Watch now | Recording of AURA - Anna Bergevin & Shane Gibson 21st September 2026

Hey, we're live now. So hey, Anna, welcome to the Aura session. This is our 3rd one. It's a, it's a standing joke with all the Aura guests now that, uh, we can't tell when Substack's actually going to start it. So we're sitting there humming to ourselves, kind of looking up going, has it started? Has it started? And then takes different time every time. But how are you? It's been a month, been a month of craziness from what I can tell. What's happening in your world at the moment?

Oh, what isn't happening? You know, just, I'm just, you know, building things at home. My husband and I are working on a home operating system, like task household management system built on a Mac Mini. So that's been really fun. I'm brainstorming some hypotheses around context layer at the hospital and like, how are we going to build that? What might work for our certain use case? Then just the ongoing work. We switched electronic health records at the hospital this year, and so we have to figure out how to stitch longitudinal data for our patients. As you can imagine, there are many— it's live clinically in the health record for a certain number of years, so there's no clinical disruption, but from a— you want to analyze anything about multiple years of what's happened with individual patients or many patients. We're a pediatric hospital, so patients who were children before, now they're adults, but if you want to look at cohorts and trends. We got a lot of data products to to build this year and figuring out where we're going to start. So kind of some big architecture things going on right now, uh, and just, it's busy. It's just very busy. How about you?

Uh, yeah, just busy building at the moment for some reason. So tend to, I've always gone through kind of cycles where I could sit back, take long, long views, right? Look at the market, look where it's going, figure out where we need to go, uh, and then you go into build cycles, you get new customers, right? And there's always, no matter how much you've built out of the box, right, there's always There's always something. We've changed our business model this year, so that's kind of changed a little bit the way we work. But yeah, definitely been far more in the build cycle over the last couple of months than the stepping back and looking. But it's kind of one of the reasons I love the Aura sessions. And the other thing I love is the things that go across the 4 of you, right? So I was talking to Chris Gamble last time, and it was all about data migrations.

Yeah.

Don't know how we got onto it, right? It was like, oh, data migrations from— old platforms to new platforms, you just raise the data migration. Context is king is always the, the one we're struggling with, um, so that always comes back. Um, but the interesting one that I picked up from you was that in my journey, I started off with a personal OS experiment, uh, because that's how I kind of started playing around with GenAI tools, Claude building and automation. And then I kind of moved and took some of those patterns into the day job data side of things. Whereas for you, you kind of started off using it for writing content, using it for your work, and then you moved into personal OS, right? So you kind of did it the opposite way.

Yeah, I definitely did. I mean, we had a certain amount. My husband has always been a— my husband's a software engineer, backend software engineer architect type. And so he's always been on like every Christmas, the joke is which home automation stuff do you want for Christmas? Because he's always trying to automate things at our house. Um, uh, and so I think that I didn't really know how to help it automate with like this mental load, remembering, planning, management, like at scale. Like I knew it could help with little tasks. We use Claude and Gemini. We have both, um, because we're, we're ridiculous. And they, they're useful for different things. But, um, so I'd use it for small things. Like my husband, one day our water softener— not water softener, our water heater went out and he went downstairs and took a picture of something that was like the pipe was looking weird. He's like, what is this? And it was like, oh yeah, that's a common failure mode. So like we've been using AI as a household, as a family researching. We're going to buy an air fryer. Which air fryer should we buy? Help me compare these things. We use it for a lot, but now it's like, how do we remind us our car registration is due in July?

Yeah.

you need to schedule an appointment to go get it in and get the safety inspection, get your car registered. So that's what we're working on now. And I think the other thing is that became really big. How long ago was OpenClaw? Back towards the beginning of the calendar year that everyone was like going crazy with OpenClaw. And I was like, you guys are doing a lot more trusting this agent and giving it access to things. And like, yeah, you're able to do cool things, but I am not touching that with a 10-foot pole. You guys can play with that. And by you guys, I mean wider internet. And then I kind of had this aha moment recently about like a way we might build this that still is privacy-centric and access-limited but gives us help still with that mental load planning part. So, uh, we're, we're getting close to an MVP. So that's my fun side gig. And then yeah, the content stuff is always ongoing.

And you know, one of the things I found is, uh, I learned so many patterns, right, that I could actually then, you know, reapply in a different way when I did it for a data work. Um, and then I found that data governance, right, actually, because like you, it was my own personal information, yeah, I wanted a way for it to be able to interact with it remotely. Uh, so I ended up going with Signal because I use Signal to communicate with family, and I kind of— it's the one I trust the most, right, to have encryption and privacy.

Yeah.

As much as I can trust anything. So I'm like, okay, I'm gonna go hook that on. I did open, like you, I didn't Yeah, I did open Claw for about half an hour, put it, installed it, looked what it did, and went, yeah, no, that's coming off. Um, but then I ended up getting Claw to look at it and say, what features could I use? Um, and then same with emails, right? Uh, I don't let it send emails out at all, and I don't let it access emails directly. Like, I actually, I found trying to get, uh, my personal OS to connect to my Gmail system in a secure way where it couldn't actually do anything. was horrendous, right? Actually, you always had to delegate it far much— far too many, um, kind of, uh, permissions than I ever wanted to give it. So right now I'm experimenting with the idea of these, uh, open source kit where it basically connects to your mail, brings it down, and it has a local database of your email content, right?

Yeah.

And so I'm like, yeah, yeah, I can handle that, I can handle that coming down. Um, and then the next one I did was task management, which I still haven't fixed, right? So I've been through probably 5 to 6 iterations of going back and saying, task management's not working for me. It's me, it's not you, but we have to do something differently, right? And so I've just found an open source one that somebody published and I've picked that up and I'm seeing if that works for me. So again, it's each part of the system I'm having to tweak as a human, right? And go, that's not working, or how do we refine that? What about you? Have you kind of built it up in the idea of Lego blocks? Pick the thing that's most important for you first?

Yeah, so what it's doing is, um, so I looked into that. You sent me, or somebody else sent me, a GitHub repo for that local database version. Somebody else, there's an MCP. I had a couple people recommend things. I looked into them, but we're still sticking with— we gave, um, um, our dogs are named after Sherlock Holmes characters, um, Enola, Sherlock, and Mycroft. So we named our OS Watson. Oh, that's cute. which if we ever open source Parallels, I can't call Watson or IBM is going to come for me, but like we call it Watson OS. And so we got Watson an email address of its own. And so what we're doing is setting up deterministic email filters of like relevant things. So things from my kids' schools, doctor reminders, basically I did an analysis of my inbox and said, what are the things that tend to have relevant information I kind of need to act on or be reminded are coming up? And then I did some things that are just known facts, meaning I need to register my kids for summer camp in February. That's when most summer camps open and we need a reminder then. So I did kind of a manual curation of kind of some metadata and a YAML file, and then it's going to have access to our Google Calendar, meaning we just share our calendars with that same Gmail address that it has access to. So it can access those forwarded emails that it's going to get, it can access the calendar, and then It dedupes everything and then its real brain is knowing what's relevant to whom. So like the school emails, I think mostly come to me, but depending on which kid it's for, we have one kid in elementary school, one kid in high school, so they have quite different kinds of action items. And it's not even per kid, it's actually things like who's responsible for registering for camps, who's responsible for making sure their lunch funds are topped up or whatever. It might be. So it kind of takes that, it says, what is this thing about? And then it routes it to who owns that responsibility area. And then it creates a digest every day that it sends to my husband or to myself. And then the task management we did not pull in. So the task management is going to go, um, from that digest, you can add it to Apple Reminders or to Google Calendar. So you can either time block something or put a reminder. Because what I just said, what we kind of reasoned, we meaning Claude and me, reasoned was like, I don't want this to become another task manager. Its job is an orchestrator and a noticer of what needs to be done. And then if I can't do it immediately that morning, it's— some of it's just informational, like, don't forget you have to take your daughter to the doctor appointment today at 1. Like, it's that heads up. And some of it is like task-oriented. Hook it into your existing task system, which Reminders works really well for my husband who has ADHD. So I'm like, we'll go with what system works for him. And Google Calendar is what works for me. I tend to be a time blocker. And so between those two, So the morning is going to be read the digest, action on anything that needs to be actioned, meaning add it to a reminder or do it right now, make this phone call later, schedule this, that appointment, whatever. And then it wasn't— it doesn't remind me again. At the end of the week, it's going to do a weekly summary that says these are all the things you're supposed to do. Do you need to add this again? Did any of these not get added, linger too long? So we will— that's our current hypothesis. Hopefully get it into an MVP in October and then we'll just see how it runs. And our goal is to look at it kind of once a quarter and tweak.

Yeah, and I think that's one of the key things, right, is that constant iteration. So going back and doing a retro effectively, right? And one of the things I found interesting, especially using Claude, is I can ask it to do the retro because it's logging every interaction I have with it. And so I can go back and say, hey, you know, tell me what you think we need to change. And yeah, typically comes back and says you actually need to do your tasks. And I'm like, Yeah, no shit, Sherlock. But yeah, there are some other ones there I didn't know. The other one I use it for is I use it for a personal CRM. So it effectively stores all my interactions. And then so when we have customer meetings, it comes back and says, oh, in the last customer meeting, you talked about this. These were the tasks you meant to have done. These are the ones you have done. Your customer meeting is coming up in a couple of days. You probably want to get those other ones done. But here was the themes and here's what the themes are likely to be next time, right? So again, it's that reminder, like a good executive assistant, right, would remind you of these things. So, um, no shit, Watson doesn't work. Uh, it does if you try and use IBM AI, I heard, but hey, that's, uh, that's a bad thing for me to say. So, um, so yeah, I find that really interesting. And then the other thing is, again, because we're using it for 2 of us, I use a lot more than Katie, but, uh, that personalization you talk about, right, where each of you have a different interaction style, a different way you want to be communicated with, a different way you want to communicate back, different levels of tools that you want to use, that you're comfortable with, your favorites, where you go, or you spend all your time. And you think about doing that in an organization that has 1,000 people, right? Like, we're not there yet, right? We're not at that hyper-personalization. But I reckon that's where we're going to end up, right?

Yep.

I reckon that we're going to have systems that personalize around each persona to begin with, so group of people, uh, and then we're going to go down to personalization for individuals, which is what we want, right? You know, maybe I do want a dashboard, maybe I do want an AI chat to my data. Uh, maybe I do want a report sent to me once a week. We've never, at least in Beyond Data, we never got to that stage where the systems were tailored for the individuals and the way they wanted to behave. We were always forcing, ah, here's the PDF emailed once a week to you because, uh, well, we decided that's what everybody wanted. Um, I mean, what do you see? Do you see that hyper-personalization coming eventually? And like, if you had to guess, how many years? Or weeks or months or days.

