Name: AI as an Engineering Leadership Multiplier
Uploaded: 2026-10-08T20:31:58.339Z
Duration: 49 min 20 s
Description: AI as an Engineering Leadership Multiplier

Transcript for "AI as an Engineering Leadership Multiplier": Hello. Hello, everyone. My name is Peter Saulitis. I work on content and digital programs for developers and engineering leadership here at Anthropic, and I will be your moderator today. Thank Thank you everyone for joining us. It is always so much fun to see all those emojis with different countries. Thank you for staying up late. Probably past my bedtime in a lot of regions, but, so much fun. Thank you all for joining. How's everyone like in Opus 5.5? I feel like I'm slowly reviving all my dead domains from the graveyard, and I hope you're having as much fun coding right now as as we are here to Anthropic. This is the second session, in the guest lecture series on leadership and engineering. If you were able to catch... A couple weeks ago, we had Will Larson on on road mapping with decisions rather than dates. Be able to see that session. Drop an emoji. Welcome back. Great to see you. If you are new to the series, this is a three part series where we are bringing on engineering leaders from the field to come on and impart some of their their wisdom and and lessons that they've learned from leading large enter... Engineering teams. And so following Sudheer's talk, that is next week, next Tuesday on the thirteenth, we are hosting Jean Sue and Kate Huston. They're gonna be going and talking about career choices and engineering today. It's gonna be a great session. And so if you've registered for this entire series, you should be able to get calendar invites and notifications to join that. If you have not, please go and register Great. And as always, housekeeping, one of the questions we get every single webinar is, will a recording be available? Yes. This session is being recorded, and we will send this out to you within twenty four hours of the session. The question bar and the Q&A panel, please... This... The the most important valuable aspect of this session with Citi, or you have his time, you have his attention, ask questions. We will have a Q&A at the end of the session. We'll get to as many as we possibly can. We have quite a few people, almost a thousand people on the line right now. So there will be questions that are unanswered, but we are going to do our best to answer all the audience questions. And then following the the session, there is a, survey that we're sending out. Please feedback is so important so we can keep delivering these these awesome sessions to you all. Cool. And I am so excited. Without further ado, I'm gonna be introducing Sudheer Anne to you today. I am so honored to have you on today, Sudheer. He is the head of engineering for the product capital at Stripe. I actually came across Sudheer from a LinkedIn post he wrote about, some custom skills that he had for maximizing his productivity as an engineering manager, team check ins, weekly updates, performance evaluations. It was so cool. Sudheer, are you still using these these skills? I know this is our January, and I think things things have changed since then. Right? Yeah. I I I would say I'm still using most of them in my day to day. Definitely have expanded upon, the the full list here for sure. Awesome. Well, it is so cool. I've gotten a sneak preview of the presentation. You know, it's great to have you, you know, talk about how you leverage AI and also human judgment and how you you you manage, engineer, leading your engineering team. So without further ado, I'm gonna let you, introduce yourself, Sudheer. I'm gonna pop off stage and let you, present. Alright. Awesome. Okay. Cool. Well, first of all, thanks, Peter, the entire Anthropic Team for putting, you know, this webinar together. There's a ton of ton of work that's been happening behind the scenes to get this ready, and it it's been really awesome, that we have this event. And so definitely excited to talk about AI and really how it's impacting, you know, engineering leadership. I think a lot of the conversation around AI and engineering to date has really been focused on software engineers, and I I think rightfully so, the the varied nature of what it means to be a software engineer is dramatically changing, you know, from even six months ago. But, you know, the reality is AI is impacting every function, and engineering leadership is certainly no exception. And so, you know, as our teams are accelerating with AI, our jobs as leaders, like, it... It's... They're not getting smaller. They're getting bigger. The service here is getting wider. And so, you know, if we keep working in the same way, all that acceleration just turns into more work for us. So today, you know, I really wanna talk about how we can leverage AI in our own work to make that increase manageable. So I'm gonna share a mixture of theory and examples of, you know, how I'm using AI to make sense of everything happening across my teams, how I use it to sharpen and pressure test my thinking, automating recurring manual work. And then I think, you know, equally as important, I wanna talk about, you know, where I draw the line, where I think it's really important to still have our own, you know, unique thinking voice and personal engagement as leaders. But first, I wanted to share just a little bit about me. My name is Sudheer. I'm currently a head of engineering at Stripe for our capital product. If you're unfamiliar with Stripe Capital, we help small businesses get access to fast and easy financing to help them grow. So, you know, if you're interested in learning about lending, please come find me after the talk. I've spent about twenty years in tech, my first ten years as an IC. Did everything from, you know, low level embedded systems all the way up to, you know, large scale distributed systems. Last decade switched into engineering leadership roles, worked at companies like Block and Gusto and Microsoft. And picture on the right, that's me and my my beautiful family. My kids, they they look like little angels here, but trust me, they're they're little maniacs, but I do love them. And let's see what else. Born and raised in Los Angeles. It's where I currently live. And, yeah, diehard, Lakers and Dodgers fan. So apologies to any Braves fans in in the audience. It's a good game last night. Alright. Without further