Using AI on Plant Data Without Losing Control: 12 Questions to Ask Any Vendor

Before connecting AI to your plant data, ask where it goes, who can access it, and how results can be checked. Download the free 12-question vendor checklist.

By Itanta Marketing · Published 2026-11-18
AI can help a manufacturing team build dashboards, investigate a problem or prepare a report. But before any tool connects to your plant data, there is a more important question: what happens to the information you give it? Process data is not just a collection of numbers. It can reveal recipes, cycle times, quality limits and the operating knowledge your team has developed over years. Logins, operator IDs and shift records may also include personal data. A useful demonstration does not answer who can see that information, where it is processed or what happens when you leave the vendor. Our free e-book, Using AI on plant data without losing control, gives managing directors, plant heads and IT leads 12 questions to ask before choosing an AI or analytics tool. Here is how to use them to make the vendor conversation more concrete. Download the free e-book: Using AI on plant data without losing control Start with where your data goes A statement such as “your data is secure in the cloud” is not enough. Ask the vendor to describe the actual flow, including storage, processing, backups and the AI step. If these basics are unclear, it is too early to connect production data. 1. Where is our data stored and processed? Storage and processing are not necessarily in the same place. A tool might store data in one region and send it somewhere else when it calls an AI model. Ask for a written list of locations and the companies responsible for each stage, including backups. Good sign: specific locations and providers, not just a broad regional label. Be careful if: the answer stops at “in the cloud” or names a region without explaining who operates the service. 2. Does raw process data leave our site? Ask exactly what crosses the boundary between your plant and the AI system. Is it a full copy of historian data, selected readings, aggregated values or tag information? The answer should explain why each type of information is needed. Good sign: a plain description of what leaves the site and why. Be careful if: full copies of historian or SCADA data are required simply “for the AI to work”. 3. Is our data used to train AI models? Using a model to complete a task is different from using your data to train a shared model. Your process knowledge should not become an unclear side effect of a subscription agreement. Good sign: a contractual commitment that your data is not used for training unless you opt in. Be careful if: the only answer is buried in general terms of service that can change without notice. Check whether you can trust the results A confident response is not the same as a correct one. Plant teams need to verify the evidence and the calculation behind a result before using it to support a decision. 4. Can every result be traced back to its source data? For a chart, figure or recommendation, ask the vendor to show the underlying tags, time range and source records. A finished output without that connection is difficult for an engineer or shift lead to check. Good sign: the result identifies the data it was built from. Be careful if: the tool produces polished charts or text with no way to inspect their source. 5. Can we see how a number was calculated? OEE, energy per part and downtime can each be calculated in different ways. The tool's definition needs to be visible so your team can compare it with the definition already used in the plant. Good sign: calculations can be inspected and checked against your own definitions. Be careful if: the method is described as proprietary and cannot be reviewed. 6. What happens when the AI is not sure? Missing readings, ambiguous tags and incomplete context should change how a tool responds. Ask what it does when the available data cannot support a reliable answer. Good sign: it identifies missing or unclear information and asks for clarification instead of guessing. Be careful if: every question gets a confident answer, even when the evidence is thin. Get security and contract answers in writing Data control also depends on access rules, contractual responsibilities and the ability to leave. Bring IT and the people responsible for privacy or contracts into this part of the evaluation. 7. Who can access our data, at the vendor and inside our company? A shift lead and a managing director may need different views. Ask how access follows those roles, how company accounts are used for sign-in and how access is recorded. Include vendor support staff in the discussion. Good sign: role-based access, company sign-in and a record of who accessed what. Be careful if: vendor staff have standing access to customer data simply “for support”. 8. Is a GDPR data processing agreement available? Machine data can be stored alongside operator names, logins or shift records. If GDPR applies and a vendor processes personal data on your behalf, review the need for an Article 28 data processing agreement with your privacy team. Good sign: an agreement is available that identifies sub-processors and their locations. Be careful if: privacy obligations are dismissed as irrelevant because the system contains “only machine data”, without checking what is actually included. 9. What happens to our data when the contract ends? Evaluate the exit before signing the entry agreement. Ask how you can export your history, which formats are supported and when remaining copies will be deleted. Get the commitments in writing. Good sign: full export in a standard format and confirmed deletion within an agreed time. Be careful if: export is unavailable, costs extra or deletion cannot be confirmed. Test the fit and the real cost A successful pilot should show more than an attractive screen. It should work with the systems you already have, solve a real problem and make the cost of the next step understandable. 10. Does it work with the systems we already have? Most plants already run historians, SCADA, MES or ERP systems. Ask which existing connections the vendor supports and what, if anything, must be moved or replaced. Good sign: the tool connects to your existing systems without requiring everything to move to a new platform. Be careful if: all data must first be migrated into the vendor's own system. 11. How do costs grow with more users, sites and data? A small pilot can look affordable while a second plant changes the economics. Ask for worked examples covering the next two or three realistic steps: additional users, another site or more data. Good sign: a clear pricing model with examples your team can budget against. Be careful if: charges depend on data volume or “usage” that is difficult to predict in advance. 12. Can we test it first, at small scale? Choose a real plant problem rather than a generic demonstration. Agree on the scope, success criteria and end date before starting, so both teams know what the pilot is meant to prove. Good sign: a focused trial or pilot on your own problem. Be careful if: a long contract is required before you can see the tool working with your data. Turn the questions into a vendor comparison Send the questions before the first demo, ask for written replies and compare vendors side by side. The e-book includes a printable comparison sheet for three vendors, with a place to mark whether each answer was clear and in writing. Use the sheet as a prompt for follow-up questions, not as a replacement for a technical or contractual review. An unanswered question about data movement should not disappear because the rest of the demonstration went well. As the guide puts it, a vendor that cannot give clear, written replies to questions 1 to 4 is not ready for your plant data. How Prompt to Insights approaches these questions The guide describes Prompt to Insights as an agentic AI platform for manufacturing: assign a task in plain language, and an agent such as the Dashboard Agent carries it out. It describes a hybrid architecture where your data stays in your environment and the AI sees tag information, alongside traceable results, root cause and a recommended action. Apply the same checklist to Itanta that you would to any other vendor. Ask for the architecture, access arrangements and contractual answers that apply to your deployment. Then use a focused trial to check a real task against your own source data. Download the free 12-question checklist Get the complete eight-page guide, including why each question matters, good signs, warning signs and the printable vendor comparison sheet. Share it with your plant, IT and management teams before the next vendor discussion. Download the free e-book: Using AI on plant data without losing control Try Prompt to Insights free for 15 days or book a 30-minute demo.

Tags: Manufacturing AI, Plant Data, Data Security, AI Vendor Evaluation, Data Traceability, Prompt to Insights

Turn Your Manufacturing Data into Decisions Instantly