I think that the big vendors will take longer because they're trying— they have to solve for everyone, right? And like solving for yourself is the easiest. That's why building this for myself is easier. Solving for a persona, like you said, is the next step. Solving for an organization's culture is maybe the next step out, you know? And so like I worked at Sourcegraph briefly, and they were globally remote and async. It was the most insane Slack environment I've ever seen because everything was just writing culture first, async first. So Slack was insane. Then I joined ResMed and later this hospital, older medtech healthcare organizations. They had teams and there was like, I was like, where are, what are the channels I'm missing? Like, this place is like crickets. And they're like, oh, it's all in email and in meetings. And I'm like, okay, this will never work. So it's like the culture, the current state, the current tech stack, it all is very dependent. What's really interesting on the personalization front is like our tools for that are currently super blunt, meaning Copilot custom instructions, Microsoft's got you, you can give it instructions for chat. Oh, sorry, it's saying my internet's being weird. Sorry about that. It's saying it's like you can give it, sorry, custom instructions for chat. So the way it interacts with you and then you give it custom instructions for how to write your emails. Like this is how I like to communicate. So you see Microsoft doing these little like personalization touchpoints. So it'll be interesting to see what becomes available. But the thing that they're not doing yet is very effectively is, hey, other salespeople like you tend to approach their sales cycle this way. or they have a personal CRM. Do you want me to adopt that pattern for you? Or someone in your company's built this tool or this MCP server, or IT has now rolled this out. This is a canonical data context data product, if those become a thing. What does that mean? And how do we— What I know from enabling AI among a non-data AI role, because at the hospital I'm on the Copilot rollout, I don't know, enablement committee. We call them champions. Is that people don't— their use case, their use— what's powerful for them is very different and role-dependent. And how motivated they are to develop their own hypotheses and test things out is extremely— You got the power spiky users, and you've got kind of some playing around a little bit hypothesis testers. And then you got the long tail of like, Yeah, I just use it to natural language schedule meetings and to summarize notes. And you're like, it can do a lot more than that, but like they need you to show them. So it's like, how are we as organizations going to like create things, create things that work, scale it, get people to use the things that work? And then how are we going to create systems that help those enablers, administrators say, hey, this seems to be taking off in this area, you should roll this out for everyone or make it canonical and make the admin of it. Just like with data products, I have this thing in data products I call that I'm trying to promote in my organization, which is gold data products and community-led data products. Meaning superpower analyst over here has created something very useful. They want to share with other people. I won't let them share their table directly because permissions, governance, etc. But I'm like, cool, hand me your code. You can always pull a request and modify it. I'm not going to stand by it as in this table will never go down, but other people are welcome to use it. If it gets useful enough to the organization, we're going to bring it into the gold tier. We're going to make sure it has testing, observability, it doesn't go down, all the things. I feel like there's almost that pattern needed for context and MCP servers and these things that people do. It's like, how do we let these power users experiment, share their patterns, make things available? And then eventually IT goes, hey, this is looking really useful and we don't want Brad to leave and the whole thing falls apart. So we're going to take it over. So like, I think that's the pattern we need to develop with organizations, is expansion and promotion.

And the problem is we've never really managed decentralization or democratization properly, right? So, I mean, years ago I did the trusted versus community, and then, yeah, originally it was individuals could only create it for themselves, they couldn't share it. But then we saw some value there, so we shared it to small groups, and then they ended up becoming quasi-data warehouses, right? The groups got bigger, and then if we didn't let them share it, then they just pushed it somewhere else, right? And it was daisy-chained on daisy chain, and we And we lost all the lineage of who was using what. You know, one table would go down and everybody would complain and you're like, where the hell is that table? And it was 3 tables down from the analyst who created it.

Exactly.

But there was value there. And same thing, I've just been a week up in Scotland. So Nigel's been doing a lot of the information product development and using our harness and I've just been watching what he's committing. And because we want to reuse it time and time again, we effectively have a bunch of shareable components. So date picker is shareable, right? But one of the feedbacks from one of our customers was in the date picker, it said last year, and when they clicked it, they expected to get all of 2025, but actually it was the last 12 months. So the date picker said last 3 months, last 6 months, last year, and they quite rightly said that should say last 12 months. So Nigel's like, yep, no problem, blah, blah, blah, it fixes it. But it basically forked the date picker straight away into that information product. And so I'll now go back in and pick it up, push it back to our shared components, and then roll it out across every information product, right? Because it's good feedback, or it's like, that just makes sense. But if you think about that way to share, get innovation, bring it back in and centralize it, it's really, really hard, right? And I don't think we've done it with BI. I don't think we've done it with data. You know, when I see dbt stuff, we decentralized it and democratized it, but we never brought back the sharing piece, right? The bringing it back in and making it trusted. So I don't think we have the patterns for it. And then with AI and GenAI, it makes it even harder, right? So especially when you talk about personas, because you— that idea of champions, right? That idea of you want to say, well, you're power analysts, right? So we know you're going to find 101 ways of using it. But, you know, you're a frontline staff member in a ward. If you find something that makes your job really valuable and quicker— sorry, more quicker, right? Because it's always valuable or more effective. Then actually, we want to roll that out to personas like you, but how do we do that, right? How do we encourage you to experiment when you are really busy? And then how do we find personas and find a way of experimenting or incrementing it out to them iteratively in a way that's semi-safe, right? And that— those ways of working, I haven't seen a lot.

No, it doesn't exist. And I think that's maybe one of the new jobs that needs to exist, which is like, What I want, I mean, with data products, I own, my team owns the data warehouse. We can build in telemetry. We can say you can build community-led data products in this spot in the warehouse so we can observe what's happening. We can at least observe the lineage even if we're not guaranteeing it to be stable, all those things. But when it comes to AI enablement, AI data products, it's like things are breaking off into silos, things are happening. Microsoft doesn't give me— it tells me who's using— I actually don't have access to the dashboard here, but I did at ResMed. It's like who's using Copilot and in which applications. The heaviest one by far is Teams because people are using it for AI summaries. Well, what I want it to do is be like, oh, this little person over here has created a tool they're using every single day in this space, in this way. I want a little paragraph summary of what it is and who else it might be useful to. And that needs to go to someone whose job is to find cool pilots and scale them, right? And that could be a data PM, an AI PM, or program manager, architect. I don't know what you want to call these people, but their job is to go interview that person and be like, show me what you're working on. This is cool. Okay, what if we took this over and we built it so you didn't have to maintain it, but other people could use it? And, or we, you know, central IT, central data has more infrastructure available to us, right? So you may have built it in a kind of a janky way, but it's working for you. And we're going, what if I built this in a real environment? What if I did this? You had the same user experience, or you have ideas, you get that like magic between you and the person going riffing together, right? Like, oh, I see what you did, but what if we also did this? Or people like this, and you interview and you do that product work. I mean, it's product work. These are mini applications. And the problem is enterprise IT does not have product managers. We have data product managers now, which may sit in IT, may sit somewhere else. But like, I've been feeling that pain recently in my org is like, I want someone in core IT who's running enterprise apps as products, right, that they're extending and building upon. And instead, what you have is architects, if you're so lucky to have enterprise architecture, who are looking at the whole system and how things interoperate. But that's not the same as exploring the opportunity spaces like a product manager. So Hmm. I think orgs need all of it.

