AI agents are now so easy to build that almost anyone can create one, but easy access doesn't mean blanket permission. In this episode, Trish and Traci explore how agentic and generative AI can help chemical facilities by gathering data for risk assessments, spotting weak signals in process trends and freeing engineers from administrative work. They also cover the risks: hallucinated sources, rogue agents and company information leaking into public tools. Trish urges listeners to fix broken processes before automating them, keep risk assessments human-led and verify every output. The goal isn't to fear AI but to use it thoughtfully.

Transcript

(Edited for clarity)

Traci: Welcome to Processing Safety with Trish and Traci, where we share insights from past incidents to help prevent future ones. I'm Traci Purdum, editor in chief of Chemical Processing. And with me, as always, is the one and only Trish Kerin, director of Lead Like Kerin. If you haven't already, subscribe to this free, award-winning podcast on your favorite platform so you can continue learning with Trish and me in this series. Hey, Trish.

Trish: Hey, Traci.

Traci: How are you doing?

Trish: I'm doing pretty well. Last week I released a free e-book on my website about due diligence in safety, with practical tips and guidance for leaders in the workplace. It's really exciting to get that information into people's hands.

Traci: I took a look at that, and it's pretty compelling stuff. It's a good-sized e-book, too. I'll link to it in our show notes.

Trish: Awesome, thanks.

AI — Where To Draw The Line

Traci: In our conversation today, we're going to talk about how easy it is to create your own AI agents to do activities for you. It's almost too easy. But easy access doesn't always mean blanket permission. The real skill now is knowing where to draw the line. You and I have been talking about the use of AI in safety for some time. But AI is moving into nearly everything so quickly that this conversation is more urgent than it was a year ago. Can you talk a little about that?

Trish: Yes. We're certainly seeing much broader use of what's called agentic AI. That means creating agents, which are bots that go and do stuff for you. It's really interesting to see how that's evolving. It has the potential to remove some really low-value work so we can concentrate on higher-value work, and I think that's important. It's not about using agents to take people's jobs away. It's about using agents to take away the dull and monotonous tasks so we can do the more important things.

We also still have the continued spread of generative AI. That's where you ask ChatGPT or Claude, or whatever you use, a question, and it gives you an answer. So there are two fields really taking off.

One reason we're seeing far more agentic AI is vibe coding. People can now create these agents without knowing how to program anything. You vibe code it by telling it what you want it to do, and it does it for you. It's a really interesting progression.

But you're right, it does open us up to some very significant risks if we don't get it right. We see a lot of people vibe coding all sorts of apps now, which is great. But are they actually following the security standards we need to keep these things safe? That's another question. It's so easy to create these things now. It's very democratized at the moment, and everybody can do it. So we almost end up with an overwhelming tidal wave of agents coming at us.

Beware Of Risks

Traci: You touched on some of the value it adds. Let's talk more about where it could introduce risk, especially in chemical facilities.

Trish: If we think about generative AI first, we've known for a while that there are risks. We call it AI hallucination: Generative AI sometimes makes stuff up that's just not real, and we need to be very aware of that. You can ask it to scour the web for the answer to a question and give you all the references, and it will write whatever you want it to write. You can even train it to write in your own voice. There's some really powerful stuff you can do. But you need to make sure the information it collects is accurate and properly referenced. Just because it tells you it has a reference doesn't mean the reference is real. It's been known to make those up, too.

So hallucination isn't just getting some data wrong. It makes up a source to justify the data so it looks legitimate. If you're going to use AI to write anything, get it to give you the sources. Then you need to physically check every one of them to make sure it exists and is accurately represented.

Then there are the agentic AI risks. We create these agents to follow a process. The first thing to understand is that you can't fix a process problem with an AI agent. You need to fix the process. I think people are too eager to vibe code a solution when the real problem is that the process doesn't work. Simply putting an agent on it won't make the process any better. It will just codify the existing process issues.

So the first port of call is to go back and look at what you're trying to automate. Is it the right process? Does it have the right steps? Are they in the right order? Are the right people being involved at the right times? Map your process properly, optimize it, get it right and then build your agent to do that process. Then test it, test it and test it under all sorts of conditions to make sure the agent doesn't go rogue and do other things.

I believe it was announced earlier this week that there was the first known instance of agentic AI hacking a system. An agent had broken into another system and stripped data from it. It was an agent out of OpenAI, and it broke into the Australian government's Medicare database, which is our public health care system.

Traci: Wow.

Trish: That was only released publicly a couple of days ago. Right now we're all asking, "What data did they get?" It looks like they didn't get anything of significance, but that's not the point. An agent went rogue and hacked something, and it was a government agency, which is interesting as well. These are the things we need to be careful about. We can't have rogue agents running around the internet by themselves.

