Responsible Use of AI in SAFe Organizations

0:00So welcome to skill number 10. This is the last skill before we get to our final exam. I'm really glad you've continued to train with me. This one is all about the responsible use of AI in safe organizations. Now AI has rapidly become part of modern product development. Teams can use it to generate ideas, summarize decisions, write code, analyze data. The possibilities are endless and used correctly it can dramatically accelerate learning and delivery. It can dramatically accelerate value delivery. It's amazing, but used incorrectly it undermines the very principles that makes agile and safe effective. For example, teams can lose ownership, conversations can disappear, decisions become centralized again. In safe, the safe framework, AI should enhance collaboration and learning, not replace them. The goal is not to avoid AI because it is an amazing tool and I love the opportunities it has given me. The goal is to use it responsibly so it strengthens agility rather than weakening it. And here's a simple rule. If AI replaces conversation, ownership,

1:20or learning, you're using it wrong. So get ready for a fun skill. Let's do this.

0:00So welcome to section number one. AI is simply an assistant. It cannot be a decision maker. AI systems are extremely good at generating possibilities. They can analyze a ton of data, suggest options, and accelerate research. However, AI lacks the context, the intent, and more importantly, the accountability required to make real decisions. So agile teams must continue to rely on discussions, a shared understanding, and responsibility for the outcomes. In a safe environment, AI can support thinking, but it can never replace it.

0:43Teams remain responsible for interpreting the information and deciding what actions to take based on the needs of the client, the PI objectives, and that entire list we've talked about over the last nine skills. AI can assist the team, but the decisions must always remain human. And in this section, as we get into it, I want to point out that yes, I am acknowledging I am using this particular image to represent AI. I know it's a little corny, but I like it anyway.

1:16So let's jump into it. Oh, I also have some AI spotlight moments along the way as well to show you that while AI is a useful tool, it does in fact make some pretty serious mistakes along the way. So let's take a look at the first one right now. It's already been proven over the last couple of years that AI produce incorrect probability calculations, financial summaries, story point, and capacity estimates. So not only is it giving you wrong answers, but it's doing it with high confidence. That's like your friend saying, yeah, rattlesnakes don't bite anyone.

1:55Go ahead and poke it with that stick, because that's exactly what you're doing here. If you're relying on data with an overly confident projection without verifying it, you're already out of the gate wrong. And in SAFE, AI might suggest story points, but the work shortest job first scores or capacity plans must be mathematically sound. If they aren't, you're poking that rattlesnake I was just talking about. So let's look at some of these points here and hop into this. I love AI. It's great at generating ideas and ideas I didn't even think of.

2:34I might come up with one or two and ask it for five more. There never seems to be an end to how many ideas it can actually generate or summarize and the options. I do a ton of predictive project management work, and you want to talk about summarizing data very, very quickly. GPT is phenomenal, but it can't understand the project context. It can't understand the human need of your clients. It cannot understand the strategic intent of your company. It cannot understand your team dynamics, how long you've been working together, who understands what, what your skills are, your strengths, who's introverted, who's extroverted, where there's conflict. It doesn't understand any of that. It only understands what you input. And if you can't successfully put that all in, basically it's just going to parrot on a neutral playing field. And again, I love AI, but it's just a tool. You must interpret the suggestions from AI before you act on them, because we've already seen in this example, it will give you outputs with a high degree of confidence and they can be completely wrong. And trust me, I've seen this firsthand. I ran an

3:51experiment I saw online and it said, ask chat GPT, how many R's are in the word strawberry? And so I asked it and it said, well, three. And I said, are you sure it's not two? And it immediately says something like my mistake. You're right. There are only two R's in the word strawberry. You can make it change its mind. It's the most helpful hype crew I've ever seen. It tells me how great and awesome I am, regardless of how great or not so great I may actually be.

4:22So you have to understand it before you act on anything. Decisions plus human judgment, equal accountability. AI will never be held accountable. So understand that whatever decisions you make from that AI tool, that data compilation analysis suggestions, that's on you, not chat GPT. You're never going to see chat GPT behind bars, but it could be you. Well, probably not that serious, but maybe out of a job. Conversation ensures a shared understanding.

4:58There's how I interpret it, my peers interpret it, my leaders interpret it, and how we collectively interpret it. If you're not having conversations, you're going down the wrong path. It is great at accelerating potential solution exploration and then implementation. It is amazing. And I've seen scenarios where I will ask AI a question and it will give me a really sound answer right away. But in more complex conversations or inputs, I've seen it take 10 minutes to wait and think about it and then give me a mathematically defensible solution. So it really does depend what kind of solution you're trying to find. You must collaborate. You must discuss whatever the insights are. They must be validated. Well, chat GPT says if I use this code right here, then I can interface with this hardware over there. Okay, let's run an experiment and let's find out. Let's not just push it into the widget without validating, without verifying, and then push the widget over to leadership. That is a recipe for disaster. You cannot outsource thinking. You cannot outsource decision-making. Remember, if accountability lies with you,

6:17then the thinking and decision-making must also be with you.

0:00So welcome to section number two, preserving team ownership. Now, one of the core principles of SAFE is that the teams own the work they deliver. They estimate the work, they commit to the objectives, they take responsibility for delivery. That's how it has to work. And when the tools begin doing this work for them automatically, ownership shifts away from the people who are actually doing the work. Now, notice I said ownership of the work shifts away. I didn't say accountability did. AI is great at assisting you with planning and estimation, but it can't replace the conversations the team must have to create a shared understanding. When teams participate directly in planning and commitments, predictability is going to improve. Ownership creates accountability and accountability drives reliable delivery. So take a look at the screen for just a moment.

0:56It looks like we're estimating here. Well, I see a 13 for a very large estimate and then some smaller estimates over here, excuse me. But then I'm looking at 34 and I'm wondering, what? What are you basing that on? And if you've ever estimated story point size, you know what I'm talking about. Anytime I start getting past eight, I start to freak out a little bit. 34? 34 would make me flee the room in terror. Well, maybe not in terror, but I would leave in disgust at a minimum. So let's clear the board and let's wander over here and look at how great AI has been. Here we go. So in this particular reality check, one of my favorites, AI hallucinated legal citations in federal court. Several attorneys were sanctioned after they submitted AI generated case citations, which were completely made up. Can you imagine walking into court as a lawyer with a responsibility to your client and to the system and having not checked the citations you were about to present? It blows my mind. But what's the parallel here? Well, AI would be great at creating backlog items or compliance references, or even documents that may not work.

2:18They may be out of compliance. They may in fact be illegal, whatever it is. So if AI can hallucinate legal citations, what do you think it can do for your product? I'm just going to let that one stand there. The teams must estimate their own work. Anything short of that is a mistake. Now, can you use AI to help you? Of course you can. But there's a danger here. The data that you input, AI is a system, and trash in will give you trash out. That's piece number one. Piece number two is, are you protecting the data that you're actually using as an input? If you're using some sort of open AI source, whatever data you're plugging into it is not secure. It can be used somewhere else.

3:07And I don't fully understand how all of that works, but I do know this. If you put client data into an open AI source, you've just violated some rules. So at a minimum, you need to be using something homegrown, something built within your company that's going to protect whatever the data is that you're putting into it. And that includes your estimations. Once you get those estimates, you need to have conversations. We need to talk about it. I may think it's a fantastic estimate. They may think it's terrible. I need to understand why they believe that. They need to understand why I believe that. And we need to move from both of these opposites back towards the middle. And the middle is called an agreement. And suggestions by AI must be validated by the team. I think one solution would be to do X, Y, and Z. Okay, can we do X, Y, and Z, team? Well, actually, no, because here they're talking about AC, and over here we're talking about DC, and there's no integration point. So obviously, we can't use this solution, but we wouldn't know it if we didn't talk about it. Teams, not tools, commit to delivery. AI can never take responsibility for the value you're trying

4:25to create and deliver. You have to do that. And as soon as you push it off, I promise you, you're the one hallucinating at this point if you think you are always going to have a positive outcome. Ownership drives accountability. Ownership drives buy-in as well. If I'm constantly letting the tool do the work for me, I don't have any real buy-in. I'm just inputting data, looking at the output, and then trying to implement. Because at some point, it's going to fail, and then I have no choice but to tell them what I did and then blame the tool for what I should have been doing. This is my fault. You're never going to see Grok or Claude or Claw, that's another one, or GPT getting fired because they didn't own it. It's not going to happen.