Here's the question, right? So if we think about data products, information products, right, and we go, in the past, they've always been very BI-centric, they've always been dashboardy, right? They've always been BI reporty kind of thing. And we move more to products that actually start actions, right? They actually either assist somebody to say, here's the next best action you should take, or actually, I'm going to take it for you with your permission, right? I'm going to do this, yes, no kind of thing. thing, um, then actually, isn't that really still sit in a data and information team, right? Like, why, why do these mini apps— we keep— I keep thinking them as software products, right? I keep thinking they need to be in a software team and an IT team. They need to be driven by software product managers, uh, and then I go back and actually go, well, actually, maybe not, right? Maybe these are just data that's incurring an action. You know, the number of times we've built data platforms and we always end up building a data entry screen, like we're doing it at the moment, reference data, right? We have a bunch of reference data for a customer. We used to always build it in Google Sheets, right? Because we had the integration with GCP, Google Cloud, it was out of the box. We'd already built it. So we're like, you maintain that Google Sheet, we'll treat it as a view of a table. We'll lock it down a little bit so you can't change the schema and then we'll fix it when you do. And that gave us a cheap, easy way of doing reference data, right? Right now we're now building the apps. We're saying, well, actually, why don't we give you something that's more magical than a sheet? We'll give you a mini app. We're now starting to do that write-back, that creation, that mastering, that write-back, which we always tried not to in a data platform. I'm thinking the data world is going to start extending out into these places that we never really wanted to go. Master data management. I want to come back and ask you about master records for patient. Right, an individual review, changeover in a minute. But yeah, I think we're going to— I think data people have an opportunity right now to move into spaces and solve problems that they used to leave for other teams, which in a large organization like you is going to cause problems, right? Because we had the natural IT boundary, give me a server, give me a service, and then we have data people doing on top of it. It's a bit like DevOps when the software engineers used to start deploying their own containers, right, their own ecosystems. and they got value out of that in terms of reducing the value stream, reducing the friction of that work. I'm going to think we're going to see data people do that. What do you think?

Oh, this is like coming up so much right now at the hospital. There are so many processes where someone is entering something into a form. There's a research form database tool called REDCap, somehow is funded by the government. It relates to healthcare research, but it has become at our hospital also a place that people can— you can create a form so easily and it creates a little database. What that means for us is every time someone wants us to bring that data into the data warehouse, every API response is unique to that little mini app they've essentially built. And it's not always the best tool for that. Like, it works. Like, they're— it's fully featured enough anyway. There are reasons they're building there, but I'm going, What if you built it as a Microsoft Form with some validation that goes into a Dataverse table and then those are more database table-like and then I can bring them into the warehouse? That was a pattern I had a hypothesis about maybe 6 months ago when my data PMs and I were like, can we get more people to adopt this pattern? It'll make it— the ingestion instead of an API call is within our Azure because we're in Azure there and we're in Azure in Databricks. Probably there's an easier problem to solve. I was talking to our leader of engineering about what if we went this route. Then it came up a use case recently where one of our top engineers built a Databricks app, which is like a Streamlit app. I don't know if he actually used Streamlit, but a mini app with basically forms and it records directly into Databricks. There's just no ETL at all. And like, I'm like, this is amazing, this is awesome, let's do more of this. And my leader of engineering is going, it's just a test, I don't know if we want to get in here, I don't know if we want to play in this space. And I'm like, but if we don't play in this space, then IT plays in this space, and then we have to react to whatever they do. And so like, I feel like it's— there's a huge opportunity space here, and I feel like data teams are going to wrestle with, do they even want— that's not— we haven't managed source systems, right? We've managed replicas, and we're essentially stepping into the source system space. And do we want to do that? And I don't know. I can tell you the PM, it feels easy to me. Sorry, I'm getting a little choppy.

That's right. Yeah, I think also we— yeah, yeah. Then we get into backup recovery levels of service that we, we treat differently, right? And volumes of users, concurrency, all those things that our data platform tools didn't typically serve well, um, but they got better at. So it'll be interesting to see, see where that goes. Um, I just want to— oh, I just want to pick up where, uh, some comments from Johnny. Johnny Winter, good to see you, Johnny. Good to see you, Nick, and Serene, and Rupesh. Um, so Johnny's saying, uh, for a while I've always promoted the idea of a data consumer choice. So a well-curated, governed set of data, but let the user interact with it where they feel the most comfortable. Um, let's wait for Anna to come back. So— oh, you're there. Oh, good. Um, what's your thoughts around that? Like, you know, have we really allowed our consumers to consume the data and information we provide in the way they best prefer, or Have we always given them cookie-cutter stuff that was easier for us?

