- 2 months ago
The constraints that defined how businesses operate for decades are changing. Organizations seeing the biggest breakthroughs are willing to unlearn what made them successful and rebuild from the ground up with agentic AI rather than bolting it onto existing workflows. AWS took that approach across the company, from rethinking how engineers develop software to transforming how marketing organizations goes to market.
Category
🤖
TechTranscript
00:15Good afternoon, it's so nice to be here. It's been such a good week at Viva Tech, like energy,
00:21startups, developers, all of the folks together, I have to say, but like today's the best day,
00:26this is the marketing day, the CMO day, so I'm so delighted to be here today as it goes forward.
00:32Now I'm going to jump in. So I have the, I guess the privilege, I sit next to thousands of
00:39developers
00:39every day, and I get to watch how they have embraced agentic AI, right? And that is the topic
00:46du jour, as it were. And by witnessing what they've been able to accomplish with agents, it's really
00:52caused us across the company at AWS to think differently about what's possible in our
00:57discipline. And I actually think other than coding, where agents have really become such a huge force,
01:03I think marketing's next. We have the next level of opportunity and challenge to embrace agents and
01:10what it looks like in our disciplines. And I also think it's really important for all of us in this
01:15room and around in this industry to be really hands-on, because it matters, because it is going
01:20to change how we work individually, but also our teams as well. Now, as I said, the developers are
01:27really that tip of the spear. So kind of learning from what they've done and applying it to AWS,
01:32about a year ago, I set a challenge for my organization to become the most AI forward marketing
01:39team on the planet. Because frankly, AWS provides agents and AI technology to the world. So our marketing
01:46is the brand face of that, right? And if we're not AI forward, why are we credible to be a
01:51purveyor of AI
01:52services for others? So it's important that we embrace this in a really big way. Well, and the other
01:56privilege at AWS we have as well is we have unlimited access to the very best AI technology. And so
02:03we have
02:04the pleasure of being able to experiment, to try it and learn, and nothing gets hold back. And so we
02:09pushed forward and jumped in all the way to see what we could learn. And today, I really want to
02:15talk
02:15about some of the very specific examples and what it's looked like, both within marketing and across
02:19the company as well, since we've had the opportunity and the privilege to kind of be a leader in how
02:24we
02:24take it and what we learn. And by the way, not everything works. And there is no silver bullet.
02:29There's no, you know, crystal ball. We just have to try and learn and iterate and get better
02:33as we do that. Now, the kind of the first phase we started on, both within marketing, but across
02:40Amazon as well. And actually, this is similar to what I've seen outside in the companies,
02:44customers I work with, is we started by looking at how we work and then adding agents into those
02:50processes, right? Helping it do some content editing or doing some reviewing or pulling
02:55together some system information. And actually, we're able to achieve in those cases, you know,
02:5920, 30% improvements, better outcomes, more effective, different cases like that, which is
03:04awesome, by the way. Who doesn't want to be 20 to 30% more effective? So saw some really nice
03:09gains in
03:10those places. But as we did this again and again, we also realized that we could bring an agent into
03:16one part of the workflow and it would speed that up or make it more efficient, but then it would
03:20hit
03:20the next bottleneck, right? So it improved a little, but it wasn't really giving us these big
03:25breakthroughs. And so we caused us to have to step back and think differently of like, rather than just
03:30using agents within our workflow, what if we fundamentally shifted our entire workflow now that
03:36we understood how agents worked and how they could affect our processes and took much more of an agent
03:42native approach as well? And again, watching from my colleagues in the development team take a similar
03:48path, I could see that pattern and apply it to us. And I'm going to walk into some very specific
03:54examples to make it real, because I feel like people talk about AI at a very high level, and I
03:57just want
03:58to make it real and show you exactly what we did. But before I get there, as we did this
04:03process,
04:03kind of a mental model or an approach kind of emerged. And so there's three things that we went
04:09through in this journey that I wanted to share as kind of context before we jump into the examples.
04:14The first one is just the absolute, absolute importance of learning, being hands-on. And that
04:21means me every single day, that means my team. Because the reality is this is changing a lot,
04:26innovation's moving fast, and if you don't have your own personal experience, you don't get it,
04:31right? And so getting my own like, aha moment when I saw what was unbelievably possible now with
04:37agents, and also having those moments of fear, right? I, you know, I've been doing marketing a
04:42long time, and I have this way that I love to do positioning and messaging that I've been refining
04:46for like two decades, and I'm so proud of it. And I use my positioning messaging framework, and I gave
04:51it to the agent, and it didn't perform very well. The content coming out wasn't very good. And the team
04:56took it, and they came back, and they're like, actually, we wrote, we gave it like pages and
05:00pages of content. We got it context. We taught it why that positioning and messaging was right,
05:04not just give them the positioning and messaging. We saw much better outcomes. And for me, that was
05:09a moment of like, oh, that's a totally different way to do it, and I have to learn that new
05:13way.
05:13And if you don't have those personal experiences, it's hard to figure out how to best use agents with
05:18the humans together and what that combination is. The next part, and something I probably talk to my team
05:23more about than anything, is the importance of unlearning. And if anyone was here on Wednesday
05:28and heard Jeff Bezos talk, he talked about this idea of dream to build cycle, and that I think in
05:34my career, so many places I've given up on trying something because it's impractical, it's too expensive,
05:39it's too hard. But now, with agents, new things are possible. And so I kind of have to unlearn
05:45these constraints that used to be true and question those assumptions and allow myself to
05:50fundamentally think differently about what now is possible, right? Most of our processes and how we
05:55work today is around human constraint, but agents change that. And so every day, I try to have a
06:00moment where I recognize and speak to an unlearning of like, that used to not be true, but maybe today
06:05it is, and embrace that. And then, of course, with that, both the learning and the unlearning, you're
06:11then in a situation where you can actually really step back and think about how do we fundamentally
06:15change the way we work in light of that, right? Once we've kind of let go of those sacred cows
06:20and
06:20thought, allowed ourselves to think a little bit more freely, we can really lean into that. And
06:24of course, that means lots of change. And so going through that process and having that experience
06:28hands-on and directly makes it easier to get to that phase. But let me make it real, some very
06:34specific examples of what we've seen, what we've done, what we've learned. Now, like I keep saying,
06:40coding and development is kind of the tip of the spear. And really, you know,
06:43agents are perfect for coding, right? It is a form of language. And so in AWS, we have a product
06:50called Amazon Bedrock, and that's where we run all of the world's models and customers use it to get
06:54their access to models. And in 2025, we needed to have a major re-architecture of this big service
07:01that customers were using, because we needed to scale it to the next level. And the development team
07:06went away and they did the estimate and they said it's going to be about 30 developers and it's take
07:10us
07:10about 18 months. And that's not unreasonable for a big architecture. It seemed reasonable.
07:15But then the team took the challenge of what if we worked completely different and embedded agents
07:20and used them in a very fundamental way, not just code along with the agents, but actually give the
07:25agents the specification and let them go. Kind of get out of code with and babysit mode and into
07:31empower and checkpoint mode from a development perspective. And something really remarkable
07:37happened. They actually did this entire re-architecture of a massive AWS service with
07:43six developers in two and a half months, right? So from 30 developers in 18 months to six developers
07:49in two and a half months. And we didn't know. We really didn't know as we went, we jumped into
07:53this
07:54foray and tried and see what would happen. So we were all shocked and delighted and amazed. And it
08:01really gave us a whole new worldview of what was possible by proving it there, right? And in a wonderful
08:06way, by the way, the re-architecture worked great. It did more AI tokens in one quarter than we had
08:11the previous two years. So the re-architecture was good too, not just fast. But I saw that,
08:18I watched that, and I thought, what does that mean for marketing? How do we embrace that same kind of
08:23exponential approach? Oh, by the way, when we rolled out the learnings from this project across
08:28Amazon, across all the developers in Amazon, we're seeing almost four and a half times productivity
08:33gains. So this is really, really meaningful if you think about the core of what we do as a company.