ado, let's let's get into things, shall we? Does this look familiar to anyone in the audience? I'm willing to bet if I removed all the labels from this graph, you would know exactly what you were looking at. This is my organization's code throughput from January to August of year... This year. So it's monthly merged PRs, and you can see this has grown by 60%. Now my org has grown a bit, but even if you normalize this by, you know, number of PRs merged per engineer, that status still, like, up 30, 40%. So I think it's really quite astounding, you know, how much has shifted so quickly. I... I've been in this industry for probably, you know, close to two decades. I can't recall a single paradigm shift or technical innovation that has so suddenly shifted what was possible. And, you know, that red dotted line there, you know, the founders at Stripe, they they they jokingly refer to that as the singularity. But, you know, we really are kinda living in unprecedented times here. It's it's it's really a huge change that's hitting our industry. And so, you know, this explosion in output, right, I think is really stress testing every stage of how we build and ship software and forcing us to rethink, you know, our playbooks. But all this, like, extra output, it it doesn't come for free. Right? It means more decisions to make, more coordination, more follow through. And, you know, a lot of that actually lands on us as leaders. And I think this picture is probably a more accurate description of what it feels like leading teams right now. There's more PRs to review, more decisions to make, more cross team coordination because now everybody's shipping code and everyone else's, you know, code base, and you have you have designers and PMs shipping code and, more incidents because now there's more things in production. And so, like, none of these things are new in isolation, but I think what's new is how much of this is happening all at once. And, you know, one of the most common things I hear from my team right now, and this is, you know, both ICs and managers, is just burnout. Right? It's really hard to keep up with all of this, the context switching across all the agents and work streams. It's taxing, very taxing mentally. And so for me, the hardest part isn't even necessarily keeping up with everything. It's really connecting the dots. Right? Figuring out what actually needs my involvement, what are the things my team has handled, and, you know, that work is eating into the time that I should be spending coaching my team, thinking ahead, strategizing. And so, you know, this is really what pushed me to look into AI and looking for leverage in my own work is really out of necessity. You know, I was I was starting to kinda drown in some of this. And so, you know, the question I asked myself is, hey. Could AI help me make sense of everything, you know, that's happening across my team? And that's really kind of the core kind of problem I wanna focus on today. So I've organized where AI has been useful in my own work into four broad, principles. The first is, synthesizing context. So using AI to connect information across lots of disparate data sources, helping you figure out what actually needs your attention. The second principle is using AI to sharpen my own thinking and improve the quality of decisions I'm making, you know, challenging assumptions, exploring different alternatives, finding weaknesses, gaps in my own reasoning. The third is, systematizing repetition. So really automating all of the toil and manual work, you know, out of out of my life. And, you know, lastly, it's really kind of a check on on these first three is kind of keeping and staying human as a leader. Right? I think there are things AI can definitely help you with, but, you know, at the end of the day, your judgment, your relationships, your voice, your accountability, those are still yours. And I think it's important that those things still come from you and, you know, not from your agent, because at the end of the day, we... We're still humans leading other humans, at least for now. And so it's important, you know, not to lose sight of that. Alright. We're gonna get into the first principle here, which is around synthesizing context. So I think, you know, probably all of us are using AI today to do a lot of summarization is probably, you know, it's one of the things LLMs are phenomenally good at. And I think that's a perfectly good starting point. But you can read a great summary and still wonder your... To yourself, what do I actually need to do here as a leader? And as a leader, what I'm actually doing when I'm reading anything is I have specific questions in mind. Right? Is there a decision that's waiting on me? Is there a dependency that's not being surfaced? Is something falling through the cracks? Is there a business risk, a customer risk? And those are the questions that I actually want AI to help me figure out, not just telling me what this says, but what's actually important and where should I direct my attention. So how do we get from a generic summary to that? There's, I think, a progression, that I found to be pretty useful here. And and sort of the key to unlocking this greater reasoning is kind of running through these steps that you see here. And I think most people probably start and stop at this kind of summarization layer, which is like, hey. What does this thing say? You know, TLDR this thing. And like I said, it's great for getting you oriented, saves time. But if you think about what we actually do as leaders, we run through a couple of more steps. Right? We're synthesizing information. What what are all these facts mean together? You might have a date over here and a comment over here, and you put them together, and you get something, you know, that means something completely different. And then we're we're interrogating information. Right? What doesn't add up? We're pressure testing our assumptions, looking for gaps. And then finally, we kinda run through our own prioritization step, like, what actually matters? What needs my judgment? What can I delegate to my team? And I think each step down this ladder gets us closer to the work that we actually do as leaders. So for me, the value of a up... Of AI goes up tremendously the further down, you know, the stack we get. And that's where I think AI can really start to become a multiplier for you once you start