← Marketing Against The Grain
Loading summary
Transcript59 lines
- [00:02]
A
Hey, everyone. I just had a great conversation with Tina Wong. She is a data scientist turned creator and builder and she is awesome at building agents. And we just had the most positive insight, which is the next generation of builders over this next year are going to be the best marketers, the best salespeople, the best business people. They're not going to be the best technical people. And so what we did is we built an episode for all of you to to show you how you can take your domain deep knowledge and build awesome AI workflows to actually solve your problems and scale the knowledge that you have. This is gonna be a great show. Let's get into today's episode. Did you know that most businesses only use 20% of their data? That's like reading a book with most of the pages torn out or paying for a coffee. That's 1/5 full. Point is you miss a lot unless you use HubSpot. Their customer platform gives you access to the data you need to grow your business. The insights trapped inside emails, call logs and transcripts, all that unstructured data that makes all the difference. Because when you know more, you grow more. