B (2:58)

Well, great to be here. Thanks for having us.

C (3:00)

Thank you so much.

A (3:02)

One of the things that we can do now that I am probably personally causing in the, in the Internet world is we can just build a lot of, a lot of product. I am always out there. Like I was thinking the other day, I was like, I'm going to tweet something where I tell PMs that they should just spend a month saying yes instead of saying no, like, let's ship some features. And I think AI has really accelerated product development, software engineering, getting innovation to the hands of customers. But the problem it has created is we don't know if those products are any, any good. So the, the perennial product problem, which is you can ship things and they can not make the difference that you hope they would make. And so I'm really excited about this conversation because you are going to show us how to use AI and even some of these tools that software engineers or product managers might be familiar with to do really deep, meaningful product analysis. And I spent a lot of time in experimentation and so I love a good conversion rate optimization. So, Tim, we're going to kick it to you to start with. How do we start at the very beginning of analyzing kind of a product and its quality and its usage? Through analyzing conversion rates.

B (4:14)

Yeah, I love this. I think everyone's talking about vibe coding, but no one's really talking about vibe analysis. And we're heading in that direction very quickly. So let's get into it. So before we do anything too technical, I think we want to share a really broad range of examples here, from the really complicated to the like. Actually incredibly simple. I think everyone knows PMs are going to have to become engineers and then we've got a lot of issues where all of you guys are going to have to come in. Our analysts as well. So I think there's a lot we can show here. So we want to start off with just a really simple use case that should be familiar to, I think everyone listening. But I think it illustrates the point there's often the most simple AI tools that can actually have the biggest impact here. I think before we get into the actual demo, I think it's useful just to pause very quickly for a second on the question of what analytics actually is, because I think once you break that down, you get a much clearer view of where these current tools can be most valuable. I think most people jump straight to the nuts and bolts from actually manipulating and crunching data, but actually it's really just a small part of the overall process. And the most important, often most difficult thing is actually just getting the right context in the first place, because that's what separates good analysis from bad. You need to know it to ask the right questions, to come up with the right hypotheses, to know what analyses are even worth doing in the first place. You need to know where the data lives and you need to be able to interpret it all very well. And the new AI tools have just absolutely transformed the process of just getting all that context. You can go as broad as you like, self serve into an unfamiliar topic just incredibly quickly. And that means you can not only deliver quicker analysis, you can just deliver much better analysis too. So to illustrate the point, I want to talk through what sadly I'm guessing is a very familiar situation where a business metric suddenly drops off a cliff and no one's got a clue what to do with it. So I'm actually, I'm going to use a real example from FAIR for this. And this happened to our new customer conversion funnel at the end of last year. So if you've ever worked in growth, everyone's going to know new customers. They're just extremely sensitive to even the tiniest little friction. So almost anything anyone does in the business anywhere can affect these kind of things. Whether it's a sign up flow, a search algorithm, a shipping policy like this all can affect these things. And if you're not careful, you're going to have to decomp the entire business. So let me show you how these things can just be done so much quicker. So imagine this problem lands on my desk. I might look at a couple of just existing dashboards that exist to say what's going on here. And you can see very quickly the issues started in September and there was another drop in December. And it seems to be concentrated in the checkout stage. But beyond that, I've really got no idea what could have actually caused that. So let's start really bored. I'm just going to share my screen. I'm going to start just by doing an enterprise AI search and we use Notion. But frankly, every document system now is going to have an AI system. If they haven't got one yet, it's coming and they are just game changers. So I'm just going to start very simply by asking Notion what happened. Okay? So the only thing I'm going to do, I'm going to just make this more realistic. I'm going to filter the date range. I don't want it cheating and looking at the answer. It's only going to have access to the things I had access to when I actually did this. So I'm going to put it up to the end of April last year, which I run out. Okay? And then we're just going to get that running. So if you think this. All I've asked is what experiments or new features launched between September to December 2024 that could have added friction to the checkout process for new retailers in Europe or North America. And I just said focus on XP docs, PRDs and launch announcements. Okay? So if you think about what I'd have done in the past, I'd have to be crawling through a million documents, doing a load of searches, going through a ton of different slack channels, trying to work out what's going on. And instead, look, I've got straight away a really interesting list of hypotheses to dig into with no word. And you can see it searched across Slack, Notion, Jira and everything else very, very quickly. And if you. Let's just pull out a couple of these. So what's happening? So let's go. So you've got clearly we launched some kind of checkout experiment around this time. That's definitely worth looking in. We've done something with a checkout blocker in Europe. Okay, lots of interesting things to dig into. Now with a couple of clicks I've got a good long list, but I don't really know what these things are. So I've got all the links of extra documents I could go click into. But let's just ask as a starting point, what is Eori? Let's pick one of them. What is Eori? So we'll just ask that it's going to run another little search and give us more things. Now you've got a little bit here, but it's going to bring up a little bit more information to just get a bit more, a bit more detail on this thing. So let's see where that goes. Okay. So very quickly it's saying give me the term what it is and you can kind of see it's okay. It's a regulation that's involved in Europe and someone's done something to start asking for more details, clearly trying to improve checkout and conversion rates and they're trying to bring that one in. But I think this is a great starting point. I've got some detail, but I think what's really interesting here is everyone knows PRD is one part of the story, but between a PLD being written and something going into the code base, a lot can happen. So to actually understand what's going on, you usually need to go one layer data into the actual technical implementation. I want to show you like a quick trick of how I do that. So I think one of the best things about these AI tools is just the ability of someone who's like non technical to access things that they couldn't previously access. And a great example of that is just being able to talk to the product code base. I'm not an engineer, I can't write Kotlin or Swift. I used to be a lawyer, for God's sake. Instead, I can run a deep research against our code base to find out exactly what got implemented for a particular feature and when. Now, I'm going to do this in two different ways. I'm going to do it on ChatGPT, which I think is very simple and anyone can replicate incredibly quickly. Everyone's familiar with it. And I'm going to do it on Cursor, which is a bit more specialized but just incredibly powerful. So I'm going to open up a new chat and I'm going to put it into deep research mode and make sure my GitHub is connected. So all you do, it's not technical to do that. You just need to say yes a few times to get your GitHub connected. The only reason you do it on on Deep Research is just because it's the only way you can actually access it. It's going to search our code base now in exactly the same way it would normally search the web on a Deep Research. So I'm just going to put in a prompt. There we go. Let's just copy that in. Now let me talk a little bit about what this prompt is doing. So I've given it a role. I've said you're a senior staff engineer and you've got expertise in all these different code bases, Kotlin, Swift, Typescript, and you are working at fair. And I've given it a task to say, please conduct a forensic investigation of the code base to Produce a comprehensive time sequence report on all changes to the EORWE collection process at checkout between June 24 and February 25. So just making sure we don't miss anything. And the rest is just a bit of detail as to what I want this to look like. So I've said I want an exact sum, I want a table with all the different PRs and commits, what they've gone into. And I really want it to focus in on the actual impact these commits had on the retailer experience. Like explain it to me in layman's terms. And then I've just put a few requirements in here just to give it a bit more context. So be precise, simple, clear language. Only use GitHub sources.