to figure this out. But I think AI can only help you get this far if it knows what you care about and how you think. So how do you do this? I think the key unlock here is really teaching AI what you look for, and I think there's three broad things you have to always keep in mind. The first is you have to give it context and and, you know, if you spend time researching AI, you know context is king. And so, specifically, your role, your priorities, your teams, what do they do, the decisions you're responsible for. If you guys use Claude or codex, they both have this concept of memory. And memory is basically just a markdown file that, persists across sessions. And so every time you start a conversation, that memory gets injected into the context window. I encode all of this information, into my memory file in Claude. And so this way, anytime I ask a question, it can reference that memory, and it it has a sense of who I am, what are my teams, what am I doing, what is my role at Stripe. And I... I'm... I have that context loaded every single time, which is great because at the end of the day, an engineer, a manager, a VP, they could ask the same question, or read the same proposal and come away with completely different questions. So how how is AI supposed to know what you want? You you have to give it context. It's super important. The second part here is really explaining how you assess the work. So the questions you ask, trade offs you weigh, the risks you're looking for. One of the best ways that I found to do this is actually to show AI examples. If you have past reviews that you've done, feedback you've given, you know, explaining why something caught your attention. One of the... Like, one of my favorite tricks is to just feed raw transcripts of meeting notes actually into Claude and kind of work work with it and say, hey. What are you noticing about how I'm running this meeting? What am I paying attention to? And so in this way, you know, you get a really lightweight version of kind of your criteria and judgment, and you can start to encode this in your own AI loop. And to be clear, you know, this isn't about getting AI to make decisions for you. I'm not trying to create a mini version of me to send out, you know, across my team, but I... It's about getting it a sense of your criteria and judgment. So the first pass you get actually reflects what you care about. You still have to check the output. You still need to form your own point of view, but you get something that's a bit more useful, than a generic summary. So I wanna show you what this looks like in practice, with an actual, skill that I built. I modeled this one around the concept of a leadership brief. And if you don't know what a leadership brief is, like, you know, imagine you're communicating with an executive, your skip lead, a VP, maybe somebody in the c suite. You're probably not, slacking that person your stream of consciousness thoughts. You're probably being very careful about how you communicate with that person. You're probably giving them a clear summary. What do you need from them? What are the things they should focus on? What are the, you know, next steps you need? And so, unfortunately, for me as a leader, I I don't get that. I'm bombarded every day with, you know, documents, emails, proposals. I'm cc'd on Slack threads that are, you know, a 100 messages deep. And so it's pretty dizzying to keep up with this. Right? And I don't have time to read everything in the moment. And a lot of times to even build context in in any one of these things would take a considerable amount of time. So I thought to myself, wouldn't it be great if I had a skill that could just generate this for me so that everything that lands in my inbox, right, I I can quickly know how to prioritize this and what to do. And so that's really kind of what I modeled this skill around is this leadership brief. And if you've never built the skill in Claude or codex or pick your, you know, your favorite harness, it's really just a reusable set of instructions for how you want AI to approach a task. And, essentially, it's a markdown file, which it... You can load on demand and call it, you know, whenever you want. So for this specific skill, there's a couple of key components. The first is, you know, gotta give it an artifact. What is the thing you're trying to get, you know, informed on? And the nice thing about, you know, AI is that this can be anything that, you know, in this case, Claude can read. It could be a document. It could be a Slack thread. It could be an image. It could be a spreadsheet. It could be a slide deck. As long as Claude knows how to read it or there's an MCP tool it can access. So there's... You have a ton of flexibility with how you can invoke it. And the second key thing here is that context that, you know, I was telling you about earlier. Right? I point it to my memory file, which has all of the information about, you know, my role and responsibilities. And we use this to seed kind of the second phase of the skill, which is this reasoning step. And, you know, what I've actually encoded in the skill is a bit more complicated than than this, but this I truncated for, you know, the sake of the presentation. But essentially, it's running through the loop that you see here, which is, you know, summarizing the data, synthesizing the facts, interrogating, prioritizing what's important, and it's using that context from the first step. And then what you get in the third stage is this really clear output, and this is structured for how I like to receive data. And I'm sure you have a format that would work for you, but I I wanna know with anything, hey. What is a quick overview? What is this? What's the TLDR, triage step? Am I actually needed? Is this something I need to pay attention to right now? Is it something I can defer to later? Does my team got it? Do I need help? So having that triage step is really important. And then, you know, I get a couple of focus areas of, like, hey. Pay attention to this, this, and this. So if I want to follow-up, I can easily do so. And then I get a really clear next step. Right? What's the most important next step that I should I should take here? So I'll walk you through what this looks like with, like, a a real world example here. It... Pseudo realistic example. I tweak some of the facts here, but largely accurate. So, you know, within my