Quick answer: An AI pilot needs data your ERP or MES can expose on a known schedule, through a named system of record and a supported API or export, not an ad hoc report built by hand. If you cannot name the system, the update frequency, and the person who owns write-back, the pilot is not ready to start.
A team that gets an AI pilot approved in a budget meeting has answered a funding question. It has not answered an engineering one. The engineering question is whether the ERP exposes the data the model needs, on a schedule the model can use, through a connection that does not require a human to copy and paste. That question decides whether the pilot runs on schedule or stalls on integration work no one scoped. Ask it in week one, not after the vendor contract is signed.
Data Access Is Not a Yes/No Question
“Do we have the data?” is the wrong question. The right question is whether the team can pull the specific data this process needs within a week, without a data project first. That means naming where it lives, who owns it, and whether it is complete and current enough for this one process. A field that exists but is sparsely or inconsistently populated is not data access. A report that requires a formal request and a wait before it arrives is not data access either.
Name the table, the module, or the export. Name the owner of that data, not the department. If the answer is “somewhere in the ERP, probably finance has it,” stop and find out before the kickoff call, not during it.
Systems & Integration: Name the Target
Point four of the readiness checklist asks a narrower question: do you know which systems the AI output must write back to? Not which systems it reads from. Which systems it writes to. An AI system that flags a quality exception is only useful if that flag lands inside the MES screen an operator already watches, or the ERP workflow a planner already works from. Name the system. Confirm it exposes an API or a supported integration path, not a login screen a human has to read.
If the plan is for staff to check a separate dashboard or chat window and manually re-enter what it tells them into the ERP, that is not integration. That is a second job no one budgeted for, and the first step a busy plant can drop.
Who Owns Write-Back
Integration work needs an owner on both sides: a named business owner with time set aside each week, and a technical counterpart who can speak to the ERP or MES configuration directly. One name. Not “IT” generally, not “the plant manager’s team” generally. A person who can answer “can this system write back here” without asking three other people first.
Reject any pilot plan that lists “integration” as a single line item with no owner attached. That line item is where timelines break down, because it is the first place the plan meets a system that may not be maintained closely day to day.
The Tuesday CSV Problem
Here is the question that exposes the real state of readiness: does this system talk to anything, or does someone export a file by hand on a fixed schedule? A team that runs on a manual export can still pilot something, but it needs to say so out loud, and it needs to name what breaks the model’s usefulness: a missed export, a format change, a person out of office.
A manual export is not disqualifying on its own. A manual export no one has named as a dependency is. When a manual export is the actual system of record, write that down as the integration method, set the update frequency accordingly, and scope pilot expectations to match that data cycle. A pilot built around real-time alerts, running on a weekly export, is scoped wrong before a single model runs.
What ERP Data Readiness Looks Like, and the Red Flags
Ready on data access means you know where the data lives, who owns it, and that it is reasonably complete and current for this one process. The red flag is data that lives in PDFs, inboxes and personal spreadsheets, or a situation where no one knows who can grant access.
Ready on systems and integration means the target systems, such as the ERP or the MES, are named and expose an API or a supported integration path. The red flag is an AI that lives in a separate chat window that staff must copy and paste from.
Score each point 0, 1 or 2: zero if it is not true or you do not know, one if it is partly true, two if it is clearly true with evidence. A zero on data access is a blocker on its own.
Run the Checklist Before the Kickoff Call
Before anyone schedules a vendor demo, get the operations lead and the IT lead in the same room and answer these questions plainly: What is the named system of record for this process? What is the update frequency: real time, daily batch, or a manual export on a fixed cycle? Does the target system expose an API, or does someone need to build one? Who is the one accountable owner on the business side, and who is their technical counterpart?
Score honestly. Any zero on data access or systems and integration is a blocker on its own, not a risk to manage in parallel with building the model. Reject a pilot proposal that moves to model selection without naming the system of record. Set the pass/fail threshold for the pilot before the first sprint starts, not after the demo looks good.
The modeling work is not where a manufacturing AI pilot meets its first hard question. The integration question is. Answer it before the pilot is approved, not after.
If you want a structured way to walk through this with your operations lead before the next pilot gets greenlit, use the AI Readiness Checklist. It scores the same ten points covered here, including data access and systems integration, so you can see exactly where your ERP and MES stand before anyone writes a line of model code.
AI Readiness Checklist (10 points)
FAQ
What does ‘data access’ mean for an ERP-based AI pilot?
It means the team can pull the exact data this one process needs within a week, without a separate data project first. You can name where the data lives, who owns it, and it is current and complete enough for the process in question. A field that exists but is poorly populated, or a report that requires a formal request and a wait, does not count as data access.
Is a manual CSV export good enough for an AI pilot?
It can work, but only if it is named as the integration method up front and the pilot’s expectations match that cycle. A weekly export supports a weekly-cycle pilot, not a real-time one. A team relying on a manual export should also name who builds it and what happens the week that person is unavailable.
Who should own the ERP or MES data question before a pilot starts?
One named business owner with time set aside weekly, paired with a technical counterpart who knows the ERP or MES configuration directly. Data access and systems integration are owned jointly between operations and IT; security and compliance sign-off sits with whoever owns that function, not with the vendor.
Comments are closed