A (11:29)

I want to call out here. You're using this prompt in the context of sort of what I would call like a business incident, right? New user signups just drop. But this is a prompt that I want the engineers watching or listening to the podcast to really pay attention to. Because if you're in the middle of a, you know, SEV1 incident and you need to trace who did what. I know so many of our engineering teams are looking either manually looking through code, looking at these specialized kind of code gen tools to do this, but probably aren't reaching for something like ChatGPT deep research to just go ahead and do this for you. And if you're a product manager looking to be helpful during an incident, this is maybe a task you can take on on behalf of your engineering team, just to provide some additional context in the background.

B (12:15)

100%. I think this is great for engineers. I think it's great for just getting people to talk better to engineers. I think there's just so much you can do here. So as always, deep researchers asking a few questions. So use discretion. We'll just answer a few of those to make sure we got it. Use discretion. And yes, please. So that'll get it going.

A (12:32)

But no, you prompt just like I do. I just say you pick, you decide, you go, I don't care.

B (12:38)

I think the fact that Pro doesn't ask you these questions make me think it's more to like, make you feel like it's doing it rather than anything else. So that's going to take a bit of time. So while that's running, I want to show you how to do this in cursor, because I think cursor is one of those tools that everyone thinks of for vibe coders, they think of it for engineers, they're not really thinking about what else it can do. And I think for both analysts and non analysts alike, it's an incredible tool. So I think more and more people are talking about the phrase context engineering rather than prompt engineering. I love that it sort of actually explains what we're trying to do here. And for me just Cursor is the ultimate context engine. You can hook it up to mcps. So basically I can hook it up to every single system in our business to get all the data I need and that just makes it such an incredibly good accelerator for getting context and doing analysis. So I actually find increasingly this is getting better results than deep research on GPT. So both are good, both are game changers, but I think this is just a little bit quicker and better. So I'm just going to make sure my MCP is all hooked up and then all I'm going to do is I'm going to drop exactly the same prompt into cursor and we'll see the two running so exactly the same prompt. So just for context, we are not even started on our. Hasn't even got off to the races at all on the ChatGPT and straight away in Cursor we're going and finding it's got a nice to do list. It's saying it's going to search all the right things in GitHub, it's going to then forensic analyze it and we'll just let this run for a little bit. You can see it's already starting to pull in the code and the pull request.

A (14:09)

Everyone, one of the things that I think is interesting to call out is, you know, I've run a lot of product engineering data orgs before engineering certainly day one, what are you doing? You're getting access to all the repos, you're getting set up with GitHub, you're pulling your, your local environment together. I know that data teams often have a similar onboarding because they're working so closely with production data. One of the things I think is going to change or if it hasn't already should change right now is I think product managers and designer onboarding. First seven days has to include access, at least read access to GitHub, getting your local repository pulled down, getting all your MCPS set up because it just code has become now a data source for anybody doing work, not just people writing code. So I look at this and I think leaders out there need to pay attention and rethink basically their onboarding process because you don't want to be in a situation like this and go like can something give me GitHub, like, can I, can I get access?

B (15:08)

It goes even beyond that. Like everyone should have access to every system and it should be from day one. These tools are just the best onboarding accelerators. We've seen it for analysts, we've seen it for engineers. Suddenly people get the context very quickly. Okay, so we're already off. It's summarized everything, it's written as an aspect and we're actually starting to write things out here. So straight away you can see I've got a nice exact summary. It's giving me a few things, but this is what I was most interested in. Okay, so I'm getting a table here for those who can't see my screen. I'm getting a table with every single PR that affected this part of the flow from. Look, it starts in July 24th all the way to. It's still going, but it'll probably go to somewhere like December or February, depending what's going to go with all of these things. Now let's just call out what this is doing. So it's given me an exact link to the specific PR that actually pushed this into the code base, it's giving me the name of it and it's giving me a summary of what it did. It's saying who was affected and it's saying what was the impact on a retail experience. Now if anyone's done this kind of thing, it's so difficult to do and actually pick through all the codes and actually understand what's going on on this. And it can just be incredibly quickly and so very quickly, knowing nothing about this feature, I can already start to get really smart on what happened. And I can see if I dive down here. Yeah, you can see there was an experiment launched in mid September, right in the sweet spot of when this drop first happened. And if I scroll through getting through to looking at December, yeah, you can see it launched all treatment, all users at Bandwidth went live. So this now looks like a really interesting, potentially smoking gun that we can debut into. And so instead of spending days talking to people about all the potential hypotheses, I can now speak to exactly the right colleagues and have a really targeted conversation, an informed conversation right from the off with them to crunch through this problem in a matter of hours rather than weeks here. So even before we've done any data crunching, this can just be absolutely game changing for us.