team, on capital, you know, we receive thousands of loan applications every week. Each one of those applications actually checked by a human to ensure the person who's applying is not fraudulent. They're not trying to run away, you know, with Stripe money. This is, of course, very expensive. We have an entire vendor team that's staffed up to handle this load and, you know, leadership as you can imagine, you know, they're trying to save costs. And so there's a push to fully automate this process, wind down this vendor team over time. So we have these two opposing forces kind of hitting each other. There's this push for automation. My team has to build this automated service. There is a hard deadline when this vendor team is gonna get cut. And so all of this sort of landed on my desk as a as a proposal. And you can see here on the left, right, the AI summary does a pretty good job. It's it's got all the information here, and you can see there's a problem right away. The vendor team's gonna be cut June 3. This automation pilot's gonna go live June 10. There's gonna be a week where where, you know, we're gonna have a capacity gap. But if I was, like, skimming through that really quickly, I might have missed something. Right? Or or it might not have connected the dots as quickly as you can see on the right with the leadership brief skill is very clearly telling me in, you know, bold letters, hey. This needs your attention. There's an SLA breach that's very likely. I get a clear sense of, wow, there's no backup plan for that week. And I get a really clear next step, which is, hey. You should talk to your leadership team and see if you can delay, you know, this timeline a bit. And so that's kind of the difference here is, you you get a much more useful first pass here when you are encoding this information about who you are, what your judgment is, what you care about. That can get you a bit further than, you know, kind of what a generic summary might give you. So to be clear, I... I'm not, like, running around, running this brief skill on everything. Like, there are, you know, there are still times when I will go and fully read something, but this can help me prioritize my day. Right? If I don't have the time to really dig into something, I can at least get a quick read, and know where to prioritize and stack rank this, in my day. So super powerful concept, like mixing in, the context and and also your judgment and figuring out how to encode that, into your workflow. Alright. We're gonna move on to the next principle here, which is really, you know, using AI to sharpen your thinking and decision making. And I think it's one thing, you know, to ask AI to examine someone else's proposal. I think it's another thing, you know, to ask it to challenge your own. As leaders, I think we often come into a situation with a preferred answer or option, but, you know, it's... And it's really easy to use AI to build a better argument for the decision you already want to make. Right? But a stronger argument is not a better decision. Right? And my perspective is the more useful move is to give AI your reasoning, evidence, constraints, and then ask me, hey, where might I be wrong? You know, you're not looking for disagreement for the sake of disagreement, but you might... You wanna see what you might be missing before committing to a decision. So here's an easy trap I see a lot of people fall into. You know, this prompt here on the left, you know, I think we should launch Friday. Does that make sense? As many people, you know, structure their prompts in this way, you're kind of already telling the model the answer you're leaning towards, and it's very likely to give you a very convincing answer and leaving you feeling more confident. But your decision has not actually been tested. And and there is real research, you know, behind this that studies on model behavior found that training can sometimes reward agreement over truthfulness. And, you know, the models have gotten better at this over time, but, you know, agreement alone is not evidence that you're right. And you can see here the prompt on the right side, you're you're asking something very different. Right? Here's the plan. Here's the evidence. Here's the constraints. What would support launching Friday? What would make you delay? You're really inviting that rigorous scrutiny. And and I think the more valuable question to ask is, you know, what evidence would actually make me change my mind? I I might still decide to say launch on Friday, but at least, you know, I can articulate why I made that decision, right, and what assumptions that I checked. So you can you can get a much more robust decision coming out of this. So here's how... You know, a couple of ways you can put this into practice. You know, I think I like to use AI as a as a sparring partner as I'm making decisions. And I think there's a couple of different angles you can use to test your thinking. The first is, you know, just challenging the assumptions by asking, hey, what am I what am I taking for granted here? And there often are hidden assumptions underneath all the decisions we are trying to make. Like, maybe we're assuming we're gonna get head count or some dependency is gonna arrive on time. So it's good to surface those and know what they are. The second one, and this is my personal favorite, is, like, argue the other side. Right? Make the strongest case against my preferred option. This can help you take opposing views much more seriously, especially when you're already invested in an answer. The third one here is generating alternatives, and this is really, like, open up the solution space. Right? We often frame decisions way too narrowly. Right? We give ourselves binary choices, but really there's likely a spectrum of options that are helpful. And so AI is really good at helping you think outside the box here. And the last one is a fun one is, like, run a pre mortem. Right? Imagine I made this decision. I went ahead and then, oh, the project failed. Right? Why did it fail? You can check what failure modes are actually plausible, what warning signs to look out for, and then what you can do now to prepare for that. So to be clear, you don't need to run all four of these every single time. But, you know, pick the question that fits and, you know, evaluate what comes back. And, the key thing here is, like, you're not always gonna get this