08:38But as we get into some marketing examples, first I'm going to start actually one where it's really
08:42about bringing an agent into an existing process, not completely retooling it, because that's an
08:46important step for learning and understanding how it works, but also maybe just sufficient for that use
08:52case. So in the case of localization, right? AWS across marketing, we publish in 16 different
08:57languages, about four and a half billion words a year that we get to translate and send out to
09:03our customers, and we want to do it quickly, right? We want all languages to get all the information as
09:07fast as possible so it's globally accessible. And the process today we had is a machine translation
09:14of the content, and then we hand it to these very talented and skilled linguists, and they get the
09:19pleasure of oftentimes really rewriting the content, because the machine translation misses a lot of local
09:25nuance, like kind of native speaking, some of the brand tone. So it ends up being a pretty heavy lift.
09:31And because of that, our linguists are spending a lot of time, and our service level agreements for
09:35this are usually translation happens in days, all 16 languages takes multiple weeks, right? And that's
09:41just a constraint we've gotten used to. That's how we've always worked. But if I think about these
09:46talented linguists, right, they're spending most of their day on more the mechanical part, fixing the
09:50basics, not doing more, leaning into more native storytelling, which is where I'd like them to be spending
09:55their time. So we brought in agents and changed our process. So we stuck with the machine translation.
10:01That's still an important role. We didn't replace that. But we added in a new layer where we brought
10:05an agent in. So the agent actually takes the machine translated content, and then the AI,
10:11the agent, applies the context, the more local nuance, maybe some kind of native ways of speaking
10:18within the language, make sure that the brand tone is correct. It does all of that. So it's almost
10:23like one single step between those two things. It's an immediate handoff, and it can get done.
10:28And then the lovely part is that when it gets handed to the linguist, who ultimately needs to look at
10:33it
10:33and kind of make it perfect, they actually get to spend their time going back and really thinking
10:39about how do we tell this story in a more local way? So it really resonates with that audience and
10:44that language or that dialect even in that front. And so they've been able to spend their training on
10:49what they really want to do versus a lot of the mechanical parts of it as well. And in working
10:53with a partner of our State Street Bank, one of our great customers, as they were able to localize
10:58more, get higher outputs from their localization in a similar way, they're seeing incredible
11:02performance, right? Over almost 300% better sign-up rates from their website and almost double the
11:08industry standard for click-throughs when they have that excellent line of last mile localization.
11:13So it's not just about making the process faster and making the linguist better used, but it's also about
11:17just better marketing outcomes as we did that. Now, the biggest one I think we have, so much of
11:23marketing is about content, right? So I have to talk about this one. And this is probably where
11:27we've seen the biggest impact and where we've really challenged ourselves to step back and think
11:32differently. So I just picked webpages for one example of the content work that we do. We have about
11:3710,000 webpages we publish each year. And the process for any webpage on any of our properties is the
11:43assembly process, the validation process, the review process, and then it pushes live.
11:47And so for the marketers on my team, the process looks a little bit like they request a new page
11:53from the campaign brief. It gets sent to the digital marketing team and the ops team, and they triage it,
12:00and then they put it in the backlog. It's back and forth to track when it's going to get done.