A (17:03)

Yeah, and it allows you to go a lot deeper than I've been able to do historically on these kinds of analyses. When you're Running these high velocity experimentation programs, you have so many concurrent experiments, you have experiments colliding with rollouts, colliding with just plain launches and just trying to decompose. What was the state of your app on any single day is really challenging. And even if you can do the manual research to get this at a feature level, like, yeah, today we launched the one one one page checkout. I think the real challenge is, well, did we implement it well? Is there anything in there that we should like worry about? Did we exclude any users from that? Like. And so I do think the ability to use code as a detailed source of truth when doing these kinds of forensic analyses really makes the difference in figuring out what's going on with your business.

B (17:52)

And they're getting smart enough to go one lever deep as well. You can ask follow up questions to say, how did it differ for different segments? Are there other ones interested? You can get so much detail just by asking questions on these kind of things without speaking to any engineers.

A (18:05)

And this gives me a little bit of some inspiration on other use cases for querying your code base in GitHub history for events. One of the things that I do very frequently is I do a very similar analysis to this, but I say what is everything that shipped in the last week from the context of a customer and then I use it to write my newsletter. So again like I'm starting to use our code base as a source of truth for our marketing materials. I don't have to proxy through like what was in the PRD or what did a PM write or any of that stuff. I'm just like, just tell me what was in the code in the code commits, because that's what I know went live. It can interpret what the customer facing, experience and intention would be and then you can create these really interesting business and market facing assets out of that. So I just think the ability to query your code base in your GitHub history for any use case, including this one, is really useful.

B (18:53)

Yeah, I love that.

C (18:54)

Great.

A (18:55)

Now what, what do we do after this? So you've identified, you have a conversion rate problem, you've identified maybe a couple sources of the issue, you're going to go talk to your colleagues, you're going to look at the code. How do we actually do some analysis? I know we said we were going to do some vibe analysis and we have seen very few numbers. So Alexa, how do we do actual analysis of data when we've identified a problem or an opportunity we want to go after?

C (19:19)

Yeah. So obviously a quite Classic analytics task I'm going to take us through. We launched a new product feature and we actually want to understand how it did. So I'll take us end to end from understanding how the feature was built, analyzing its performance and then producing a summary that could eventually go to our exec team like Tim kind of touched on without AI, especially the context. Gathering would mean hours spent digging through all the specs and PRDs, writing SQL queries from scratch, and then spending a lot of time writing and editing a doc. So with AI I can pull context similar to what Tim just did directly from the code base. I can generate queries and I can draft a draft a synthesized doc. And so I am going to start sharing my screen.

A (20:11)

And while you pull that up, I have to say people think that why I got into AI in a deep way was because I thought it was so fun to code and it was actually it made my SQL so much less ugly than it used to be. It was like my number one use case. However, many years ago I was like thank God, now I don't have to bother my colleague with my disgusting SQL. I can't bother AI with my horrifying SQL and it can make it a little bit more efficient.

C (20:40)

Yeah, I mean even just chatgpt for the last couple months has been a game changer for SQL queries. The problem with ChatGPT is you had to spend a good amount of time giving context like the exact table names, the exact field names and so using, I mean it's not its sort of most marketed use case, but using cursor, which is what I'm going to show today to actually create, edit, write SQL has been pretty game changing, especially because it's so context aware.

A (21:10)

And I will talk about that.

C (21:11)

Cursor can take three to four minutes to run some queries. I'm going to just kick off this prompt and then I'll explain the context and what I have done. While that's running, I will set the stage. Last month in July we redesigned the signup flow for a new payment method that we have been piloting and this process of signup is successful when a customer links their bank account for the payments. And our old flow had been live for a few months, we had a hypothesis that we could improve it. So we redesigned the flow because this is a pilot, we actually didn't have enough retailers or users to run an A B test. So I just needed to do a pretty straightforward how is this performing before, how is it performing after? Historically again that would have meant a Lot of digging through documentation or more realistically just pinging an engineer to ask questions like, okay, what did we build? Who sees it and why? What front end events are omitted that I can use to analyze this. And while I do work closely with our engineers during the MSPEC phase to figure this out, those details are easy to lose track of. Especially like we're often coming back to analyze things, you know, weeks or even months after the feature launched. I will say that I probably would start with notion AI context building similar to Tim, but we already showed that, so I'm skipping straight to the code base and if we go up to this prompt, my prompts are way less pretty than Tim's. I don't spend a lot of time on them. I feel like with cursor you can always iterate. So I wanted to understand the setup wizard, which is what we called this new flow. I told it to research our code base and I essentially asked who, what, where, when, why? If we go to this answer we can see, okay, it is looking into the code base and I'm not an engineer, I don't really know what this means. But we called this in our code the first run user experience. And it tells me about some flags cannot be sub users. There's just a lot of detail here and it's telling me when users see this flow, what happens during the flow, the order of steps that happen, that's pretty important. If I'm going to analyze a funnel, I need to know in what order did things happen. And then if there is a success event when the setup is complete, then it gives me a bunch of events that I can use to analyze it. This is already such a game changer. In the past I would have leaned on secondhand sources like notion to piece together how it was built with cursor. Like you were saying, I can go straight to the source and have it translated into natural language and that just gives me a lot more confidence because it reflects what's actually live, not what someone remembered to write down.

A (24:02)

One thing I want to call out while you're going to your next step is one of the steps that I see skipped by engineering teams is good event tracking when they release a feature because you know, you start up front in the PRD and you like define a tracking plan and then it gets to implementation and people forget. Should be a front end event, should be back end event. And one of my favorite follow up AI tasks after something has been released or it's in code review is I do a quick prompt And I go, is this is everything appropriately tracked in this feature? And I get either Cursor or Devin to go in and put in all the right events and make sure that the schemas are normalized. So for all the data analysts out there, be annoying and do a PR for your own events on new features so you're not, you know, stuck with what the engineers built for you.

C (24:50)

That inspires me to I can take the NSpec and just put it into any AI tool and say, what front end events do I. Or what events do I need to ask for to be able to measure the success of this effectively? Because right now I'm just doing that in my head. That is not something that I have done.