Probably a mix. I can't figure out why my camera's on, but I'm here and I can hear you, so we'll just keep talking. One of my jobs as a product manager on a data platform is to talk about user experience. One thing I've learned is We have to think about what experience they want consuming the data and what patterns we want to support. And by that, I mean, my data engineers do not love the idea of people consuming the raw data as it lands, right? People are always like, but that's faster. Can I just get that data? It's already in there. And we're like, oh, we really don't want to put a dependency straight on that table. Should we create a view? And then you can hit that view. And then later, if we need to move things, we can move things. And so And then some people don't know SQL. So you're going, oh, can I create a view? And you're like, select star and just download to CSV. Or should I do that as a pass-through, as a table within a Tableau dashboard? Should it be a chatbot that you just put in some parameters and it spits back out the CSV export? I think we just have to think about what's the— what I'm always trying to say is, why do you want that? And why do you want it in a spreadsheet, right? Tell me a little more. What are you planning on doing with it? They're like, well, then I upload it to this other system. I'm like, okay, well, let's talk about that system. Why should you be in the loop loop, or if they're like, oh, I like to create reports and I only know how to do that in Excel. Okay, well, let's talk about how we keep that fresh or working for you, or can we— I think it gets into the who wants to consume data, what are their skills, whose skills are going to evolve, whose skills do we not expect to evolve. If you have a title of analyst and you've typically been in a drag-and-drop reporting tool and now I'm saying, well, you got to learn SQL and AI makes it easier than ever to write SQL. That's different than me saying, random person in finance who works in the FP&A tool and in your own spreadsheets, come learn the warehouse, come use this tool. Why? The only reason they would learn it is to pull data from us. Is there a different way we can get them their data? I think it's a holistic question of who are the people, what are their current skills, why do they want to consume it that way, and trying to say what is the ideal least fraught pattern to support this process long term. And sometimes that's gonna include people needing to change behavior, and sometimes it includes us innovating in new delivery methods we haven't thought about.

Yeah. And observability is the big problem a lot of the time. So, you know, when I, when I talk to a customer and they always say, can you give me the Excel export to Excel button? And I'm like, Yep, we can add that. It's really, really simple. But why? You know, what are you going to do? Because if you're doing something outside of the value stream, we should just codify that, right? We should make that easy for you. We should automate it. So if you tell me what you're trying to achieve, then we can look to how easy is it to build that in so you don't have to do all that manual work. And a lot of the times the answer is, I don't know, right? They want to grab the data to play with it, to explore it, to understand it. And once it's in Excel, you know, we have no observability of what they're doing. When we had our BI tools, you know, they were really, really poor at observability. Like, we could give self-service creation of reports and dashboards, we could see what they created, but we couldn't see why they were creating it. With the GenAI tools, in theory, you know, they're typing in text now a lot of the times to ask it to do something or have a conversation with it, and that's the gold, right? We talked ages ago about the the transcripts of meetings being the context that we never get back again. And so we want to store that raw context and then summarize it in many different ways for the tasks we want to do. And it's the same with the work doing. But, you know, in a non-enterprise environment, people are using Claude and it's a local copy effectively, right? So we have no visibility of the conversations. And there's that gold. I don't know if it's different in an enterprise. And Johnny Winter saying that the back to our point about observability of who's doing what ages ago, but the level of observability is definitely there in Microsoft land for that kind of pattern. But the issue is they build their shadow IT outside of the platform you're observing. Because you're in Microsoft land as well, so, like, if I was using Copilot or whatever the Microsoft GenAI tooling is, is it logging everything that a user asks? And is that sitting back in a massive log store that I could mine and See every interaction, every piece of text that was asked, every answer, every reasoning path that the whatever LLM they're using used to answer that question? Is that level of observability baked into those enterprise platforms at the moment or not?

I know it's in there because it's in there for compliance reasons. Um, my thing's being slow again. Let me just toggle this camera back off and I'll just talk. Um, yeah, it's in there for like compliancy reasons, and so I know it can be accessed, but then you have the question of have you set that expectation with the user that they know it's going to be accessed and mined in that way? Um, like your chat experience with your Your local LLM, while you may abstractly know IT too, but you don't think they are other than looking for, I don't know, really illegal behavior. You know, they have some sort of like cybersecurity, they have things like that, but you know they're not looking at the substance of your email to understand your behavior. And if you do find out they're doing that, it's kind of like keystroke tracking. It feels really invasive. So if we wanted to mine this behavior, I feel like there's a ton of questions over ethics and consent and do people understand? Are we de-identifying it? Even if we de-identify it technically, are they still identifiable based on the content or the substance of their interactions? Then will people stop using Copilot in the same ways because they're afraid someone could read what they're writing, like an actual human might read it versus it being lightly observed? I'm more interested in— and the other thing is transcripts, which are a goldmine, the retention policy is not great. And so I've been really interested more recently, and I'm testing a pattern, and so is one of my PMs, of personal context layer, personal knowledge graph, meaning how I set patterns where I'm like, I worked on this topic for a couple of days. I'm going to create a synthesis for future me and stash it. And so it's not storing all the raw context, but it's storing a synthesis and fairly detailed summary of where I ended at this end of spike of discovery or what I learned. And so can we— and that if we enabled a behavior like that where you're contributing to your personal context or team context, right, then it's a more known consensual behavior to publish out knowledge or publish out context. And then it's more just about, like, how do you make that easy? How do you make that an easy button? Like, hey, we sense you got to the end of kind of a topic you've been working on for 3 weeks. We've done an automated synthesis of everything you worked on and created this article. Can we publish this to the team knowledge base? So, like, it's low effort, and it's like, yeah, you can publish that. And then when it's published, either publish saying, you know, this was auto-generated, or you— or the person that says, I actually, yes, but I want to do a deep dive, and then they publish it under their name. Like, that difference between auto-generated metadata and manually drafted curated context. So I think it could be really useful to have tools that are like promoting and saying, is it okay if I push the synthesis? And then knowing whether or not the person just said yes, or they said yes, and I'm going to take the time to read through it top to bottom and make sure it's accurate. And those are 2 different kinds of publishing out. So anyway, I have a lot of hypotheses and context and not a lot of conclusions yet, Shane.