5:17Participation improves predictability. It can't just be me and the tool. It can't just be one developer over here and the tool. We have to do it collectively. And if we do it collectively, we're going to get better at inputting these requests, generating these ideas, these hypotheses, and then collaboratively looking at the outputs. But if one person is always doing it and they have a flawed input model, we go back to trash in, trash out. Now, AI insights are great, and you can use them to make estimation decisions through discussion. But it should never be taken as 100% correct. For example, you may ask AI how long it takes you to mow a yard. You may tell it how many steps a minute you take. You may tell it the kind of lawnmower, the horsepower, whether it's gas, diesel, or electric, the length of the grass, the size of the yard, the humidity, the temperature, whatever it is. And what it's going to do is give you a neutral answer minus any unknowns, any uneven surfaces in the yard, any storms that might pop up, any falling sticks or trees or animals running through your yard, or gopher burrows, whatever it is, gopher holes, excuse me.

6:35I don't know what a burrow is, but actually I think a burrow is a donkey, but I'm off topic. It's just an insight. It is not the automatic answer. And finally, if the team wants autonomy, they have to take responsibility for the estimates that they are putting out for these story points. They have to take responsibility for what they're committing to, because anything short of that means they're going to lose autonomy and their ability to make local decisions.

0:00All right, so welcome to section number three, where we are going to protect built-in quality. Now, AI can produce a ton of outputs very quickly, talking about code, documentation, and analysis. You can do all of this in seconds. And while it's great for accelerating development, it also introduces new risks. AI-generated outputs can appear very convincing while containing hidden errors, while containing assumptions. Without proper validation, defects can scale rapidly. Safe emphasizes built-in quality, which means verifying the correctness continuously throughout development. And AI may accelerate creation, but quality still requires verification. Teams have to review and test and validate everything before it becomes part of the solution. And I love this graphic. Can you imagine being the team that used this code from Mr. AI bot over there, and it resulted in a failure? Can you imagine how much rework there might be because of that? Can you imagine how much time they may have lost because of that?

1:16And again, this is a hyperbolic example, but it happens. It has happened. And used incorrectly, it can happen again. So let's jump over and take a look at our AI in the spotlight moment, our reality check. This one is a good one too. A biased healthcare risk algorithm. It basically said that Black patients required less healthcare based on spending. Black patients were spending less money and or they were seeing fewer Black people, which must mean they require less medical care. Ergo, Black people are super healthy, probably healthier than anyone else. You can infer all kinds of outcomes from this. The assumptions are endless. Why does it matter here? Because AI trained on historical system data can underprioritize certain features, users, or risk categories. Imagine what happens if you use some sort of AI input to tell you who your client market is going to be for whatever app you're about to put out. And if a bias is injected into that learning model, what do you think is going to happen with the output? It's one issue

2:36to have a code defect. It's a whole other animal to inadvertently show a bias towards an entire demographic, whoever it is. It doesn't matter. Black people, white people, men, women, it does not matter. So you have to be very careful. So let's jump into these. So here we go. AI output has to be reviewed always. This isn't a one-sigma, three-sigma, six-sigma moment. It is 100% necessary to review the output. Can you imagine what happens if you set a three-sigma standard and you have, I think it's 2,300 opportunities for problems to get past you? That's bonkers. 100% is the only standard you can play with here when you look at the AI output. And it's fast. But the faster it is usually means the greater the exposure. Easy example. I love chat GPT and I've asked some very complex questions. Sometimes it spits them out so fast, I look at it and immediately know there's no way it could have done this this fast. Plus, I'm kind of a skeptic when it comes to this particular tool, but I love using it. But in other situations, other versions of AI, it will say

3:55thinking. And you can see tiny print, compiling data, analyzing data, assessing data. Which recommendations make the most sense? And I can watch it thinking in real time in as much as it actually thinks before it gives me the answer. And when it gives me one answer, I'm always suspicious. But when it gives me three or four, I feel better about it. Plus, it generates more ideas to ask further clarifying questions. And that's really the name of the game here. You have to set it up to ask clarifying questions and ensure you're giving it everything it actually needs to make an informed, unbiased recommendation. Teams are responsible for validating correctness. Imagine if you got output from one AI model and then took it and input it to another AI model and asked for it to verify correctness. That's bonkers. I don't even understand the risk implications there, but it feels pretty bad. If accountability lies with the people, then the peoples must validate correctness. And if you have weak acceptance criteria, you're going to have poor quality verification. You must have very clear, measurable acceptance criteria for any definition of done,

5:14for any increment to be considered properly working, for incremental delivery, anything short of that, and quality is going to be skewed. And AI can, of course, help you understand what the quality standard might be or is. It can also help you understand what experiments you should run to verify it, but you can't let AI do it for you. This one's easy. Teams confirm the solutions meet the requirements. That's like when you ask AI a question and it gives you a completely ridiculous answer. I call them one-offs. The answer has nothing to do with what you asked it.

5:55So you may see a great solution, but it may have nothing to do with your requirements. So you have to be careful there. AI can write some amazing code, but you must review it. AI makes mistakes. We already saw it hallucinated legal citations. Do you not think it might hallucinate some code in whatever code language you're using? And the answer is bet on it. And you have to be transparent about how you're using it. What were the inputs? What were the outputs? What were the suggestions?

6:29What is your justification? If you can't explain some of that, you will find hidden issues. Transparency, inspection, and adaption. Those are the pillars that agile, scrum, and safe is built upon. You take away any one of those pillars and the other two fall down. And lastly, built-in quality requires people. Human judgment. You cannot give it up to AI. Go to any automated assembly line for auto manufacturers. You have tons of robot arms moving things around, and they're welding, and they're moving doors, and they're dropping in bolts, and they're doing some amazing work. But all the way down that line, you have people inspecting the outputs. Nothing has changed there, and nothing is going to change here. The people are responsible for built-in quality.

0:00Welcome to Section 4, where we talk about transparency and how to maintain it. Transparency is fundamental and safe. The teams, the stakeholders, must clearly understand the work that's being done, why, and how. Now, AI is great at generating artifacts very quickly, documentation, if you will. But those artifacts need to be discussed. They need to be understood. And if you're shortcutting that, you no longer have transparency. Work might appear complete, but if it lacks a shared understanding across the team, how can it be considered complete?

0:36To maintain alignment, AI-generated artifacts must be visible. They must be reviewed. They must be understood by all of the people responsible for delivery. Transparency ensures that everyone within the ART, the Agile release train, stays aligned. And I've spent a lot of time in the last nine skills talking about how alignment delivers autonomy, and trust, and decision-making, and shared value delivery. Yada, yada, yada.

1:08You know what I'm talking about. So let's move forward. I'm off that soapbox. This is a very interesting reality check as well. Flattened probability interpretations. When I saw this one, my eyes bugged out of my head just a little bit, and I'll tell you why. AI has shown problems telling the difference between probabilistic statements and deterministic recommendations. Now, probabilities are percentages. 80% probability, 90, 15, whatever it is. Deterministic, that's usually one number, 10, 5, 4, whatever it is. Imagine what happens if you're using a tool that can't tell the difference. Then the probability of getting anything done according to AI can always be 100%. And in SAFe, it can distort the confidence votes, the portfolio forecasting, or risk-adjusted estimates, which is unforgivable because it's going to impact capacity. It's going to impact velocity, and it will almost certainly impact delivery, the last issue that you want to have. So here we go. Buckle up. It's about to get bumpy. All of the work must be visible on the board,

2:26wherever it is, digital or a real board, a chalkboard, piles of paper, all of the above, none of the above, whatever it is. If I don't clearly understand what we asked AI, what recommendations it gave us, our decision logic to use one of those recommendations, the outcome and or the output, whether or not the test failed, and that list goes on and on, then we don't have visibility. Go back to what I said about the three pillars of Scrum, transparency, inspection, and adaption. If we don't have transparency, how can we clearly inspect it and then adapt to whatever the outcome is? We can't. It's bananas, which means we have to look at every generated artifact. And by artifact, it could be a burndown chart, burn-up chart, a risk-adjusted estimate like I had earlier. Go to predictive. It could be the risk plan.

3:31It could be the work breakdown structure. It could be the estimates for any particular task in the schedule. It could be the cost estimates. Can you imagine the impact of generating an AI-driven risk register with recommended actions and cost and assigned risk owners and equipment needed? It sounds great, theoretically, but the reality is you can't do it. Hidden work reduces art alignment. I talked about that. If we don't have transparency, at some point along the way, we are going to break the train. That cannot happen. If there's anything in the backlog, everyone needs to understand what it means. Now, sure, there are some story points that may have large estimates because we clearly don't understand the work yet, but we're going to break them down further and figure it out. But if there's an artifact there or a code fix there, tech debt or some other kind of enabler or enhancement, and we don't understand it, why is it there in the first place? It simply shouldn't be. Eventually, we will produce something, and stakeholders will ask us how we went from point A to point Z.

4:43Our decision logic needs to be sound. It should start with, well, we input this data into the AI tool. It gave us these recommendations. We tried these experiments, and we confirmed that only one of them would do X, Y, and Z. Plus, it meets the acceptance criteria in the contract. However, you made your decision, if that was it, should sound reasonable. What it shouldn't sound like is, well, we relied on AI, and this was the cheapest path, so that's what we did. Yeah, I know it didn't work out, but it was the cheapest path. I was trying to save you money.

5:19It doesn't matter if you're trying to save people money if you're not delivering value. Delivered value and increasingly higher value through incremental delivery is the name of the game here. Documentation. We must have it. Everyone must understand it. It must be visible. I'm going to leave that one to stand on its own. And never mind the fact that transparency prevents rework. If we don't understand what led us down this path, we can't prevent rework from happening in the future. Go back to code generation. Unless we review 100% of the code being generated by the AI tool, we have no idea how many mistakes may actually be there. If we have some crazy experiment where we say, just go ahead and check every other line. Okay, that's 50%. Is that enough to avoid rework? I think we know the answer to that one. And finally, the art, the agile release train, depends on visible work. If it's invisible or tightly held or questionable or even sketchy, how are we going to maintain trust? And the answer is, we can't.

0:00Okay, as we start section five, I'm going to stay off screen for just a moment because I don't want my big fat mug over the word responsible. This is all about using AI in discovery and the planning portions of your project regardless of the project type. So here we go. Now AI is extremely valuable during discovery and research and planning because it can help teams analyze a ton of market information or exploring options and identifying potential risk. It's really amazing. During early exploration, AI can accelerate the learning process, but discovery activities generate hypotheses, not final answers. So don't think just because you're asking AI what you should do that it gives you a recommendation that you should do it. You should not be doing that. You should be using it to help you figure out what your opportunities might be. That's very different. AI can assist the product team in generating ideas, analyzing data, but the prioritization and commitment remain human responsibilities, and I've talked about this in previous nuggets. Look, teams must validate the ideas through experimentation and feedback. That is a rule that cannot be broken.

1:19So let's move forward here. Let's take a look at our AI reality check. This one is pretty bad too. Let me grab Mr. Penn. All right, I've got Mr. Penn. We have misaligned fraud detection. Now back in 2023, they discovered that some AI fraud detection algorithms for some platforms had a racial bias because they were canceling orders for people within a certain demographic or a certain location or within a certain annual income. So AI can penalize new customers or transactions in new markets because they're just not part of any predefined patterns, and if you can't make the leap to this one, here's an easy one. How many times have you been on travel, maybe for the first time in a couple of years, and you use your credit card or your business card and suddenly you get an alert on your phone that says suspicious activity for this card, this location. Press Y to accept.

