Legacy modernization

How to Add AI to a Legacy MES or ERP Without Replacing It

The short answer

Most plants that ask about AI already run an MES, an ERP, a LIMS and a set of spreadsheets that hold the real exceptions. Those systems work. The problem is that the knowledge needed to answer everyday questions is scattered across screens, stored procedures, database views and the memory of a few experienced people.

Replacing the stack to "make it AI-ready" is slow and risky. A better first step is narrower: choose one decision, understand exactly how today's system supports it, and give an AI assistant a reliable, explicit foundation for that decision only.

Step 1: Choose one decision, not a platform

Start from a question someone asks every week, for example:

  • Why is this batch still waiting for release?
  • Which orders are at risk because a material lot is on hold?
  • Which deviations look like ones we have already investigated?

A good first decision has a named owner, a visible cost when it is slow or wrong, and data that already exists somewhere. Write down who decides today, what they look at and what "good enough" would mean for an assistant.

Step 2: Trace the decision through the existing system

Follow the business action from the screen the user sees to the code and data behind it:

Layer What to record
Screen or report Which fields the person reads, which filters are applied by default
Service or stored procedure Status logic, exceptions, hard-coded plant or product rules
Database views Joins, filters, time windows and units
Source tables The full population, keys and history
People Unwritten rules, workarounds, known data problems

This is where most surprises appear. In our public industrial demonstration, a convenient queue shows 8 batches while the full register holds 40. An assistant that reads only the queue would give a fluent but incomplete answer. You can watch that step in the industrial walkthrough.

Step 3: Make sources and meanings explicit

Before any model is involved, agree on:

  • The authoritative source for each fact, and why it was chosen over alternatives.
  • Business meanings, such as what "released", "on hold" or "available" mean at each site.
  • Ownership and versions, so a changed rule is visible rather than silently absorbed into a prompt.

Keeping these decisions in a small, versioned knowledge asset (a schema, a set of rules, a graph of production objects) is what makes the result maintainable. If you later change the model or vendor, this asset stays. We explain the idea in more depth in why a digital expert asset is more than an agent.

Step 4: Pilot read-only assistance first

A sensible first pilot reads from the existing system and prepares an answer for a person to approve:

  • It cites which records and rule versions it used.
  • It states what is missing or conflicting instead of guessing.
  • It never writes to the MES or ERP and never touches control systems.

Define a small set of evaluation cases from real history (de-identified where needed): cases where the right answer is known, cases with missing data and cases where experts disagree. Run the assistant against them before anyone relies on it.

Step 5: Decide what comes next with evidence

After the pilot, you should be able to answer three questions with records rather than impressions:

  1. Did the assistant find the right sources and state gaps honestly?
  2. Did it save the owner time on the agreed cases?
  3. What would deeper integration require: new interfaces, data quality fixes or a partial modernization?

Only then is it worth discussing write-back through the system's own validated interfaces, or a phased modernization of the underlying data.

Common mistakes to avoid

  • Starting with a chatbot over all documents. It feels productive but hides which answers are grounded.
  • Treating a view as the truth. Views often carry filters that were right for one report and wrong for a new question.
  • Putting business rules in prompts. Rules become invisible, untested and tied to one model.
  • Giving the pilot write access. It turns an evaluation into a change-control and safety project.

What this costs and how long it takes

For one workflow, a discovery that produces the source map, knowledge model and pilot proposal usually takes one to two weeks. A focused read-only pilot usually takes three to six weeks. The main variables are the number of systems involved, the quality of existing documentation and how quickly the system owner can confirm meanings.

If you want to test this approach on one of your workflows, describe it to us and we will tell you honestly whether it is a good candidate.

Questions & answers

Do we need to migrate our MES or ERP before adding AI?

No. Most first pilots run beside the existing system with read-only access to its database views, exports or APIs. Migration is a separate decision that discovery can inform.

Which workflow should we start with?

Choose one recurring decision with a clear owner and a measurable delay or error cost, such as batch release, deviation triage or material shortage checks. Avoid starting with closed-loop control.

How long does a first discovery take?

For one workflow, typically one to two weeks, depending on how many systems it touches and how quickly read-only access and the system owner's time are available.

Can the AI write back into the MES?

Only after a read-only pilot has been evaluated, and then through the system's own validated interfaces with explicit approval steps. Control, interlocks and safety systems stay outside the AI's authority.

LET’S START A CONVERSATION

Let’s start with your workflow.

Tell us what you want to improve and which systems you use. We will reply within one business day with questions, a view on fit and a practical next step.

Discuss a project