A (25:05)

Yeah, don't do it in your head. That's the subtitle of How I AI. How I AI. Yes, don't do it in your head.

C (25:13)

So with this next prompt, I, again, not the most sophisticated prompt. I'm just saying I want to understand at a high level how this feature has been performing. And I give the quick context of, you know, our goal is to make it better. That's pretty obvious. I just want to spell that out. And I, like Tim, am giving a fair amount of discretion to the cursor agent. I'm saying, okay, come up with the ideal output fields. I have some ideas, but it's up to you. And then two, I do find that telling it explicitly to create a file, it sometimes forgets to do that and just writes the SQL directly in the conversation sidebar. Use the MCP connection. Like I went through all this trouble to set it up, I want it to use the snowflake MCP connection and then actually QA the file. And that's what's so powerful about this cursor agent. And the snowflake MCP is not only is it writing the SQL, which is what ChatGPT has been doing for me for the last year, it is running it, looking at the output, and then making its own sniff test sense check decisions, which is just so cool. Okay, and then another thing I want to call out as we are running this, the reason why I have a fair amount of confidence that this is going to work relatively quickly is because I and our data team have done a fair amount of work to create what's called a semantic layer. And so first, our amazing data engineering team six months ago decided we were going to create a general company semantic layer. And a semantic layer is essentially just a translation for an LLM of our business. Terms, tables, fields, filters, metrics, et cetera. And AI can look at those files and understand what our tables mean. This general one covered our most used generic tables, orders, items, users, et cetera. And so they connected it to a custom GPT and anyone in the company can go ask pretty basic questions like what was the average order size in Europe last year? And get an answer really quickly. And so that's been a huge unlock to save our analytics team time of like we're not answering these questions for people they can self serve. It's just democratizing data and saving us a lot of time so that we can focus on more deep analysis. And for deeper analysis we needed something more than just these basic tables. And so I, with a lot of help from one of our data engineers, she built a specialized semantic layer just for my scope as a test. So I was the first one in the company to do this, but we're planning on kind of rolling it out to all of the areas of scope. And basically this semantic layer just defines the tables that I use the most. The joins, the filters, the metrics. And because it lives in our code base, it's in our data science repo cursor can just tap into it and it just makes the zero shot ability insane of running SQL.

A (28:19)

I've seen a couple of these and I don't know what yours looks like, but they really just look like defined terms, tables. This table means this, this field means that if you're trying to query average order value, this is how you do it. And it's almost your documentation in a little bit more of a structured form around common queries. And what I think is nice about this is its ability to be managed by code. You can change it, you can update it, you can add new things. I also think for the data engineers out there, it reduces a little bit of needed complexity on the data warehouse setup because previously you were creating these like aggregate tables and these like defined metrics and you're hoping people were writing queries the right way. And now you can define these canonical queries and know that no matter kind of like what your tables look like, they're going to get to, to the right answer. Which I think is quite nice on the data engineering side.

C (29:11)

Yeah, so this is an example of like what you were talking about. It's just a very structured JSON file and from what I understand, I did not do this, but I had the engineer explain the process to me and honestly LLM's helped a lot with creating this. You know, he fed in details about our data warehouse and just a million queries that I had previously written. And it kind of helped spit out this type of thing. He also used LangChain to like change the names of a bunch of the reports that we had into Question form. Because obviously when I'm querying this, whether it's through a custom GPT or cursor, I'm often asking a question. And so I thought that was pretty cool. Like translating it to a question makes the semantic layer work so much better.

A (29:55)

Oh, this is going to be my next project. This is so fun.

C (29:58)

Amazing. Glad to inspire. So to go back to the actual SQL that was run and I will actually just run this, See hopefully this.

A (30:11)

And just in case people missed this, you did call out the Snowflake mcp, which was what we're seeing right now, which is a programmatic way to hook into running queries in your Snowflake data warehouse. So you can not only generate the SQL here, but instead of copying and pasting it and going into like Snowflake Cloud and running it or whatever your visualization tool is, you can just run it right here. You're getting your tables right here. So again, like you're eliminating that context switching, you're eliminating the copy and paste, and you're getting your data right here.

C (30:39)

Yep, exactly. And so I am. Oh, this is interesting. This. Actually, I am looking at this and I think it showed a mistake, but I asked it to QA itself. Normally this does a very good job, but one of the quick QAs that I do for something like this is I want to see no skip steps. Oh, actually, you know what I remember from the context, this is a temporary. This is a step that only some people see. But usually when I'm looking through this, if we were not doing this demo, I would spend probably a lot longer QA ing this. But I just want to see drop off. That makes sense, right? I don't want to see 0, 0 and then 1 or then 0. And so that's just a quick QA that I can do. It's not the AI's name on this analysis, it's mine. So I do that. The other thing that I have done to really make sure that I can QA this effectively is I, in my cursor rules, I tell it to comment every single CTE so that I know what the and sorry, CTEs are like sections of SQL that often are created when you're writing SQL. And I just want to know each step of what is happening so that as I'm looking at the SQL I can say, okay, the agent said it's doing this and like looking at this code, I can actually tell that it's doing this.

A (32:04)

So engineers, cover your ears because engineers hate, hate, hate, hate, hate when I say this. They hate it. I love over commented AI code and let me tell you why, because when you are not writing this code, you really need to understand the thought process behind how the code was designed. And having AI comment the code that it writes gives you a natural language way to understand if your understanding of the implementation matches the actual technical implementation of the code itself based on the AI's own reasoning. Fine, delete it if you want to, I don't care. I know all the arguments against overcompeted code and I think there's a lot of benefits for human review and it's also great context for AI when they go back and work on it. So engineers, you can now uncover your ears. You can yell at me on Twitter if you want to, or on X if you want to, but I do the same thing where I say go ahead and comment in the code so I can understand how you decompose these step by step.