2:22Press N to be connected to a customer service rep because I've seen it a bunch, and I've also had people call me. Hey, this is Capital One Bank. We noticed card activity in Eugene, Oregon. Is that you? Oh yes, it is me. I'm out there for the next couple of days. Okay, then we'll take the flag off and we'll let you go about your life. It happens all the time. So let's jump into this one and see what we've got. Obviously, it's going to accelerate research and analysis. I cannot understate how fast AI can sort through piles and piles of data to pull out trends, to analyze, to make recommendations, to show you patterns. It is amazing. In another example, I set up a way to take a list of commands and directives and drop it into ChatGPT and then drop someone's LinkedIn profile link in there, and it would spit out a two-page report on any particular person.

3:24And I did this back when I was in sales, two and a half years ago. What they liked, what they disliked, any patterns, places they traveled, conversation points, and stuff like that. It only took about five minutes per person, and it was wildly fast. But let's keep going. Obviously, it's not a one and done. AI and ChatGPT and large learning models can do a whole bunch of stuff all at once, simultaneously. So instead of exploring a single option, or perhaps one at a time, you can set the parameters and let it explore multiple options. And by multiple, I mean as many as you can think of, it can probably figure it out. I've casually asked it for 30 examples for various things in project management, and it takes about 0.3 seconds to start spitting it out.

4:15Wildly fast, but I'm not trying to sell you on it. You already know it. Those insights are going to support product decisions. Maybe it's a make-or-buy decision. Maybe you're trying to decide if this particular piece of unique code needs to be made in-house, or maybe you can buy it off the shelf and drop it in and deal with it as tech debt later on. Whatever it is, this can help you figure it out. It can also help you figure out availability. You may have something unique to do, and it's wildly available in the marketplace, or it's not. So then you have no choice but to build it. Remember, you are going to develop hypotheses. You're going to experiment and confirm or refute.

4:59It is the only way to validate whether your hypothesis is correct or incorrect. This is science. Nothing's changed. Planning activities must be collaborative. It can't be just you or just them. It is not the me anymore. It is the art, the agile release train. It is the we. We are all doing it. If there's a trade-off, we have to talk about it, and that might go back to the previous example of build it in-house or buy it off the shelf. That may go to the heart of testing something in-house or pushing it out to a vendor, whatever it is, and that list is endless as well. Data is going to help you justify your decisions, and one of those decisions is going to be what the priorities are.

5:47If you're looking at a backlog and you're trying to figure out which one of them is a priority as it relates to fastest deliverable value, maybe AI can help you with market research, which demographic and which location is using which brand of soda or buying what kind of car or using what kind of phone app, whatever it is. There's a ton of data out there. If you can pull valuable information from it, and finally, continuous learning is the name of the game. You may get wonky results with AI at first until you learn the input patterns and what kind of language you should be using to teach it to help you get the appropriate results. You may not get it right the first time, but I guarantee you that over time, through open collaboration and conversation, you'll dial it in a lot faster. I know I have.

0:00Okay, in section six, we're going to talk about leadership responsibilities with AI. And let's face it, change and control starts at the top. So if we don't have clear guidelines or guardrails or a vision, it's going to be crazy time lemonade. Leadership plays an important role because they're the ones introducing and or allowing AI within the organization. And without clear guidance, AI tools can unintentionally recreate traditional command and control structures. And I'll explain what I mean in just a moment. Leaders have to ensure that AI adoption strengthens, not weakens, team autonomy. Responsible leadership encourages experimentation while maintaining very clear guardrails. The goal is to empower teams with new capabilities while preserving the principles of agility. So let's jump into this one. I am stoked. We're almost to the end. Here we go. This is a really neat reality check too. This is all about what the courts have ruled. And the courts have ruled that professionals are accountable for AI generated content submitted under their authority. So even if it's fake, even if it's wrong, even if it's

