidance, not legal advice.

Use a dedicated read-only account. Never the admin login or a shared password. On SQL Server, the db_datareader role can read all data from all user tables and views (opens in a new tab), which is more than an AI tool needs. Grant access only to the specific views your questions require.
Leave out what the AI should never see. Payroll, employee Social Security numbers, bank details and anything under a customer confidentiality agreement stay outside the views you share.
Know where the data goes. Confirm which AI service processes the data, whether it keeps it, and whether your terms allow training on it. Use a business account, not a personal login.
Plan for hostile text. ERP records hold text typed by customers and suppliers, such as order notes and attached files. OWASP describes indirect prompt injection (opens in a new tab) as an AI accepting input from external sources, such as websites or files, that alters its behavior. Read-only access limits what that text can make the AI do.
Check the platform underneath. Microsoft's lifecycle pages show extended support ended for SQL Server 2014 in July 2024 (opens in a new tab) and for SQL Server 2016 in July 2026 (opens in a new tab). If your ERP database runs on either one, plan that upgrade before you open a new connection into it.
Protect the shop floor. Point heavy questions at a reporting copy or run them after hours, so a big query does not slow the people entering time and shipments.
Ask your vendor about access. Confirm with your ERP vendor or reseller how outside tools may read your data under your support agreement.
Settle controlled data first. If your ERP holds defense information covered by CMMC, decide how that data is handled before any AI tool sees it. Our guide to CMMC for Wisconsin manufacturers explains the levels.
Keep a log. Record who asked what, which query ran and what came back, so mistakes are traceable.

Frequently asked questions#

Can you add AI to an old ERP without replacing it?#

Yes. An AI layer reads the ERP's data through a read-only database connection, an API or scheduled exports, then answers questions and builds reports outside the ERP. The ERP stays the system of record, and replacing it becomes a separate decision on your own schedule.

Is it safe to connect AI to our ERP database?#

It can be, when the connection is read-only, limited to the views your questions need, and run under a dedicated account with logging. Sensitive data such as payroll should stay out of scope. Anything that writes to the ERP should need a person's approval first.

Does AI work with JobBOSS or Epicor?#

The approach does not depend on the brand. It needs a supported way to read the data, which for many systems means a SQL Server database, an API or a scheduled export. JobBOSS releases 2018 through 2021, for example, run on Microsoft SQL Server according to ECI's system requirements. Which route fits depends on your version and support agreement, so confirm it with your vendor or IT partner first.

Do we need a data warehouse before adding AI to our ERP?#

Not to start. A read-only connection to a reporting copy of the ERP database, or a scheduled export, is enough for the first questions and reports. A warehouse starts to make sense later, when you want to combine the ERP with other systems such as accounting or shop floor data.

How accurate are AI answers from ERP data?#

They are only as accurate as the data and the query behind them. Ask the tool to show the records or query it used, and compare early answers with reports you already trust. Cryptic field names and inconsistent status codes are common causes of wrong answers, so write the data dictionary first.

Talk to us about AI for your ERP#

If your ERP is paid for and full of answers nobody can get out, start with AI for manufacturing from Coulee Tech, which turns plain-English questions into live dashboards and scheduled reports from systems like Epicor and JobBOSS. Bring the system and version you run and the one report you have been told isn't possible. Our manufacturing IT team supports the servers and networks underneath them, with local engineers across the Coulee Region and the Chippewa Valley.