level of scrutiny or pushback from your team. Right? They might be busy and doing other things, but AI is always there to help you make these decisions. It's never gonna tell you no, and so use it, you know, and and maybe you can even go build a skill that can do this for you. Okay. We're gonna get to, I think, probably my favorite part of this presentation, which is, I call this automating the toil out of your life. You know, and this has been the promise of AI, I think, since, you know, things got going is, you know, we can use AI to eliminate all the manual kind of drudgery in our lives, and we can focus on all the high leverage stuff. And I think in practice, I don't know that we've quite hit that, that goal quite yet. And the the reality is it's it's really hard to automate the things that humans do because we're we're very messy. And so, I wanna talk about what types of work are are actually good candidates for automation, and how you can build this in in your own day to day. I think there are broadly three principles that I look at, and this is through a lot of trial and error that I found that that make it a good candidate for for automation. The first is... I mean, it's obvious. It's gotta be recurring. It's gotta be something that happens every day, every week, every month. If it's something you're doing every six months or every year, there's probably not enough, ROI to to warrant building an automation. The second is that, you know, the context is sort of scattered and there's an element of, like, say, data gathering where you're you're pulling data from a bunch of different sources, maybe, you know, docs or Slack or Jira. You know, we know that AI is, you know, phenomenally good at sort of this data collection part. And the third, and this is really the most important part, the analysis that you're doing is repeatable. Right? So either you're running the same checks, you're asking the same questions, ultimately, you can distill kind of your logic into a set of rules or some kind of a deterministic workflow. If the thinking is very different every time, that's probably a good sign that that's a unique judgment that you should hold on to. Right? But if if all three of these are true, you've got a great candidate, and I think that's where AI can really help you automate, that toil. That's what it's really good at. So I'm gonna show you an actual example, around something that I, spent a lot of time with, unfortunately, which is incident tracking, you know, definitely not my favorite, task. But I wanna talk through, you know, how I used AI to really automate this part of my, workflow as a manager. And, you know, when I first got to Stripe six, seven months ago, I was actually tracking all of my team's incidents in a Google Doc, so very low tech. And so every day, I would go through all of our open incidents. I would gather updates from every Slack channel, figure out what was stuck. I would go chase people down. It worked, but it took a ton of time. Right? And everything depended on me remembering to do it. Like, I was very much the bottleneck. So I got tired of this, and the first thing I did is I I just built a dashboard with AI, and it's super trivial to do that now. And that was great. I got this three sixty degree view of every incident. I could filter by the state of the incident. I could see what teams, you know, we're we're we're doing what. I could see what the next step was. And so this was a huge step up. Now, like, I can very easily follow-up with people, but I was still the one who was checking it every day. I was still the person who was nudging people. Right? I was still the bottleneck. And so I thought to myself, like, how do I remove myself from this loop? Right? And, ultimately, you know, what was the analysis that I was doing to determine, you know, when to follow-up, when to nudge somebody? You know, there there was, like, a crude logic to it. And so the next step was really building an automation and building a daily report, that goes out where I could actually encode this this rule set into a prompt. And and that's sort of the third iteration, which I'll show you next, is now I have a daily report that goes out. It groups incidents by, you know, the action that's needed, what's the next step, and it tags the owners directly. So people are getting the nudge without me in the middle, and that's really the part I encourage you to think the most about. You know, don't stop at a tool or a workflow that you have to remember to execute. Right? I think the real win is when this all happens without you you in the middle and you in the picture. So here's what, like, the actual report looks like. You know, this is fictional data here, but you can see there's, like, this needs attention bucket here that's not an actual incident state at Stripe. That's like a severe created state of of me encoding my very crude logic of, you know, hey. This is something that we should go look at immediately. So this report, like, is visible to my entire team. All the engineers and managers see it every day. And so you can see here, like, Jordan is in big trouble. This incident has been stale for eighteen days. You know, what's going on? And, you know, anyone reading this can can easily see it and follow-up. And in this way, like, this takes me a couple of minutes to read every morning, and the follow-up happens very quickly. And more importantly, I don't even need to be in the picture for this to run. Right? Now this is set. The team is handling it. AI is handling all the chasing. If I really want to go follow-up on something, I absolutely can. And huge time savings for me. Right? This used to take easily twenty, thirty minutes a day. Now it's, you know, less than two or three minutes tops. And the other nice thing is, like, the rep... The workflow that generates this report, it's visible to my entire team, and anybody can edit it. Right? And so we keep iterating on this over time, and we keep making it better and better. So, the value of the workflow is compounding, which is really great. Okay. Alright. We're gonna get to the last part of the the, of the discussion today, which is really on, you know, how do we say human as leaders. And, you know, here's how I broadly think about it. I think AI is really good at the work that gets me ready and helps me prepare, and it's a really good assistant in that way, and I'm happy to hand off, you