Why industrial AI requires a data ops foundation to scale

When I talk to prospective industrial customers, the conversation goes one of two ways: “We don’t know where to start with AI,” or “Oh, we're already doing that.” The first group cannot figure out how to combine all their systems and projects, while the second has usually bolted AI features onto their enterprise resource planning system.

Both groups have missed that AI is not a single data point you can quickly pull. AI is a combination of data points you must combine to truly understand, say, predictive maintenance.

See also: Siemens entry aims to make agentic AI more affordable for SMBs

To build the right model for a particular machine, you need more than just its sensor data. You also need your ERP data, your manufacturing execution system data, and dozens of data points from different sources that the business generates every day.

You can try using disparate horizontal AI solutions, but if you can’t bring them all together into a verticalized AI solution, you won’t be able to scale beyond a handful of industrial experiments. You need all of it, in context, flowing through a common architecture. You need a data ops platform.

The hidden cost of inadequate infrastructure

Most companies think they first need to organize all their enterprise data before they can start with AI, but that usually only delays value for years. The most successful AI programs start with a focused business problem, such as reducing downtime, improving maintenance, or increasing operational efficiency, before connecting only the data they need for that use case.

There's always a starting point. Figure out your KPIs first: What are you really interested in and what will provide you the right ROI? What are the things that you have tried time and time again, and really need to make work?

Modern industrial AI platforms can incrementally ingest, contextualize, and unify data over time, so you don’t need a perfect data foundation on day one. In fact, AI itself can help accelerate your data organization and normalization, allowing you to deliver measurable ROI quickly and scale from there.

See also: Rockwell report: Days of ‘experimentation’ are over, DX is here to stay

When some customers, like a major yogurt manufacturer or an energy company, initially push back with stories about what they’ve already tried, we dig into how it’s working.

Typically, it’s not going anywhere because they don't have the right information or technologies. They’re in the business of making yogurt or generating power, after all, not optimizing sterilization-in-place and other cycles. That's where data ops experts come in.

What happens when you roll out AI on your own

The yogurt company I mentioned, before it gave up and hired a data ops vendor, tried to adopt AI on its own. I’ve seen manufacturers hire AI experts from the industry or from consulting firms that have worked in the same industry, gather all their resources, make the investments, and say, “Let's try to do it.”

Then they run into the same two major roadblocks. This is not specific to the yogurt company; I’ve seen energy companies, chipmakers, and all sorts of industrial customers experience the same issues.

The first roadblock is decentralized operations. The corporate oversight is there, but it's not as strong, resulting in mini projects across the board. Factory A is doing well on its AI adoption journey because it has made some investments.

Unfortunately, you cannot replicate the results across all the factories because the scenarios are different. Factory A’s team didn’t think through how its investment can benefit the organization at large because it’s only focused on optimizing its own factory for its own use cases. They built a POC that cannot scale.

A data ops platform doesn’t solve everything

Even when organizations previously failed to adopt AI or didn’t see the success they wanted, a data ops platform is not a silver bullet. We once ran a fairly successful POC with a customer

I say “fairly successful” because we raised the alarms where necessary, built models to examine certain operations, and created dashboards that showed our value, but the customers’ employees didn’t take action on our recommendations. An insight that doesn't trigger an action is just noise.

If the business operators don’t act on what the data ops platform surfaces, the POC isn’t fully successful; it worked, but the customer didn't see its full benefits, such as financial gains or overall equipment efficiency (OEE) improvements. It should have, but the expectation-versus-reality gap kicked in. That’s what you need to move beyond the POC and sign the enterprise deal.

See also: Integrating IT, OT, and AI for real-world competitiveness

The vendor can, of course, put together KPIs and make sure that everything is tracked, but you can’t overlook the human element. Your employees must be completely on board. Old habits die hard.

Remember: You still have the same people who witnessed the previous AI project fail. You suddenly have the right technology, but you cannot expect their minds to shift right away. It takes time. Throwing money at the problem is not enough.

When customers come to us, it's usually because assets are failing, they don't know why, and their margins are shrinking. On the asset side, the goal is straightforward: maximize OEE by slashing mean time to repair (MTTR) and mean time between repairs (MTBR).

On the process side, everything we track is about improving workflows. Why does a cleaning cycle take three hours after one batch and six hours after another, and what steps will standardize the process to eliminate the variance?

See also: Stories of AI adoption: Wolfspeed uses ‘three-stage maturity model’ for agent deployment

There are other KPIs, of course, but the key is identifying them, agreeing on their definition, and determining what success looks like at the end of, say, a three-month POC pilot. You also need to have the key watch points along the way so that at the end of the period, everyone can agree it was a success.

Successful AI deployments in industry take time and support

The fastest path to ROI is to start small and scale in phases. Most customers begin with a single high-value use case on one production line, asset group, or plant, such as predictive maintenance, downtime reduction, energy optimization, or operator efficiency. This allows a vendor to demonstrate measurable business value within weeks or a few months instead of waiting for a multiyear enterprise rollout.

Once the first use case proves the ROI, you can expand the same platform, data pipelines, and AI models incrementally across additional lines, assets, and plants. This creates a compounding effect where each new deployment is faster, cheaper, and delivers value more quickly because you’ve already established the data foundation, workflows, and operational knowledge.

To see complete change and hit your KPIs, the data ops vendor needs to continuously deliver, work with you, and train your workers to take the recommended actions. Technology gets you to the table, but execution keeps you there.