C (33:03)

Yeah, it's pretty, pretty awesome. It's also I even have a custom GPT in chatgpt to comment code I've written before, I just insert code and then, you know, if I'm ever handing off dashboards to someone, I really don't want anyone to be so confused that they have to bother me. You know, my goal is to have it be quite self serve.

A (33:22)

Look, those lines of code are not going to expand themselves. Let's get some comments. This episode is brought to you by Brex. If you're listening to this show, you already know AI is changing how we work in real practical ways. Brex is bringing that same power to finance. Brex is the intelligent finance platform built for founders with autonomous agents running in the background. Your finance stack basically runs itself. Cards are issues, expenses are filed and fraud is stopped in real time without you having to think about it. Add Brex's banking solution with a high yield treasury account and you've got a system that helps you spend smarter, move faster and scale with confidence. One in three startups in the US already runs on Brexit. You can too@brex.com howiai so I'm going.

C (34:18)

To kick off my next prompt, but basically we're going to skip ahead a couple hours here because up until this point my goal was to get this kind of clean base query that I could use for dashboards. In Mode, which is Fair's BI tool, a lot of what we are doing as the strategy and analytics team is creating creating tables that then can be used for pretty charts to tell a story. Let's pretend that I spent a few hours with Cursor refining queries. I actually did one for the old flow and the new flow. I actually did do this. This is also a real use case like Tim's and then I built some visualizations in Mode. What's really cool is that there is actually a modemCP and I can tell it to view a dashboard directly. For those who are listening here, we have on the left hand side our legacy flow and on the right hand side our new flow. You'll see that there's one step that is only present in some of the entry points as a split by entry point. And basically it's just showing what is the overall success rate and success rate by step for each of these flows. This is what I have pointed the mode MCP towards in this prompt. So if we go back to this prompt and I'm just going to tell it to run this tool, I'm telling it again, hey, go look at this Mode dashboard and use this mcp. I also give it the direct SQL that I wrote with Cursor that's powering that dashboard. I'm just asking it for some detailed takeaways and next steps. I give it a little bit of context and I tell it to ask clarifying questions and use the MCPS if necessary. The mcps, I think. I'm not sure if we've defined it yet, but Model Context Protocol I believe is what it stands for are so powerful. I think that that's when this has felt like magic the most. At first I assumed that they were similar to APIs where everything needs to be defined. Like some engineer on both sides needs to go defined endpoints that there's a very specific structure. It seemed like a lot of work. These models just know what to do. It's just wild to me. I will say that there's a lot of work on our data engineering side to get some of these MCPs set up. So I think Ben on our analytics platform team has just spent a lot of time on this. I don't want to minimize that step, but as the end user of them, it just feels magical every time it can just access something. If we go into the results over here, next key takeaways and next steps. Cool. Looks like we did a good job. Yay. Fair. And it gives a pretty detailed list of the funnel analysis, insights and concerns, actionable next steps, et cetera. This is already a pretty good output to start with. But at the end of the day analysis like this only matters if you can communicate it clearly. You need to convince people of whatever you are trying to communicate. So we also have a notion MCP and I'm going to ask Cursor to create a doc that captures our findings in a structured way.

A (37:55)

And I want to pause really quickly because we have done this in maybe 15 minutes where you have taken a problem kind of like a pre and post analysis of a feature change. You have written SQL, you have not used a WYSIWYG analytics tool. You have written straight up good SQL, traceable SQL to do a funnel analysis of that on a daily basis. Very interesting. You have made a dashboard for it so that your business users can use it. You have then done a meta analysis of that dashboard using the MCP to actually read the dashboard, do a first pass analysis, create a summary not only of the results but of recommended next steps and then you are going to publish that to your business using notion. Now I have to say I have worked with a lot of data teams and most of them spending their time saying what is the priority of this analysis? We have a backlog, I need data engineering. And fine, here's the dashboard. Like it's like the ones that like get promoted three times in a year that go that extra step where they're like, and here's the analysis and here are my recommended next steps. And I made it pretty so you can share it with your boss. And I just think like I was watching this and I was like oh man, I'm going to promote this data analyst. Like they're pretty, they're pretty, they're pretty good. And so I just think the ability to level up the quality of your work and think through the interesting things. The interesting thing isn't like did I write this SQL join correctly? The interesting thing is like have I thought through all the edge cases? Do I have any creative ideas on what we could do next? Can we improve this analysis for the future? And so I really like this end to end flow because it just shows how you are leveraging up into higher strategic tasks as opposed to spending your time sort of in the tactics.

C (39:41)

Yeah, I mean I totally agree and we are almost done but like you said, we need to communicate this. And so one thing that we have done on strategy and analytics is our chief Strategy officer Dan. He really cares about synthesized writing and all the leaders on his Team care about synthesized writing. And so we worked with him a couple months ago to actually create some guidance on how to write at fair. Like fair is very much a vertical doc culture, pre read culture. We're not creating a lot of slides, we are writing a lot of docs. We have this use answer for structure, key principles doc and then we also have a template for what docs should look like actually in this prompt. You'll see I tell it to follow these rules that are in these docs. And that's like another thing I love about SQL or sorry about Cursor is you can just tell it what rules to follow in a variety of ways.

A (40:43)

Okay, Alexa, I'm going to give you an upgrade here which is you should reference these files in your cursor rules.

C (40:47)

So you don't have to. I actually answer that question. That's a great. I should. I mean, I wanted to, you know, show the full flow, but the reason I don't is because it would have actually done it in the previous step. Oh yeah, because it would, it would have, it would have known and then I wouldn't have gotten to talk about it. But yes, I will, I will do that once we are done.

A (41:05)

It's showbiz, folks. That's what this is.

C (41:08)

And so the last thing is I am going to pull over the doc. This is one that created from a previous time. I did this just because I wanted to highlight in yellow. I gave instructions in this prompt to.

A (41:21)

Tell me what to add.