12:04And then they spend about four hours per page assembling, pulling from the CMS system,
12:10pulling the content, making sure that everything works. And then after that four hours, it goes
12:14into the review process, and it makes sure that all the links work, that the images are accessible,
12:18kind of all the quality pieces of it. And then if any of those things fail, it goes back to
12:22the
12:22cycle again. So that's the process. Again, that's what we've been doing. It's fine. It's how we've been
12:27working. But it also means so many of my talented, skilled marketers are mechanically putting pages
12:33together rather than doing beautiful storytelling. And I would love to shift that ratio. So this is
12:38a place where we said, let's actually just step back and think completely differently. If we were
12:43to write content development with an agent-native approach, what would that look like? And that's
12:47one of our key work streams that we're focused on. So as we've done that, we looked around and we
12:53built a lot of our own technology at AWS. But in this case, we actually decided to partner with a
12:57company, Gradial, for their content agents and it's worked really, really well. And now my marketing team
13:02can actually just use natural language, just prompt in exactly what they're trying to get
13:07done. The marketer just describes kind of the need in plain language. And the agent actually
13:12takes that. It learns from what they're like, identifies the things. It actually pulls directly
13:16from the CMS system, which components it needs, pulls the content from the content capabilities,
13:21the images, and it actually puts that together. But it doesn't stop there. We actually also at the
13:27same time do the review process. So it used to be this linear process. Now it's a simultaneous
13:32process. So we're constructing and assembling the page. And the agent's also then also checking
13:36it against our, make sure from an SEO perspective, make sure everything's accessible, make sure it's
13:40on the brand standards. And it's happening at the same time. So we literally collapsed that into one
13:46single thing. And then it can publish directly from that CMS process as well. So taking something
13:51that was a very linear and manual process and literally made it a single agentic process in a
13:56really remarkable way. The outcome is, well, previously it took us, there's just the assembly
14:01and through the review process over four hours per page. It's now taking us 10 minutes to get this
14:06done. So if you do the math, that's 95% better. It's a really, really wonderful improvement, which
14:12means now for my folks on my team who usually maybe spend 15 to 20% of their time thinking
14:17about
14:17great storytelling, differentiated value prop campaign strategy, they can actually shift from about 15% to
14:23about 80%, right? That's a huge difference in how we're deploying our human capacity with our agent
14:30capacity as well. Now, just to show a different example, because it's not all in marketing, of
14:35course, across the company, we're doing lots of things with agents. And another wonderful, like big
14:39shift we've experienced is in the legal department. One part of our Amazon legal organization, literally
14:45across the world, we have a million employees, always scanning for different compliance, different
14:50regulations, and they change all over the world. And that previously has been a very manual process,
14:55highly skilled legal professionals scanning around, trying to make sure we're staying current,
14:58never falling behind, because if something changes and we're not keeping up, that's business risk,
15:02right? So it's both manual, but it exposes us to risk. So they, again, step back, how could we use
15:07agents to do something in a much more effective way and use our human capital more effectively as well?
15:13In this case, our legal team decided to use Amazon Quick, which is our agentic teammate offering that
15:18Amazon provides and sells. And it automated that whole process, right? It goes, it's actually
15:24constantly running an autonomous agent, constantly scanning the landscape for different compliance
15:28changes, different regulation changes, different aspects that might have a legal implication for
15:32Amazon. It automatically synthesizes that. It automatically sends regular updates to the
15:38appropriate legal team that it might have consequence upon. And so the whole legal team,
15:42instead of going in from a very manual and reactive perspective, is now getting it in a proactive
15:48way. And again, not only has this shifted the work experience for these folks, but it's also
15:53really reduced business risk, right? We're not going to find ourselves at risk of being months and months
15:57behind an important compliance topic, rather than have it proactively pushed to us in that way.
16:03So the best, this is my favorite, so I hate it best for last, which is around data and insights.
16:09As marketers,
16:10all of us know, it's very hard to have exactly what you need to know right on your fingertips. Like,
16:16I know what my campaign did. I know why it did that. I know how I want to adjust it
16:19based on that.
16:20When I came to AWS about a year and a half ago now, I'm just going to be honest. We're
16:26not going to
16:26share this. We're friends here. But we had 22 data science products, individual products we'd built
16:32over three years. We had 4,000 tables across 15 schemas sitting in our data warehouse. We had 862
16:39measurement documents that we had performed over two years looking at outputs. And the team had a
16:45SLA. They would turn around ad hoc. If you had a question like, hey, how'd my campaign do? It would
16:50take us 28 days to perform that, right? Like, we're a tech company, you guys. 28 days. We also were
16:56maintaining 1,500 individual dashboards that all needed care and feeding. And there was a backlog on
17:03this team of over 2,000 dashboard requests. So you can imagine how unhappy this team was.