Traci: Think about the ramifications of this happening at a chemical facility. I remember years ago you and I talked about cyber issues, when a nefarious character broke into the wastewater system at a hotel or something. Everything floating around in my head right now is going, "Oh boy."

Trish: Yes. Now we have human hackers who could compromise cybersecurity at your facility, and we also have to worry about AI agents. These aren't human hackers. They're little programs running in the background that can replicate themselves. They can do all sorts of things if they aren't built with the right security in place. We're in a totally new world.

If you have a facility with high-hazard products and high-hazard processes, you need expert advice in cybersecurity. You also need expert advice in AI: how it works and what the opportunities are. I know a lot of companies are limiting the use of AI. You can't put company resources into any of the generative AI systems, because at some point that basically releases them onto the internet for everybody. So you need locked-down systems. Some organizations have built their own agents to run internal search engines. Those agents are restricted to the company's own servers and don't reach the broader internet.

I was talking with a company last week that has been piloting some generative AI work within its business. But it has very strict lockdowns on who can create agentic AI systems. A core team does it. If you need an agent, you put up your business case. Then the experts build it with all the security features needed to keep it safe and to protect your company's intellectual property.

Because, as I said, the moment you put something into generative AI, you've educated it and fed it information. You can claim copyright on the output. But once you put something in there and ask it to do something with that information, it's out there. You no longer have control of the original information.

Traci: It used to be thumb drives. Now look what happens.

Trish: Yes.

AI Do's And Don'ts

Traci: Let's dial it back from the whole Terminator, John Connor doomsday predictions. Let's talk about some dos and don'ts, and the best ways to use the AI we have to help us do our jobs better. One of them is don't hand the actual risk assessment over to AI. AI is really good at helping you prepare for it, but what should facilities do?

Trish: Absolutely. You can responsibly use AI tools to prepare for your risk assessment. Say you have an internal agent that's authorized by your company and searches only your internal systems. You can have it collect all sorts of relevant information, like incident history, trending data and maintenance history, and bring it together. You could even use generative AI to search the internet for similar incidents around the world. Again, check that they're real and relevant. AI can scour all sorts of databases and pull this information together much faster than a human can. So there's real value in using AI responsibly to build your input data.

But the risk assessment itself, actually sitting down and having the conversation, is different. Take a hazard and operability study, for example. You have the core expert team in the room, discussing each node and each guide word and working through the scenarios. That's where humans are far better than AI. We've seen it, and we have experience and knowledge. We naturally connect two pieces of information that we might never have told the AI to connect, because we've experienced something similar.

So having that conversation, with great diversity of thought and different subject matter experts coming together, is a really important part of a risk assessment. A risk assessment is a team-based activity. It's not something you can do sitting by yourself, and it's not something you can get AI to do for you. It's a workshop activity. It involves discussion, debate and interaction to reach the best possible consensus decision on that day.

Traci: It's the humans who have the creativity and insight to work through that. AI is the diligence partner. You can have it scour 20, 40 or 100 documents and pull things out. That's where we can use it best, and then on the human side, we apply what we know best. With that in mind, what kinds of trends or signals are agents actually good at catching that humans tend to miss or catch too late?

Trish: They can catch a lot of the very subtle, weak signals, and I love to talk about weak signals. We can program agents to notice slight deviations in a trend. We wouldn't necessarily notice those deviations because we aren't seeing the larger set of data.

For example, say we're tracking a parameter with an upper limit and a lower limit. Its history shows it's almost always smack in the middle of those limits. Give or take a little, it's very consistent around the center line. Then, all of a sudden, one or two data points start to drift away from that center line. We might not notice, because we're only concerned with keeping it within the upper and lower bounds. But AI can notice that the result is slowly trending down toward the lower bound or up toward the upper bound.

We can also have it look at other trends at the same time, and this is where I think it has great value. Sometimes another parameter starts drifting at the same moment the first one does. We wouldn't necessarily spot that, certainly not in a reasonable time frame. AI lets us start to see correlation. Now, correlation is not causation, but correlation ought to be investigated for causation, because it might be causal and it might not be.

At our upper or lower bound, we've got a real problem. AI can detect those faint, weak signals well before we get anywhere near that point. It can tell us we're heading toward a problem while we still have time to fix it. That is incredibly powerful in a processing plant. AI flags that it's seeing a slight trend change. Then you ask it to look for any other slight trend changes happening at the same time. That's a really valuable use.

Traci: What are some other do's and don'ts we should be considering?