know, a bunch of things that used to take me a ton of time to do. But my team at the end of the day, it's important that they know that, hey. When I make a call, I'm the one standing behind it. Or when I'm talking to them, it's actually me. So, you know, I really think delegate preparation, but not your accountability and voice. And I think, honestly, this is probably the part that's easiest to get wrong right now because the better these tools get, I think it's really tempting to let them do a little more and a little more and a little more. And so I wanna spend just a few minutes on where, you know, like, I draw the line, you know, in my day to day. And so the first is don't... I mean, this is obvious, but don't take AI's output at face value. If you think back to the loan example that we showed a couple of slides ago, right, the brief skill said, hey. You're you're gonna miss your SLA. Well, before I go and take that to leadership, I'm probably gonna check those numbers myself. Right? Because if that's wrong, it's it's not AI walking into the room with a bad number. That's me. Right? And so important to know that at the end of the day, you are always accountable. The second is don't, you know, outsource your perspective. I think AI will happily give you an opinion on anything, and it's usually a very reasonable one. And that's what makes it very easy to adopt. But if someone on my team is asking me why we're doing something, you know, my answer can't be, you know, that's what Anthropic told me to do. Right? I think that's a very dangerous place to be. So I have a pretty simple test, and it's really like, hey, could I defend this decision without AI's output in front of me? If I can't, I I don't think I'm ready to make that call. And the last one here is around delegating accountability. I think there are moments, especially as a leader, that it really does matter that something comes from you. Right? If I'm giving difficult feedback or I'm asking my team to change directions, they need to know that I've thought it through, and it's really important that that carries my unique voice. So I've I've tried to encode this into a kind of like a decision tree of of sorts here of, you know, when to hand off to AI and when to delegate. And, this is my own personal view, but I think, like, broadly, I ask myself two questions. The first is, you know, does this thing need my unique voice or presence? You know, going back to what we talked about in the last slide. If I'm giving somebody hard feedback, telling my team about a big change, if I'm recognizing somebody for doing really great work, those need to come directly from me. Right? AI can help me prepare for these things, but I don't think it can be the one that's showing up in those situations. If the answer to the first question is no, the second question I ask is, hey. Can I actually evaluate what AI is giving back to me? If the answer is yes, okay. I think that's actually a good case to hand off to AI. I'll check the results afterwards. The one caveat here is our ability to critically evaluate things is not permanent. Right? So if we stop staying close to the work, we essentially lose or atrophy the skill that makes the review useful in the first place. So I think you have to be careful about getting in enough reps that use day sharp. And then I think if the answer to the second question is no, that's when you really need to slow down. And I think this is the one that people miss a lot is, like, you know, the less we know about something, the more tempting it is to distrust AI because it it can sound really confident, but that's also when you're least able to catch a mistake. So in these cases, you know, my recommendation is either a build context first so you can you can understand or get somebody with actual expertise to go take a look, especially if you're going to use that output to make a decision. I think it's really critical, that you do either of these steps. And, you know, in my day to day, I do a lot of data science analysis. I am not a data scientist, and so I could not tell you if the queries are correct and whatnot. And so if I were to use that analysis to make a decision, I'm gonna clearly label it to others and say... Tell them, hey. This was generated by AI. I have not validated it. We should probably get a data scientist to go and look at this first, before we make a decision. So I think having that disclaimer is really important, and making sure that you are remaining accountable in all cases. Alright. Let's, like, make this real with, like, a... An actual, situation, you know, we all deal with as leaders. You know, where does AI fit in something like performance management? I I actually do think there is a place for for using AI in in something as human as, like, a difficult performing... Performance conversation, and I think it it it is really in that preparation phase. Like, I will use... I think it's it's actually a good use of AI to, like, make sure your feedback is specific and actionable, and, you know, you can even role play how might this feedback land, and I think you can better calibrate, you know, the things you're trying to say and find examples and and whatnot. So there's there's definitely value in this preparation step. But when it comes time to actually having the conversation, say, with your direct report, this one I think you you kinda have to shut the laptop off, and and really, like, you can't be going into that situation with the script. Right? That's a very human conversation that has to happen, and that's where things can actually change. Right? Your your report might tell you something you didn't know. Maybe there's something going on at home or their personal life, or maybe there's something you don't have visibility into. And I think when those things happen, you have to be willing to change your read in real time and, you know, always remember there's another person on the other end that's sort of watching and listening and reacting to you. And so, you know, leaning too heavily in AI in these situations, actually, I think can be very counterproductive to your to your goal as a leader. So I think understanding where it fits into, you know, all of these aspects of leadership that we have to do is really important. And so I think, for me, I largely keep AI at that preparation phase. But when it comes time to doing the thing that requires me to