1:18plagiarism, even if it's copywritten or whatever it is, it's on them. You're never going to see ChatGPT or Grok or Claude behind bars, but it could be you. And as it relates to SAFe, SAFe leaders are responsible for the outcomes of AI-assisted summaries, art level updates, regulatory content, portfolio management, portfolio projections, ROI projections, whatever it is. AI is a fantastic tool, but it should never be mistaken for a one-stop shop that's infallible. That is crazy talk. It cannot happen. So let's get into this. If we don't have guidelines, we're just going to make them up as we go. It's really about authority. And there's only two ways to figure out what your limits of authority are. Ask or exceed them. One of them is proactive and one of them is reactive. You're almost never going to be fired for asking what the guidelines are, but you might be fired for violating the guidelines you may or may not know even exist. So my suggestion here as it relates to AI, be more proactive. Guardrails ensure that AI supports agile principles, transparency,

2:39inspection, adaption, decentralized decision making, and that list simply goes on and on. Now earlier I said AI can actually introduce traditional command and control infrastructure, and now we're going to talk about it. Rather structure, not infrastructure. It's not making any buildings, but it could be making decisions, and this is what I'm talking about. Let's say we have a product management VP or maybe a C-suite leader who has some sort of AI-driven portfolio manager, whatever it is, and he or she looks at it and goes, wow, this is suggesting that if we just do two more increments here, we can increase ROI by this much. Or this one is suggesting that this should be a priority over that without any backup. Or perhaps it's saying that this team's capacity has fallen below that team, and I need to do something about it. When you have control over these kinds of tools, it's easy to use them as a sounding board instead of people, and when you rely on AI, it's going to give it to you in a sterile, neutral way. But it shouldn't be used to make hard-line decisions like that. It's just giving you insights. It's giving you options. It shouldn't

3:57be giving you definitive directions to move in. Teams have to retain their autonomy. If we trust them to do what's right, then we can trust them to do what's right with AI. If we give them guardrails and we trust them, let them experiment, let them fail, let them learn from it. They need to maintain autonomy because as soon as you move back to a command and control structure, you're not an agile anymore. You're not building high-performing teams. You are putting your thumb clearly on their head, possibly at the behest of AI. We need to be safe to explore the tools, and we have talked about this. Psychological safety, the ability to ask questions, try new things, to fail, to push back respectfully against ideas, whatever it is. But we still need to explore different tools. To do that, we need safety.

4:51It's not rocket science. This isn't the only place we need safety, but it's definitely one of them. Anytime you bring a new tool into the system, the art, in this case, people are going to be suspicious. You have to let them figure it out, let them make a few mistakes, let them get better. And as long as you give them the autonomy to do that, they will. I feel like I just said this. I'm going to let that one stand on its own. AI has to be modeled responsibly at all levels, and here's what I mean. If you give them guidelines where they can do X, Y, and Z, and then explain their logic, and someone walks into your office and asks you a question, and then you use AI to answer that decision without explaining your logic or even understanding it, that's irresponsible use. Or if they perhaps walk in there and see your screen open, and you're using AI to produce some not-so-family-friendly images, that's not responsible use. If you're going to tell people what to do, you have to do that as well. I believe it's called walking the walk and talking the talk.

6:04You have to do them both, or you really can't do neither, not as a leader. And adoption should strengthen the teams, not weaken them. It should strengthen their agility, their ability to adapt, their ability to be transparent, to inspect, to experiment, to learn, without fear of any kind of retaliation, as long as they're staying within the guidelines set by leadership.

0:00Okay. In this final section, we have ethical and customer impact considerations. Look at that bottom statement. Assess AI practices as part of routine inspections. We already know about the feedback loop. We should be using them here, and if we are using an AI tool, we have to adopt that into our practices as well. For example, check out this routine art inspection checklist here. We're going to review the stories. We're going to measure outcomes. We're going to assess the usage, evaluate the risks, and see what improvements might exist. Here's the interesting part. If you use AI, you should be able to clearly measure whether you are getting better or worse, and if you're getting worse and you can pinpoint it to AI use, then you need to stop. If you're getting better and you can pin it to AI use, then you need to clearly understand exactly what part of the art that is benefiting from AI use. If you don't clearly understand what's making things better or worse, then you shouldn't be doing it. There's a huge impact to your company. Maybe some sort of legal violation. There's a huge impact to the customer. Maybe some sort of proprietary