Trish: I've said it already today: Check every piece of data AI gives you. Go over it with a fine-tooth comb. I use generative AI for simple things. Say you've written something and need a new version of it for something else. It's really hard to rewrite your own work, and AI is really useful for that. I get it to write another copy in my voice, in a different style or with different information. But I still read every word and edit the entire thing, because you might miss something or get something wrong. You want to make sure it's doing the job properly.

If you're going to ask it to pull data from some information, first make sure that information is accurate. In the very early days of computer programming, there was a saying: garbage in, garbage out. We still need to remember that with AI. AI can't take garbage in and make it beautiful coming out. If it was garbage going in, it's still garbage coming out. Sometimes we need to go back to basics. Is the information we're putting in right? Check that before we ask it to do anything.

Then read what comes out of it, and test it, challenge it and verify it. Those are the really key dos: Make sure the right information goes in, and check what comes out. Don't rely on it solely. Use your human intellect, because you're right: We bring the creativity, the critical thinking and the experience. AI can read data, but it doesn't have human experience. We need to bring that to the table.

I think we need to not be unnecessarily afraid of AI, but we do need to treat it with respect and be thoughtful in its use. It's not for everything. Occasionally I use generative AI to create an image I need for a presentation. But continually generating cool pictures with AI uses enormous quantities of energy and water. We don't see that drain on our resources, because it happens pretty quickly, so we don't notice it. But it uses enormous quantities of our natural resources.

Is it worth it? If you had to watch enormous quantities of water run down the drain to produce that one cute cartoon image, would you still want it? That's something to keep in the back of our minds. You can get free access to these AI tools, but they're not free. We're paying for them with the environment and our resources.

The Cost Of AI

Traci: One of our frequent contributors did a comparison of what these data centers consume. They're competing for the same things chemical facilities compete for: water, resources and energy. So it's interesting to think about it from that perspective. You won't only be dealing with AI in your own operations. You'll also be competing with it for resources within the community. Things to think about, for sure.

What can be done with the time that gets freed up by using AI properly?

AI Frees Time

Trish: Oh, so many things. That's an endless list. But in all seriousness, think about people who are tied up doing data processing as part of their normal job. I started work in the chemical sector last century. I'm that old. When I started, we were seeing administrative roles removed. It was too expensive to have administrative staff do admin work, so the admin tasks were all given to the engineers. That really doesn't make a lot of sense at engineering hourly rates compared with admin hourly rates. But we went through that evolution, and we're well and truly down that track now.

I know from personal experience that engineers spend inordinate amounts of time processing expense claims. Imagine an agent that tracked my expenses and filled out the form for me to verify. I must verify it, but then I sign off. That would be a much more effective use of my time than spending hours processing my expenses. That's a trivial example. But it could free me to look at plant trends I haven't had a chance to look at because I've got too many administrative bits to do.

Then I get the opportunity to study what the agents are telling me about different trends. I get to apply my critical and creative thinking and experience to it and say, "Oh, I think we could be having a problem with this or that." And then I get to go investigate it.

So it's not about removing the role of the engineer. It's about removing the admin tasks so the engineer can do meaningful, useful work. That work helps optimize our process or makes sure it's still running safely, with no potential incident precursors we just haven't noticed yet. It could even help us recover our process if something does start to go wrong.

Everybody has something in their role they really want to do because they know it's important. But it's not urgent, so they never get to it. Using AI to take off the administrative burden allows us to do those important but nonurgent tasks.

Traci: Absolutely. I've used it for many things on my side, and it does free up time. It helps me organize my emails and figure out what I need to tackle first, which is a brilliant use. That frees up time for the other things I need to do, like edit your column.

Trish: Yes, exactly. That reminds me, I've got one due shortly, haven't I?

Traci: Yes, you do. That was my little nudge.

Trish: Better get writing.

Traci: Anything you want to add on this topic?

Trish: As I said earlier, we need to not be afraid of the AI world that's coming. We need to engage with it intelligently and thoughtfully and make sure we use it responsibly. There's a lot of potential power here, and there's also a lot of potential for misuse, so we need to do the right thing with it. But we don't need to be scared. We need to move forward, thoughtfully.

Traci: Well, you always help us move forward thoughtfully and safely. And thank you for reminding me of GIGO, garbage in, garbage out. I haven't heard that since my college days, back in the last century as well. Unfortunate events happen all over the world, and we will be here to discuss and learn from them. Subscribe to this free podcast so you can stay on top of best practices. You can also visit us at Chemical Processing for more tools and resources aimed at helping you run efficient and safe facilities. On behalf of Trish, I'm Traci, and this is Processing Safety with Trish and Traci. Thanks, Trish.

Trish: Stay safe.