17:10Not the most happy team. So, and as we looked at it and we were doing triages of how do
17:14we prioritize
17:15and how do we, and I was like, this is never going to work. Like, it's just not going to
17:19work. We have
17:20to fundamentally think differently. So we did. This was one of my favorite examples because it's unlocked
17:25so much. So the team took this challenge. And very real here, we have one product manager,
17:32we have one data scientist, and we had one developer who took the challenge. And they said,
17:36great, we're going to think completely differently. We let them go off next to outside the team and just
17:40figure it out. And they just inverted the whole process. Like, we're not going to wait and ask for
17:46each campaign. We're just going to create something that can be used for all of that. And they created
17:50this tool called AIR, our agentic intelligence and recommendation system. And true story, I have
17:55it bookmarked in my browser called AIR, the magical tool, because it really, really is. That's how I
17:59feel about it. And it simply is a prompt that any of the marketers can go in and ask any
18:05question.
18:06Like, I want to grease my pipeline and AI in France. What do you recommend? And it has both the
18:12what
18:13and the why. So you could actually kind of get this insights of what happened, but why it happened.
18:19Because we also loaded not just the data, but the causal relationships, the analysis our economists
18:23and our data science did, and they can correlate that with the agent. And so you're not just telling
18:28my campaign did X, you're saying because it did Y, and here's what you could do to make it better.
18:34So it really is an unbelievable unlock. And I've had so many people across the team email me and
18:40Slack me about how much this is unlocked for them. They're spending hours just asking questions
18:44because it's so useful. And it's something we've never been able to write at our fingertips. So
18:49really, really, really big unlock on the data science aspect of it.
18:54So bringing it home in terms of kind of some key themes of what's spanned across and the learnings
18:59that my team and I have had. First, it's so important to get hands-on. I said it before,
19:04I'm going to say it again. Like, you need a culture of experimentation. AI is changing fast.
19:09How it works is changing fast. If you're not hands-on, you're not going to be in a way to
19:13lead your teams. You're not going to know how it works for real. And it's easy to read it.
19:17It's different to do it. And so everyone has to experiment. We have learning days on my
19:21organization. We have our marketing academy, always training, kind of just keeping it as a
19:25learning culture on that part of it. And part of that is also knowing that a lot of the stuff
19:29you
19:29try is going to fail. And that's okay, right? Failure is necessary on the path to mastery. And so making
19:34sure that that's part of the journey because you learn so much from what doesn't work.
19:38And then, of course, really thinking about the ways that agents work and the way that humans work,
19:44right? And I know a lot of folks on the panels earlier talked about this as well. But as I
19:49look
19:49at it, like humans are beautiful storytellers. We connect to that human experience and we can
19:54unlock those ideas. And agents can take that and help us tell it in lots and lots of ways and
20:00in
20:00more personalized ways and more language ways. And so finding that combination. And every day I really
20:06think about how do I get more of my human capacity on great marketing, great storytelling, differentiated
20:12content, not the sea of sameness, and use agents to take away some of that drudgery. But understanding
20:17that is really important. And then the last is really giving yourself and your team the permission
20:24to step back and rethink. Like literally start with a blank piece of paper. And sometimes it'll be like
20:30my experience in our data science and data, but be a place where there's no other choice. It's just
20:35not working. So why not throw it away and start over? But other areas like content, it'll be maybe
20:40a little incremental. You'll start with adding agents. And then over time, you're like, you know
20:43what? I think we know enough to really reframe this completely. But again, creating an organization
20:48where there's psychological safety, where fear is removed so people can try and learn and iterate
20:54because there's no silver bullet. It really is about just trying it and getting better. And by
20:58the way, once you get something working, the AI gets better. And so you can actually come back and
21:02retool it as well. So hopefully some real world perspectives on how we've worked. I'd say the one
21:08thing that everyone in this room can do that's changed my job and my life probably is Amazon Quick.
21:14It's my personal, authentic teammate. You can use it. You don't have to have your IT approve it. It just
21:19works. But it keeps everything safe and secure. So it's fine for work. And this is what all my team
21:25lives in. It's fire across the whole organization because it's just your thought partner. It's your
21:30campaign helper. It's all the things that you might need. It's your calendar manager if you want that
21:34as well. But again, thank you so much. Thanks for letting me share my insights and hopefully gave
21:38you some things to think about.
Comments