When an AI pilot in a plant runs late, the cause is rarely the model. It is usually something more ordinary: nobody can get read access to the historian, asset IDs in the CMMS do not match the tags in the PLC, or half the complaint records are scanned PDFs. All of these can be found in advance.
This checklist covers what to look at before a pilot starts. You do not need perfect data. You need to know where the gaps are, and to choose a first project that avoids the worst of them.
1. Inventory the systems
List every system that holds data about the problem you want to solve, and for each one note who owns it and how it can be read.
- ERP: orders, materials, inventory.
- MES or SCADA: production events, run times, changeovers, line states.
- CMMS: work orders, failure history, spare parts.
- Historian: time-series sensor data from PLCs.
- Time clock and HR: schedules, punches, rosters.
- Quality systems: batch records, QC results, complaint logs, training records.
- Email and documents: where complaints and procedures actually arrive.
For each, the useful question is how data can be read: an API, a read-only database connection, scheduled exports, or only by hand. "Only by hand" is not a dead end, but it changes the plan.
2. Check how much history there is
Models learn from the past, so the length and continuity of history matters. As rough guides for the modules on this site: predictive maintenance and changeover optimization typically need six months or more of history, and complaint triage needs at least 90 days of batch records with lot traceability.
Look for gaps as well as length. A historian that lost three months during a server move, or a CMMS that was only adopted on one line last year, is common and worth knowing about early.
3. Check that identifiers line up
This is the check that most often surprises teams. Data from different systems can only be joined if the identifiers match:
- Asset IDs in the CMMS against tag names in the historian.
- Lot numbers on a complaint against batch records in production.
- SKU codes in the MES against the ERP.
- Employee IDs in the schedule against the time clock.
Take ten real examples and try to follow each one across systems by hand. If you cannot, a model cannot either.
4. Check timestamps and time zones
Sequences matter in manufacturing data: did the temperature drop before or after the seal failure? Confirm that system clocks are synchronized, that you know which time zone each system uses, how shift boundaries are recorded, and how daylight saving changes appear in the data.
5. Check the quality of what is there
- Sensor dropouts: flat lines or zeros where a sensor went offline.
- Free-text fields: failure codes recorded as "fixed it" or "see notes".
- Closed-but-not-done records: work orders closed in bulk at the end of a month.
- Paper and scans: records that exist only as images.
None of these rules out a pilot. They decide how much cleanup comes first, and which problem to start with.
6. Decide access and ownership
Agree, before any connection is made, what the automation may read and what it may write. Reading sensor data is very different from creating work orders or sending emails. Decide who approves access, which service accounts are used, where data is processed and stored, how long it is kept, and what happens to it when the engagement ends. Put it in the agreement.
7. Choose a first pilot where the data is ready
The best first pilot is not always the biggest problem. It is the biggest problem whose data is good enough today. Minimum starting points for each module:
| Module | Minimum data to start |
|---|---|
| Predictive maintenance | Networked PLCs, a historian logging sensor data, CMMS failure history, 6+ months of sensor history |
| Staffing automation | An electronic time clock with an API, a roster, supervisor phones for notifications |
| Visual quality inspection | Cameras at inspection points, consistent lighting, 500-1,000 labeled defect images |
| Changeover optimization | Changeover start, end and SKU pairs in the MES or SCADA, 6+ months of history |
| Training compliance | Current training records in any format, certifications by role, an SOP library |
| Complaint triage | Lot traceability, 90+ days of batch records, email access for complaint intake |
| Production scheduling | An ERP or MES with a scheduling module, documented line rates and a changeover matrix |
The NuruState assessment runs these checks on site and ends with a pilot recommendation that takes data readiness into account, so the first pilot starts where it can succeed.