Same, same here. We're all learning as we're doing, right? I mean, one of the things I keep coming back to right now is that use case around Spirit Airlines where the company got liquidated. So it went under, got liquidated. There was a bidding war for all the context of the organization. So all their emails, all their Slack chats, every piece of conversation.

Wow.

And a recruitment company, I think, bid $7.5 million, something like that, for it, but Google won it for $10 million. Now, in theory, what it said was that the data was going to be cleansed before it was given to Google. And yeah, it probably will be, but you think about that, right? You're an employee of Spirit Airlines that's been sitting on Slack, slacking off your boss, and potentially a recruitment company could have got access to that. as part of that liquidation.

Yeah.

Um, and what rights do you have as an employee? Because I don't know about you, but every time I ever signed an employee contract, it pretty much said everything that I thought about even when I was asleep was owned by the company. Um, and you know, anything you do in a corporate system is owned by them. So I suppose legally they had access to all those Slack conversations and those emails that you probably didn't want anybody to see. So I think we've got that problem. The second one is this, again, this experimentation. So, you know, as Nigel and I are working, because we're remote, you know, he's in New Zealand, I'm in the UK, we really, we catch up, you know, probably about an hour a week in video time. And we always work asynchronously. We always have done that for 8 years, even when we lived 15 minutes away, because that's how we designed the company. But now we're using the GenAI tools a lot more. Yeah, we started this pattern where I would get it to do a brief effectively, uh, when I got to end of an iteration cycle, and I'd send it to him.

Yeah.

But he wouldn't really read it because it— I don't know about you, but when you get these AI-generated briefs, my brain turns off, right? Like, I, I don't know why, but I start reading them and then I just go to sleep. And so he was like, yeah, just stop sending them to me. And I'm like, yeah, okay, but what do we need, right? What is that handoff, um, when we can't make it a human handoff? because of the time frame and just the way we work. So now I'm wondering, is actually, is the AI-generated brief, uh, still valuable but for the ability to go back and ask questions? Yeah, like, or summaries, like, okay, end of the week, give me a summary of all the things Nigel did and all the things I did and, you know, where the overlaps are. And I suppose one of the challenges is because at the moment we're, you know, we're writing code or help getting assisted to write code, we have a Git repository. And so the context is always there. We can always go in and say, well, what, what PRs were released this week for that part of the product? Uh, why, why did we ask for that to be built, right? And we can always kind of reverse engineer the context of what we did to a degree. If we're not writing software and we don't have a Git repository and we don't have that track changes on the things we're creating, So it's an organization with just communications. Does that still exist? Like you said, you know, would anybody be comfortable mining all the emails and all the transcripts from the meetings for the whole organization to get back those summaries when you need it? And like you, I don't think we've solved that problem yet. And I think that one's going to bite us in the bum for a while until it becomes such a big problem that somebody does solve it.

Yeah, I think it's more— yeah, I I have the same feeling of what an AI-generated summary, why it feels so bland, it can feel so generic. But if that summary is not for Nigel, but it's for Nigel's copilot, meaning as Nigel goes about his day, its ability to go, hey, Shane touched on this yesterday. Here's what Shane said. It only brings the party or some of this relevant to you. Then it's, it's like, it's, it's, um, consensually created bidirectional context sharing, but not necessarily with an expectation that the other human read it, more that the other human has access to and the other human's Copilot, um, can retrieve relevance just like it retrieves relevance from— I mean, Copilot, Microsoft Copilot does this all the time. I'll be talking about another topic and it will go, this actually really relates Like if it heard this conversation you and I are having, it would be like, this actually relates a lot to the hospital context layer work you did 3 weeks ago, or this relates a lot to your data catalog migration you guys are working on. How will this context piece mix in with the metadata pipelines and this other bit? Like it has this ability to see relevance and retrieve it and just mention it. So imagine a system where we do provide daily summaries of our work, And so we're aware they're happening, we're able to see it, we're able to have a setting going, yeah, I've watched enough of those summaries to not worry you're going to tell it I, you know, complained about my boss to my friend at work. I know you're just going to summarize relevant context. I actually no longer check those, but they can keep going. And then you know that you're also receiving those for your coworker, but no one's expected to read them. They're more about creating shared context instead of right now Microsoft is And Claude and ChatGPT Enterprise, any of these tools can only work on personal memory, meaning it's a conversation with you. And how do we start to create ways that we share organizational context with each other? And then how do we on top of that run processes that are like, send it to the governance team being like, hey, there's been a lot of work recently happening around this. I feel like there's some new concept terms we could you know, certify. There's some wiki articles we could add to the data catalog around how this works. And they go mining for the context because they're incentivized to clean it up. And then they take it back to the people who are authoritative and say, hey, based on your context summaries, we want to put this in the catalog and we want to, you know, have you as the steward on it and say, is that okay? So it's like you're not asking people to generate context they don't care to generate. You're basically just creating shared memory, and then the people who care to curate or responsible for curating can create even more useful summaries of context that then humans might read, right? Which is like, okay, I need to— I want to get visit volumes, and my ChatGPT goes, uh, do you mean visit volumes or encounter volumes? Those aren't the same thing. It goes, I don't know. And it goes, here's what the catalog says the differences between those. Oh, I really want visit volumes. Okay. It has this inclusion exclusion criteria. Are you okay with that? Yes, I agree with that. Okay, here's the number, right? Like, so how do you create these context stashes, retrievals, curations? It's not all one thing. This is why I think context layer is like a silly thing for a vendor to call anything, because I'm like, it's an architecture layer. It is a whole web of things we're gonna have to create. How about you call it whatever it is, which is what kind of context are you auto-creating, allowing people to create in your specific tool, but don't act like you've solved this for the entire enterprise because that's ridiculous.