C (41:23)

I think one thing I want to get across is this. I don't think that cursor yet or AI can zero shot executive ready doc yet. That is where I think that we still need to do 3 to 4 revs of editing, adding, analysis, making sure this makes sense. These tools have so much context, but we still have some context that is just this je ne sais quoi. Humans are still valuable. And so this is like a pretty good start. And I think what's cool about Cursor is like I cut out some of the middlemen. I got to this point like really, really quickly. But we're not just creating like AI slop docs all over the place. We are just accelerating how fast analysts can do things like this. We, you know, and the other thing that's really helpful about I would run this through that guidance three or four times. It can be hard when you're so in the weeds of an analysis to take a step back and make sure your story makes sense. And so that's what LLMs are really good for. So it can cover my blind spots.

A (42:33)

Well, you know what's more painful than running this three times through your guidance is sitting three times with your SVP of strategy and having them tell you this makes no sense and you need to go back and edit stuff. So again, I think what a nicer way to get to a higher quality output than a yes than having to.

C (42:52)

It saves me time and it saves the leaders on my team time and hopefully improves the quality. You know, it's fundamentally improving how, you know, we are doing work on analytics team.

A (43:05)

And one thing I want to call out for folks that are maybe listening and not watching is Alexa. My friend here is smiling. This is fun. This is like interesting and it's fun. You're not sitting here saying, I have no role to play anymore. The machines are going to take over. You're saying, man, it was really boring to like dig through tables and write all this SQL that I know how to write and I've done it a couple times, so let's let the machines do it. And now you're able to focus on interfacing with the business having impact. And it's just, I, I think it's fun. Every time I get in these tools I feel like it's magical. I feel like it's really fun. And so I want to call out. We got smiles across the board here on how IAI AI.

C (43:45)

I didn't show this, but the type ahead, like if, when you're actually editing the SQL, that's also so fun. It's just fun. It knows what you want to do. So yeah, this whole process is very fun.

B (43:57)

I think what's so powerful on this, it's not just like making the good analyst just incredible, it's also democratizing data. So this is something that can be done. SQL can be written by people all over our business, whether in sales, designers, anyone else can write this. So the people with the context can do analysis just like this. And, and then the analysts can do the really complicated stuff where these tools could help them get really into the.

A (44:18)

Weeds for people early in their career. I've said this before and I mean it to be true. If you want to know the inflection point of Claire Vo's career, it is when she learned SQL. True. I mean truly, I became unstoppable at that point. And so lowering the barrier to entry on data analysis is just going to create a whole bunch of really high, high impact folks. Awesome. Okay, Alexa, so we just saw how cursor can do end to end. Funnel analysis all the way to the proverbial front door of your SVP strategy. Tim, let's talk about another kind of analysis, which is experimentation analysis. My favorite.

B (44:54)

Yeah, you should have close to your heart. So look, we've talked about the big picture, we've talked about like a really detailed sort of actual analyst of how they do their day job. But I think one of the other things these AI tools are just so good is just accelerating process, like automating away some of those routine lower impact steps in the analytics journey. So as a good example, we want to show you a quick agent we build which automates the process of writing up experiment results. So across fair, we might be running, I don't know, hundreds of a B tests on the product a month. And each of those experiments needs to be monitored, assessed, documented, and that just takes up so much time for analysts. So if we don't stay on top of this very quickly, it's our team that can become the bottleneck and slow down our launch velocity, which is the last thing anyone wants. And I know this is something that's happening up and down the country around every single tech company, so we thought it'd be a good example just to demonstrate. So let me show you how I built this. One thing I want to really, really stress here is just how straightforward these things are to build. Once you've gone through the pain of setting up cursor, getting your MCPS in place, actually spinning up any new agent you can think about, it's just so quick and so non technical for anyone to do. So it all runs off a cursor rules file. So if you don't know what these are, they're literally just a type of file, an MDC file that these agents know to look for and know they're likely to contain instructions. They're really easy to set up. It's basically plain English. So you just write a simple one line entry as a description of what it is. So format for writing experiment result using EPO data. EPO is just the experiment tool that we use. It basically takes our data, does a bit of analysis, slaps a UI around it and writes it out for us. So you then select when you want to apply. I just selected apply Intelligent. I trust the model to work out when it needs to use it. They do a pretty good job. And then other than that, it literally is just writing out what you want the agent to do. Now this might look a bit complicated. I'll generally write this in a few minutes in plain text. What I wanted to write. I'll ask Cursor to then tear the thing down and I'll rewrite it a couple of times and just get it right in the format I want. But ultimately it's just a step by step guide of what I want this thing to do. So I've just said, for those who are listening, I've said if you're asked to write up experiment results, do the following things. So ask the experiment name if you haven't already got it, and then go collect the data you're going to need. So use the EPOMCP we've set up. So go talk to our experiment space, pull in the actual results of the experiment and then use our notion MCP that we've already talked about to go pull in all the other context that you might need. So any other documentation that's going to help it interpret that data and write up this report. And then I've got a little bit down here, you can see telling it exactly what kinds of documents to look for. So PRDs, experiment docs, technical specifications, that's what it's going to help it look for. And then I ask it to basically write out those results in the format I give it. And then I'm pretty prescriptive about the format I want because I want this to do it really consistently in the format we want with really tight, tight takeaways. So actually I've asked it to create it in just a local file on my cursor on my computer and that just means I can actually look at it before it goes create to the notion docs can take a peek, refine the prompt if I need to, but that's just a fallback. And then ultimately it's going to turn into another notion doc so everyone else in the business can see it and it's going to do all this incredibly quickly and we'll essentially just see what this thing looks like in reality. So let's just run it on an experiment result. So I've just said please write up the experiment results for and I've given it the name of the experiment, which is vertical product tile images. And straight off it's gone off and it's found. It's written itself a nice to do list. It's found the EPO result, so it's just called the results. It's found its results. Great. It's found the rules and now it's going to start working this all up for me, which is great to see. And then while it's doing all that, we'll just have a look. So the format we've gone through we can just show here. So basically the rest of this is all just showing exactly what the format this thing is going to look like. So I've asked it to give me the document links exactly what I want. If I click into more context, a brief summary of the experiment and then the key bit, the actual metrics that it's got from epo. So it's going to show me the actual results, the confidence intervals, it's going to pull out the most important ones and it'll give me a nice little color coding for it. And then I just want the actual answer from this. So I actually want it to do the work of interpreting what we should do next. And so it's written the takeaway section. So I want to clear. Should we roll this out, should we roll it back? What should we do? And give me the reasons why, like why are we doing this? And are there any other interesting insights that you found that we should call out from this? So let's see. Right, so let's have a look at what it's doing here. It has found everything we need. It's starting to write out the doc, which is nice to see in this little thing. I'm just going to go ahead and queue up. So turn this into a notion. So as soon as I've run it, while we look at the actual results, it will start writing the notion doc. And let's have a look. So straight away in a second while it's running that I've got a write up with all the right context I need. So it's got the links I needed, it's got the context, it's pulled the right data. Good. The nice thing is this result. So this was just literally sharing vertical images, rather square images, like a really standard growth experiment, like which one performs better. And you can see a nice stat sig lift of about 3.5% for the treatment. And then it's pulled out some other interesting business metrics and let's have a look at these takeaways. So it's saying, great, roll it out the right answer because of that lift. And it's also pulled out some interesting things. So it said our data science prediction models are also actually positive. So it's saying not only have we got more retailers, there are actually higher quality retailers, the ones we've got. So this looked good as a first pass. This looks great.