be with... In front of another human, I think that that should remain as is. Alright. I wanna spend a minute talking about AI writing. There's a lot of debate about this. So I I wanna share, you know, where I've landed, on this one. I'll start with the story first. You know, a couple of months ago, I I was leaning in heavily into using AI to accelerate all of my writing, and so, like, you could just pass really rough thoughts, and AI will give you a super polished, you know, response. And so I was, like, slacking my team, a bunch of AI, you know, enhanced, messages and whatnot. And then I I remember I got feedback from an engineer on my team, someone I really trust and respect. And, essentially, his feedback was, you know, there's a disparity between, the amount of time it takes me to read the message and the amount of time it took you to, like, create this message. And so, like, that disparity is almost, in some sense, a little bit disrespectful. And, you know, we've all felt this. You... I'm sure you've all received, you know, an AI slop PRD or something. And, you know, your first instinct is probably not to read that. Right? You're probably not gonna want to read that because a human hasn't, you know, put their touch on it. And so, that that really stuck with me, and it made me realize, you know, your writing reveals a lot about who you are as a person, your word choices, the punctuation, that you use, the emojis you use, like, that's a part of your personality as a leader. When you farm that out to AI, in some sense, you know, you're you're almost stripping away some of your humanity. And so, I'm very careful now. Any message to my team, I'll just write it. You know? I think the time savings you get with AI, they don't outweigh sort of the cost of losing kind of your own personality as a leader. And so, the second part here is that, the writing process itself is a critical part of how we think. Right? And it's actually a critical part of how we work out what we think. And, you know, the first draft of anything, right, whether it's like a strategy doc or a design doc, right, it's likely gonna have a lot of things that don't add up. Right? And and that's friction. Right? And that friction is telling you that your thinking, is not finished. And the danger of handing, you know, AI to, danger of handing, this over to AI at this point is that you get something polished before you've actually worked through the reasoning in your head. And there's a difference between, you know, reading a draft and thinking, you know, yeah, that sounds right and actually working out the argument yourself. Right? So for anything important, I will, definitely write my position first, and then use AI to push on it. Right? What's unclear? What am I missing? Where might somebody disagree? I think AI is also great for research and feedback, you know, and and and also doing, you know, grammar and editing. But the position, like, what you send should be a position you've worked through. Right? In words, you're you're willing to stand behind. And I've linked here... There's a really good article by... I hope I'm saying I'm saying his name correctly. Eric Grunewald, that is is really thought provoking on on AI writing, which I encourage you guys to check out. Alright. Cool. Well, we are we are at the end here. Peter, I'm happy to open up for Q&A I hope you all have found, the talk useful. Yeah. So here, that was great. This is Jot and NoSound. Fantastic presentation. I think, the reaction, everyone is asking for your skills and, copy the reporting. Super useful, so dense with with amazing insight. Yeah, I think so many good questions. I think we'll probably have time for maybe five or six. One was really interesting. I think, gee palazzolo. You're talking about your own process, for your skills and memory and and improving your own work as a leader. How do you cultivate that culture with your team of staying human while working with agents? So how do you dispel that that culture to your team and not gatekeep all the amazing things that you do? Yeah. That's a great question. I I think first and foremost, we have a pretty open and transparent, culture in my team. We're constantly sharing, like, here's my workflow, here's what I'm using, here's the skills I'm using. And the beautiful thing is, like, you can kinda take that and make it your own, and everyone on the team kind of does that. So I think having the open and transparent sort of culture is super important, and I think, you know, we used to have, like, weekly AI sessions where we would just demonstrate, here's what I'm doing, here's what I'm using. So I think that's what I would recommend is just, like, building that into your culture of just sharing openly. I don't think I would want everybody to do things exactly the way that I do. I think it's important that you kind of tailor it and personalize it to how you work as a leader. But I think the ideas are kind of the things that I do wanna share is, like, here's sort of, like, how you can think about it and how to use it. But really, you know, the unlock comes from experimentation. Right? You actually have to sit down and try and work things out. Like, the first versions of all of the skills I showed you today were terrible. It's like, they got better because I kept at it and I kept iterating at it. It's great. I I I have to ask this question because I think it's, you know, it it... It's not within what I wrote down, but it just came in from Ian. Just your thoughts you shared, you know, you have you have kids growing up. How how are we supposed to train the next generation of engineers? When they're growing up right now, that might be, AI is doing all the heavy lifting. They'll never train that muscle. Do you have any sort of perspective on kind of the next generation? And maybe even, you know, as you look to expand your own team in hiring and what you're looking for today rather than what maybe you were looking for as engineering leader three, four years ago. Yeah. That's a deep question. And I I honestly... I'm... I kinda wrestle with it. I think there is some element of, like, you have to do the thing to to be an expert at the thing. And so, AI does shortcut some of that learning. So I think for my own kids, you know, I I wanna be cautious about when I would introduce that to them. I think it is important that they learn the core skills. And I think for