Yeah, yeah, yeah, yeah. I call it the context plane because I think, yes, you know, it goes across everything and there's so many things in there that we kind of know some of it and we kind of don't know others. And then, you know, personal memory within the LLM you're using, right? How do we share that? And I keep coming back to team behaviors. You know, when I saw high-performing teams, data teams that uh, were bigger than one squad, right? They were bigger than 5 to 9 people. Uh, there was always these, uh, these people that were the glue, you know, the people that would typically sit in another squad, uh, but their sole job would be to go out and educate the other squads on things they didn't know, uh, and then sit and learn what those squads were doing and find something reusable and then communicate that with the other squads, you know, and go, hey, Squad A over there, you know, uh, they've done this, right? You look like you've got the same problem. Why didn't you just reuse that? And those human glue people, you know, they were really, really valuable. And so, yeah, I can see the LLMs or the GenAI tooling becoming an assisted AI version of that. Yeah. Because if it has observability, it should say, well, look, you know, the team over there has kind of dealt with that problem. I think what they've done is the same context as you. Let's check the context to make sure it is the same and it is a pattern, not an anti-pattern. And then, hey, here's what they've written, or here's how they've worked, here's how they've solved that problem, here's the patterns they've been using, would you like to experiment with them and see if they work for you? Because humans couldn't have that observability, right, unless you're on a factory floor and you were looking down at the machines working in a specific way. When we got away from that and we're working in knowledge, you know, we all work slightly differently, and it wasn't observed how that what value stream worked. So I'm with you. I think, you know, I think we'll get there, but I don't think we're there yet. And I don't think the vendor buzzwashing is helping us, but we get it every wave, right? So we're not surprised.

No, I mean, it's fine. They can keep doing it. They're marketers and they're doing a job, but I'm just, I laugh every time. But I think it's very, if you think about those glue people, it's really interesting because people are like, oh, the glue people, they're going to be redundant. Like, no, they're just going to be more informed, can cover more projects. Like right now, my project managers, program managers, the EPMO, they can't possibly cover everything important. There's just not enough of them, and we can't staff more of them. Similar to data, we just have as much headcount as we have. But imagine, um, a program manager who gets context synthesis sent from various groups that are working on shared problem spaces, and they read those or their AI synthesizes and says, all right, we're heading into this cross-functional meeting. What's the pre-read we should send to everyone? You know, the headline of news. All right, this squad accomplished this this week. You guys should be aware of that. And that they can then become the judgment layer on top of that shared context to pull the headlines up, to create a curated pre-read, and then to have better agendas when we have meetings. So I feel like I can— I kind of play that layer right now among my product team, right? I start my weekly product meeting and I'm like, Hey Jasmine, you should know Barry worked on this last week and I think this is really relevant. You guys should have a one-on-one and talk about this. I'm almost like connecting the dots. And another question I'm getting more now, which is really interesting, is, um, people want transparency into our roadmap, uh, the data team roadmap. Data is quite centralized at the hospital, not completely, but quite. And so they're all thinking, we all have a lot of needs. What are you guys working on? Where's my stuff in the queue? Is one question. But another question, a group was like, I really want to see, like, someone the other day I was talking to, and her job is antimicrobial stewardship. So, like, how do we— anyway, and she was like, what do you have in your backlog that mentions antimicrobial? Like, who across the hospital is talking to you about data projects that are in my area? And she's not asking— I mean, sometimes people are like, can I stack rank those? Because I want to— and I'm like, no, you can't stack rank it. Or maybe they can, depending on their authority. But I want them to know so they can go approach that same requester and go, hey, I think we have overlap. Let's do a shared request. Those become better data product requests because we're hitting 4 birds with one stone. But if they can't see what's in my backlog, they don't know when to raise their hand and go, hey, me too. Can I join this discovery? Can I join this conversation? And so, that's the part I'm most interested in solving next is like, how do I help them know what I'm working on so we can pull in more interested stakeholders without bloating? The discovery, but you know what I mean.

And again, it comes back to personalization. So, you know, natural reaction is you have a product backlog, there's a bunch of tickets or things that you could get built. You may use one of the tools that allows the rest of the organization to come in and upvote, you know, like the vendors allow you to do to a degree. But actually, again, that's us forcing our interaction style that's easy for us onto those stakeholders. Coming back to the thing you mentioned before, You know, if there was a data product that actually just looked across those and said, oh, this person's really interested in these things, actually these requests coming through looks like it's in their sphere, right, in their boundary, why don't we notify them that, hey, there's a couple of really interesting things over there that you may be interested in, here's why we think it's interesting for you, go have a look, right? And they can look at it and they go, yeah, nah, or they might go, holy shit, yeah, I want to get into that. And now you're coalescing these groups of people with less effort, right? And then that's where we get value, right? We know when we build a product that is used by more than one person in the right way, then there's value there. So, yeah, again, a bunch of ways we can—

If it's going to slow us down too much, right, and there's an urgency, then it can be like, great, I don't want to include that in the MVP, but I'm going to come talk to you as soon as MVP is up and you can be in our next iteration. Or it might be like, oh yeah, that's an easy one for us to fold in. And that's kind of where the data product work is, is in saying there are multiple versions. We don't just launch a data product and it's done. Or a new request comes in and we're like, oh, 75% of what you're talking about is already done. Here's the next 25.

Yeah.