1:16information leak or personal information. This list is out of control. This stuff can get out of control very quickly, and there's a reason they have cybersecurity insurance for hacking, penetrations, exposures, and stuff like that. AI is a great way to fall into that trap, and I think I'm remembering that the rapper Eminem is currently suing someone for $140 million because they used his entire library to help train one of their models, and they didn't ask him. I think that applies here as unethical use. So let's jump into this one with our final AI reality check.

1:59There was a tool called Compass within the U.S. criminal justice system, and it had a bias. It overpredicted the likelihood of recidivism, meaning the potential to offend again to commit crime for black defendants far more than white defendants. The impact is fewer black defendants were getting out earlier. They weren't getting parole. They were being penalized by a biased AI model. It's like whatever you think is a bias, and if you build a tool without establishing a neutral platform or a neutral standpoint, you are going to inject your own bias into it, and the reason it matters here is because a risk algorithm might overscore or underscore a group of people. Maybe you think that women of a certain age bracket will be early adopters of this new feature, and you're wrong, or maybe you think that this particular demographic won't like this feature, but then they actually request it. Those are biases, and it happens all the time with people.

3:10We have internal unknown biases. We just have to deal with it, and so if we're not using the appropriate mindset and having conversations with others so we can balance these biases out, you're going to end up with these kinds of problems. So I think I've said enough about that. Let's just rock on. You have to consider the customer impact before you use AI to make decisions, before you use AI to prioritize what you think the customer is going to find more valuable.

3:42You may expose some sort of proprietary customer information or capability within open AI, and now it's just out there, and if someone finds it on the back end, and I don't know how all that works. There's a bunch of brainiacs out there doing magical stuff with AI, but if they find it and then try to benefit from it, who do you think is on the hook for it? That would be you, because you're the one that caused it, and like I said, the system may introduce bias if not monitored. I've seen this firsthand. I've asked chat GPT about which states are the best, which religions are the best, which age brackets are the best from a generational standpoint, Millennials, Gen X, Gen Z, Boomers, and stuff like that, and when you read some of the answers, you can clearly see the bias. Now you might say maybe you just asked the question wrong or the data input was wrong, and that's possible, but the bias isn't hard to see, at least not for me.

4:46I talked about this. I'm going to let it stand. You cannot use an open source system for your project, for your product, for your customers, for your clients. You simply can't do it unless you get express written and signed permission to do so, where they very clearly understand that their information is going to be put out in an open AI landscape on the interweb, and I'm willing to bet if you explain that to them, they're going to say nope, not having it.

5:18Don't assume. Assumptions are beliefs you hold to be true or false in the absence of proof, and if they're wrong, if you can't validate them, you cannot implement them, so it's either a yes, this assumption is valid or not, and if it's valid, maybe then we can implement it, but if you can't validate it or refute it, it's kind of hanging out in the middle, then you absolutely can't do whatever you're considering, and if the art is going to maintain trust, we must maintain transparency, and that includes with our customers. They need to know what kind of AI tools we are using, how we are using them, and what we're doing with the output. If you can't explain that, you're probably going to get into trouble in some manner. Ethical considerations guide responsible delivery, and here's what I mean. Honesty is something you have when everyone can see you. Ethics are what you have when no one can see you, so just because no one's around doesn't mean you have free reign to start inputting data or using AI in a manner that you shouldn't be doing, because the organization retains accountability.

6:29They can be sued. They can be dragged into federal court. They can be dragged before Congress. So could you. It could cost them millions and or billions, whatever it is. This is a brave new world as it relates to AI, and responsible innovation enhances long-term value. You can't roll this out in various localities and not do it at the art level. You can't have some sort of haphazard AI system at play within the art. You need to clearly understand the impact within the company and outside of the company to your clients, customers, and end users, and if you don't understand that, then you're not responsible enough.