C (50:50)

And just to call out one thing here personally, we have a standard format for doing these where you have to type the confidence interval and type the emojis. And that is so like work that is not valuable for our team. And so it's pretty awesome that like it came up with takeaways, but it also saved us five minutes of like fiddling around with emojis and decimal points.

A (51:15)

Yeah, I mean AI as a translation layer between a SaaS interface or a SQL query into natural language in the format that you like, that your boss likes. That's just a time saver in and of it of itself. So I, I love using AI as like the universal Format Translator.

B (51:34)

So as you can see, I've just asked the notion link. It should produce the notion. So let's just open that up and let's put it on screen and look straight away, I've got a nice document I can share around with everyone with all the right color codes, the takeaways and even as a little bonus, let's see, it's on it always has trouble getting things in a little toggle, but right at the bottom here, I've even asked it to spit out a slack with an even more summarized version. So I can just drop this into the right review channels and straight away this can go and get approved. Now, are we going to do this for every complicated experiment? Probably not. There might need to be a bit of analysis, but for the simple ones, straight one shot, even the complicated ones, this accelerates you. But also anyone in the business can start doing this, which means we can pass more and more of these things down to engineers, PMs, other people to write this kind of stuff and do the analysis for them, which again can just massively accelerate our launch velocity affair, which we're really excited for.

A (52:24)

Yeah, I'm sorry and I know this is my brand, but I feel like AI is just accruing to every task. Sorry pm, it's your job now. So I do like that, that little trend that's happening. This is amazing. Love it. Have done these kinds of analyses before. They have not been this easy to read and they certainly haven't been generated in 90 seconds. Really useful tool for experimentation analysis. A call out to the experimentation tools out there that I know and love. If you have not made an MCP for access to your data, you are limiting your customers. And so I do think sort of AI integration of SaaS tools is going to be a way that teams start to evaluate the quality of tools that they're working with. So just something to think about if you're out there building data analysis tools. Okay, we are going to wrap up very quickly with a final. We're going to do a bonus. We usually only do three use cases, but yours are all so good. We're going to do a speed run through a bonus use case, which is actually designing and analyzing kind of unstructured data in a user survey. So, Tim, you're going to whip us through how you could use AI to make surveys and survey analysis a lot better.

B (53:33)

Yeah, I'm going to do this really quickly. We don't spend time with this, but let's just show. I think it's just another one of those incredibly common analytics use cases that everyone has to do. And they are just so time consuming. You've got to design the survey correctly, code it into a survey platform, then analyze all those lots. It's really time consuming, consuming. But end to end, AI can just like transform the whole process. Let's show another one. I'm just gonna stop. I'm not gonna run these. I'm just gonna go straight to my backup. So let's just start on design. So what? I love doing this. I think you can do it on curse. You can do it on many things. I think ChatGPT projects is really good for this and again, incredibly accessible. Everyone knows how these work. It's just a great way of giving context. So if we switch over to this one, which ChatGPT is lovely and taking a bit time to load, you can see in files. What I did was give it a bit of background information. So what is our bit of business? So this was a survey we want to design on Fair Direct tools. So that's our tools that we give all our brands to help them accelerate their sales with their own customers. And so I've given a ton of information to the model that just says, like, what actually is Fair Direct? What are these tools? What's the strategy? And then whenever I do a survey like this, whether I'm doing AI or not, I'll start with hypothesis. That's ultimately what you want to test. And so this is nice. If I just open up those hypotheses. So this is what I fed it into. I just gave it a list of simple hypotheses on what we want to learn. We do aligned. We've got everyone aligned on some hypotheses. There's 14 in here and they're really simple. I'll just call that one like, higher sales on Fair leads to more usage of these tools. Things like that that we ask. Now I've just given that into it and all I did. If I just look at this prompt that we ran. So this was a simple prompt. All I did was drop it in saying, you're a specialist at doing these customer insight surveys. Design me a 10 minute survey for the thousand brands to test those hypotheses. I said, these are the inputs I've given you. Here's a bit of design requirements that we want. And I asked for three things. I said, turn those hypotheses into a full questionnaire that we can go ask our customers. But also, don't just do that. Give me the coding file that turns that questionnaire into the actual, in this case Qualtrics, the platform we use to actually run these things, can actually design that straight away in one click and give me an analysis plan for some work to do there.

A (55:46)

I have to pause you really quickly because this whole episode has been Tim saying, I just did this really simple prompt and then you see this like 1000 word, hyper structured, very organized prompt. And Alexa's like, oh man. I would just go in there and be like, maybe a nice survey, please.