And so the more visibility we create, it's interesting 'cause I read Your Best Meeting Ever earlier this year and I wrote about it. I saw you'd shared that. But one of the things she mentions is, We invite people to meetings by default because we think the topic is relevant to them. But you don't need to invite them to the meeting just because it's relevant. You could send them a summary, right? Like, have respect for people's time. And sometimes it feels disrespectful to not include them because you're like, oh, this is relevant and they should be aware and I don't want to be rude. But also that ability to be like, I'm aware you're having this meeting. I just want to see the notes after. I have this problem where I can't be in every meeting that's relevant. Enterprise data, I'm too horizontal, there's too much relevance. I am constantly picking which of 1, 3 meetings I'm going to go to for the next 30 minutes. But I often will be like, Microsoft just introduced this new idea of you can follow a meeting, meaning it will send you the summary after even if you don't attend. This ability to say to someone, hey, we're not including you in the meeting, or I put you on— right now it's just required and optional. You can imagine adding people to a calendar invite as a follow, meaning you're not even invited as optional to show up. We're just going to send you the summary after, which is our way of saying we're not even asking you to come. We just know it's relevant. We want you to have the context after, like that ability as an organization to keep the meeting small, keep it effective, but automatically share context after it. That's like a paradigm we don't talk about right now, but I— I would rather not be invited to every meeting where my awareness is needed. I only want to be in the meetings where my active participation is needed.

Yeah, I love that meeting post you did. I, when I was back working as a consultant in enterprise organizations, meetings were the bane of my existence, even though, you know, I was a contract consultant, so I got paid per hour. And, you know, back then I used to get so grumpy with them. I used to tell the customer how many packets of chocolate biscuits they just spent. wasting time having a meeting where we didn't actually decide anything. And, you know, I kind of broke it down. I wrote a post about it, about the 5 things a meeting should— one of 5 types of meetings there are, and that we need to be very specific about what the purpose of the meeting was. And then that meeting should be driven to that outcome.

Yeah.

Are we there to make a decision? Then there's a certain way we need to run it. Are we there to brainstorm? That is a different kind of meeting. It is run a different way. And then the invite everybody, And even now with just the two of us, I find that I CC Nigel into all the emails of our customers because I just want him to know what's going on. And he doesn't CC me in, and I kind of like, oh, I lose the context. But thinking about what you just said, that idea of following is a great idea, right? It's like he doesn't need to read the email. He just needs to have the— know that it happened and that there may or may not be something relevant to him, or he can find it again later. And so maybe, yeah, that idea of a follow for those emails and for those meetings is a really good plan. I'm going to go play with that one because I can see that one changing our way of working quite a lot.

Yeah, it's a very— then you get the whole summary and I don't even always read the whole summary. The way Microsoft does the summary, it'll say where you were mentioned and you can jump to that part and hear that 2-minute clip of when you were mentioned. You're like, oh, that's why it's relevant to me, or I'll just ask Copilot. I'll be like, so it'll send a standard AI summary and then you're able to go, what parts of this are relevant to me? And because you work, if you work with Copilot all day, it knows what I'm not just attending or sending emails about, but like what I'm talking to it about, like where my brain is, what I'm working. It's knowledge work and it has visibility into my inner knowledge because I copilot with it so much.

Mm-hmm.

And so when I say, what about this is relevant to me? It's able, like an assistant who was there taking notes, just to say this one bullet point is really the only thing that's relevant. So I don't even read the whole AI summary. I just, I follow it and then I ask it what's relevant. And then sometimes I jump to that part of the meeting and listen to 5 minutes and go, okay, okay. And then I ask Copilot a few follow-up questions. Was there an action item? Did they decide what to do? And it'll be like, yeah, no one agreed to anything. And I'm like, okay, I'm gonna go send that message right now about the action item. part. So I think they're onto something with this idea of following personalized summaries after. What I want it to do though, right now I get a notification that's just like, the recording is ready and here's the summary. And it's still too much context. What I want it to do is my personalized notification to go, recording is ready. Here's the part I think is relevant to you. And then you could toggle, look at the full summary if you want to. But mostly those notifications, now I'm following too many meetings. I'm getting too many, too much context notifications. So I want someone— I should send someone at Microsoft. This is me as a PM. I'm always like, who's the PM over this? I wish I could send them this 1-minute clip and they could hear my user feedback.

Well, if they've got logging, just keep writing into the chat and in theory it might pop up on their side to say that this user over here— I mean, Google's always great at that with Google Cloud. Like, you know, when we start, you know, when you start using one of their early adopter features, you know, we know that they're closely monitoring how we're using it. And then, um, you know, they come back and they go, oh, look, yeah, why'd you do that? And we're like, well, it's because we wanted to use it for that. And they're like, oh, we hadn't thought about it. So that, that product feedback, right, is, is really good. Um, Yeah. And that idea of like hyper-personalization, let's close it out with that one. So it should actually come back and say, hey, here was a meeting, there are a few things that were interesting, but nothing you need to worry about, read it when you feel like it. Or actually, there are a few things that you're working on right now, so you probably want to go look at it to bundle them in. Or there's some stuff in there, it looks like it's urgent and you're not working on it, and it doesn't look like you're planning to. You need to drop everything and go read it now, right? That, like, like a personal assistant would do, right? You know, hey, you need to go focus on this because, you know, these people are banging at the door. Um, that's how we should be treating these tools. But having said that, I know you need to actually run off to a meeting. Hopefully it's one of those, uh, those good meetings that have a Good intent, good, good lead-up, and good outcome.

I don't even know exactly what I'm heading into, but, uh, you know, 9 AM on a Monday, it's something. Um, this is always so fun. Sorry my camera was wonky this time, Shane, but hopefully, uh, people still got something out of it. Good to see you, Johnny. Good to see you, Nick. Okay, I didn't see who else was on. We have a few friends joined us live, so, uh, always a delight. So I'll see you, uh, next month, if not before.

Excellent. See you next month, and we'll see what the Uh, top of mind after 30 days.

Awesome. Thanks, Shane.

Excellent. All right, catch you later. Bye.