engineering, it's no different. Like, how are you going to know how to do an expert code review if you'd never written code? I think there is an element of that that has to has to happen. But maybe it's no different than, you know, when calculators were introduced. Right? I'm sure people had the same the same thoughts and the same questions. But, yeah, I do suspect, like, a lot of things are gonna shift up a level and the, like, the abstraction layers that we're operating at today, we're gonna move up a level. And so things will change and things will shift. But I think for now, while we're kind of in this transition period, definitely for junior engineers, we do encourage them to spend time, you know, not leaning on a... AI as heavily, like, actually doing the work, struggling with it a little bit, really getting your hands dirty. I think that's much more important in the early phases. And then, you know, once you know what you're doing, yeah, by all means, you can use AI to really accelerate. But, yeah, it's a it's a tough one. I think it's something the the industry is trying to figure out. Right. And and obviously, you know, it's it's really lifted, productivity you showed, sick... 60 increase in VRs. And how how do you personally as an engineering leader keep up with the the increased output of your teams? This is a question, that came in. My feeling is that just keeping track of things done and making sure teams are on the right track takes so much more time given their increased output. How do you manage that? Yeah. I mean, a lot... I mean, there's a ton of workflows I was not able to share just due to time, but that's a huge part of just... That's where AI actually can be really helpful is just collating, you know, all of the things that are happening, giving you a really curated look at in terms of what's changing, what are the key things to keep track of. I think without AI, I don't know how I don't know how it would be possible. I would be spending a lot of time digging through a lot of channels. So I do think, yeah, there's definitely... That's where I... It would be hard to imagine keeping up without having some of these tools, you know, at your disposal. Right. And and that's sort of a segue into question from Paul. How how can we really tell and and maybe speak to your experience at at Stripe? And, how can we really tell the AI increases in engineering teams productivity? Higher LLC, more PRs, if they are sloppy, they don't seem to try to count for the business. So just kind of a... Maybe how do you measure, you know, quality outside of just the pure output numbers? Yeah. I am not a... Yeah. I would agree. I don't think PRs merged is really the metric for are we moving the business. Right? There's there's definitely something further upstream that is is actually of value, which is, you know, could be the features we're shipping. But, really, it's like, what's the customer and business impact we're able to drive? And that I can say for certain, you know, we are seeing a definite increase, at least within Stripe. You know, what is possible and the speed at which we can create is dramatically different than it was six months ago. But, yeah, I think just relying on an, like, PRs merged as, like, a... The sole metric is definitely not the right... That's not the right one. We have a couple questions, going back to kinda your own personal, style working with Claude. Any strategies that have been proven useful in overcoming the syncopancy problem with AI when using it as a sparring partner? I thought that one was an interesting one. Yeah. You gotta be really careful not to tilt towards, like, what your preferred option is and holding, like, a really neutral, ground in the prompt, I think, is really, like... To me, is really kind of the key, because I think, like, I showed in that one slide where you're kind of... If you're looking for reassurance of the decision you're trying to make, that's where you kind of can get some of that syncency, in the output. So it's kind of trying to hold that neutral neutral ground, I think, is important. And and another one was... I thought was a really good question that came in from Veronica. It was about memory and kinda how do you maintain this, you know, on a month by month basis. If your AI keeps drawing on the same memory of your usual context and past examples, doesn't that bias it towards rehashing the past instead of reasoning freshly about new problems? Yeah. I mean, you can put a lot of different things into, like, I'll I'll use Anthropic Claude as an example. Like, you can put a lot of different things in the memory, and so, like, I kinda just give it a high level of, like, who I am, what do I do, what do my teams do, you know, like I... This isn't a full dossier about me and my life. It's it's really just kinda high level things to orient it, towards like what what would matter for, okay, an engineering leader at Stripe, you know, like... So I think I would say you don't need to go overboard with this. You kinda just need to give it enough clues that it it it can get a better sense of, okay, this is what this individual, is gonna care about and this is how I can tailor the app. And Claude's pretty smart, you know, at figuring that out once you give it a little a little nudge. Great. Well, with that, I think we are running in. We wanted to end ten minutes early. Sudheer, thank you so much for joining. That was an incredible session. Again, to everyone on the line, we'll be sending a survey out, so please provide your feedback. We collectively in the emojis on the screen. Thank you, Sudheer, for, for the presentation. It was incredibly useful for for any engineering leader today. Up next in the series, like I mentioned before, next week, next Tuesday, we have Jean, Sue, and Kate Hudson who are gonna be talking about DRI in your career, looking at different perspectives of, find... Navigating your career in today's AI era. Should be an amazing session. Again, sign up for that if you have not. But thank you so much, Sudheer. We're gonna wrap with that. Have a great rest of your day. And, again, we're gonna follow-up with the recording, and I think a few folks were asking for the slide deck itself. I think we can make that happen with the PDF version. But, again, thank you so much, Sudheer. Thank you, everyone, Yeah. on the line for your questions, and have a great rest of your day. Thank you.