B (56:07)

I love it. So I'm a big believer that 99% of my prompts are going to be one line. And then if I'm going to send a model, a big model is going to do work for 15 minutes, I'll probably ask another model just to turn my one line into something more, more detailed.

A (56:21)

I want, I want the AB test of Alexa. You run this exact same GPT with a tinier prompt and you tell me if you get the same quality.

B (56:30)

See what happens? See what happens. Maybe I'm just, I don't trust it quite as much as Alexa just yet. Okay, so what do we get from that? So very quickly, from a list of hypotheses, I've got straight away a really nice first pass of a survey. Now it's going to ask a load of questions. It's about the right length. Like this can just massively accelerate the process. And then once we've got that right, it's also given me that coding file which I'll just scroll on screen. These things are painful to write. So just having this, a one liner to tell exactly how the system should prompt this and write it out is just like, saves hours of time for our research operations team. And it even then translates that into an analysis plan that says this is what the outputs from that are going to look like. So straight away this whole thing can go from a list of hypotheses into something we could probably get out to our customers by the end of the day now that's like shortens this enormously. But what happens when you get the results back? That's the other thing this can do. And so again I'll do this incredibly quickly and just show you the final result. But I did a very similar prompt as well. So all I did, I'm going to show you the file I dropped into this, just show you how painful this is. So I just gave the same hypotheses and look how bad this is. It's the raw output from coraltrix. These usually take a lot of cleaning. It's one line for every respondent and then one column not just for every question, but for every possible answer to every question. So these things are incredibly dense for anyone who's worth them and they take a bit of time, a bit of playing with. So the only other thing I gave it was a sub helper file, which was basically that sort of coding file that I just showed you. So it's the what's the question id, what's the question language, what's the answers? And then is it I just add these two columns which is like, is it a demographic question or an answer? And is it a single choice or is it a multiple choice? That's all I gave it. And then I've written another one of my fun and simple prompts. So task here, just analyze the survey results, find the right most interesting things in this data and then judge the predefined hypotheses. So I want a table that basically says like follow these hypotheses. Was it right or was it wrong? And then again I always end on little quad, check, rest. I don't want it to go away 15 minutes before and come back with something that isn't very useful. And let's have a look at this just very quickly. So I've got a nice little summary out front and then there are my 14 hypotheses. Oh, and it's got a nice table that says proved neutral, disproved for each of them. And it's even because I asked it to, giving me a nice confidence score. So I said one, it's really confident in this. Five, it's not very confident at all. And you can kind of see the different levels throughout this. And then beneath it, I've got for each of these actually the specific analysis that I asked to do. So just throw all the insights it found to back up those findings. So like, is this the only analysis we're going to do on this survey? Like almost certainly not. But day one, I'VE got the results, I've thrown it into this and within a matter of minutes I've got a much, much better intuition of what all that day is showing. So while I might go and do some analysis on this, I can be so much more targeted on exactly what we want to look into and where I want to spend my time. And straight away we can start sort of sharing some of these findings out with people very, very quickly.

A (59:40)

Oh, no. So I'm reflecting now after this episode. Like, okay, I've told everybody to ship a bunch of features and now I'm going to be like, do a bunch of analysis. Like in my mind I'm like, oh my gosh, I'm underusing AI to actually understand my business. And it's so accessible. And if I can just write 17 point prompts like Tim, I can get really high quality insights. But I do want to call out, just reflecting on this whole episode in your four workflows. What I love about what you're showing us is, is so many people think that AI is an input to producing a thing, but haven't done that that full circle back to analyzing the thing, sharing the thing, communicating about the thing. And I think you're showing both sides. You can create with AI and you can analyze and communicate with AI. And I think looking at both sides of that coin is really useful. Okay, we are going to do the one and only lightning round question because we have gotten long on this episode and I want to get you all back to all of your agents and MCPs and analysis. We're going to go back to prompts one last time. We're going to figure out your personality around prompts. Alexa, Tim, when AI is not listening, when your MCP will not call the tool, what is your prompting technique? Alexa, what do you do?

C (60:57)

I think mine's pretty straightforward. Where I think the problem that I run into most frequently is that I'm clearly running out of context. Like a conversation has gone so long that it's starting to be wonky. And so while I think level one is just starting over, what AI is best at is summarizing. So I'll say, hey, summarize what we've done so far in this 30 turn conversation and then use that to start over. Because I've heard other episodes people say you want to figure out where it got off track. Clearly I'm a pretty efficient person. I don't, you know, I'm not Tim. I'm not like writing out the entire prompt for 20 minutes. Like I don't have time for that. I just want to say, hey, summarize what happened. We're going to start over, but I'm going to give it that summary so at least the new conversation can get some context from the old. Great.

A (61:50)

And Tim, what about you?

B (61:51)

I've got so much stick to my prompts. It's all AI. It's all AI. What does my chat did? So I generally will go and open up three windows on cursor and I'll do three chats with three different models and put the same prompt in and go get myself a cup of tea and see what comes back. That's the British stereotype in me and getting my cup of tea while I do it.

A (62:09)

But yeah, you run the AB test.

B (62:11)

Is what you do.

A (62:12)

Okay, I love this Tim Alexa, where can we find you and what can we be helpful with?

C (62:17)

You can Find me on LinkedIn. My full name is Alexandra and ways to be helpful Our strategy and analytics team is hiring across the board. Our team partners super closely with PMs and our go to Market team. We make strategic data driven decisions. Super fun. We have tons of open roles so if you like experimenting with AI, we are very AI forward. So you can learn more@fair.com careers and.

B (62:44)

You can find me on LinkedIn as well. And I echo that as well. Like come join us. If you love AI, come join us and show us how we can do it more here.

C (62:50)

Okay.

A (62:51)

We will link to your careers page in the show notes. Alexa Tim, this has been so fun. Thank you for joining How I AI.

C (62:57)

Thank you for having us for having us.

A (63:00)

Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify or your favorite podcast app. Please consider leaving us a rating and review which will help others find the show. You can see all our episodes and learn more about the show@howiaipod.com See you next time.