Can you trust your data enough to use AI?

Fifteen years ago, utilities used data to make decisions, but anyone could still open a spreadsheet and adjust the numbers until they told the story they wanted. Adam Chartrand thinks the same problem has come back with AI, at a bigger scale. When he scrolls social media, he catches himself unsure what's real and what's generated. So in regulated industries, the first question now is how you know you can trust the data before you hand it to AI, not what AI can do.

After a summer break, Adam and Jared Westrem come back to the mic in sweaters, with a new Tap Into Data neon sign in the office, to talk it through.

The skepticism is reasonable

Adam has spent the last couple of weeks visiting organizations in regulated industries. One was a former Crown Corporation that doesn't fully trust the AI side of inspection yet. The conversations covered docks, remote deployment, full autonomy, and thermal imagery, and the questions kept coming back to the same thing: what should we believe, and how do we explain it to the rest of the company? On the security side, someone asked what happens if a sensor trips on its own and launches a drone.

He has seen this pattern before. When he and Jared first met in wind, inspection was a quantity-versus-quality contest. Machine learning arrived, and the industry tried to double its output, but the models flagged defects incorrectly. People had to step back in before the technology could catch up.

Serve the data up well, and people can review it

Jared came back from Commercial UAV Expo in Las Vegas, where he presented on dock workflows, automation, and scalability with Zach from Xcel Energy (Zachary Roach, a past guest). His early experience was similar to Adam's. He worked with a third party, a former DoD contractor, on one of the first at-scale machine learning deployments in wind.

His view now is that the industry often overstates the need for AI review. If workflows are set up properly and the data is easy to click through, the right experts can review it efficiently. That might be a wind technician, a blade engineer, a solar reliability engineer, or a distribution or transmission lineman. AI can then help with things like catching what a tired reviewer missed. In Jared's words, "It's a tool that adds productivity, adds functionality, adds resilience, but it's not the sole purpose."

When the infrastructure is the power grid, he doesn't see anyone switching off human review any time soon.

Bad data in, bad results out

Adam has watched a client switch suppliers because the new one promised AI could do the work. His concern is that the supplier doesn't have the expertise or a setup that can scale. You can use AI at the core of your business, he says, but you still need people to check that the data is accurate. You also need permissions that control who can change what.

Jared has seen how quickly models go wrong. Training needs large sample sizes and consistent, accurate labeling. Put one reviewer in the loop who doesn't hold to the same standard, and a model detecting blade damage, solar panel defects, or insulator issues can start getting worse instead of better.

Don't start AI-first

This is the episode's contrarian point. The popular approach is to have an idea, give it to AI, and let it do the work. Jared mentions an upcoming 72-hour experiment where a bot is supposed to build an entire company and product, and adds, "We'll see how that goes."

He suggests the reverse. Start by capturing every data point in the end-to-end workflow: what the flights look like, what reviewers do, what counts as good and bad imagery, and what counts as good and bad damage calls. Then add tools one step at a time to remove friction. Utilities like Xcel are capturing hundreds of thousands of structures and millions of images. If those images are tied to the asset and reviewed by a professional lineman, you can start with strong detection models from day one.

Adam's version is to set a solid foundation before relying on AI. He runs an inspection a few times, checks that the results are right, and only then asks AI how to do it faster. He also uses AI to compare a new data set with the last one and this month's P&L with last month's. As he puts it, AI is "a mechanism that supports what you're doing, not the other way around."

The bottleneck after the data comes in

Jared raises the question he thinks most organizations haven't answered: then what? Back in wind, a site inspection report led straight to deciding what to fix first given budget, constraints, and how many technicians were available. Today, data can arrive within a day. Then it stalls in a manual process of talking to four people and following a procedure, and the cycle repeats next year. He sees that internal action step as the place where automation and AI are needed most.

Adam describes a client task that used to take 16 to 18 hours and now takes about an hour and a half. They still checked the new results against the old boots-on-the-ground method. Crawl before you walk, he says. Even when a drone launches, flies its mission, and lands, someone in a command center is still doing QA/QC on what it brought back.

He also warns about large enterprise AI training rollouts. A plan to train 10,000 people in a year often doesn't add up once you count the available hours. By the time the training reaches the calendar, the AI tools it covers have changed. Gurmeet Mangat's more focused programs, built around how people actually work, come up as the alternative.

The episode ends with a commitment to weekly Friday recordings in Q4, a call for guests, and the story of Papa Pete, the friend who first said they should do a podcast.

Press play above to hear the full conversation.

Key takeaways

• Regulated utilities still want a human expert reviewing every image, and on critical grid infrastructure that won't change soon.
• If inspection data is served up in an easy-to-review format, experts can review it efficiently, and AI can help them without being the only reviewer.
• One inconsistent reviewer in the training loop can quickly make a detection model worse.
• Capture every data point in the workflow first and add AI step by step, instead of starting AI-first.
• The real bottleneck is often turning fast, usable data into action inside the organization.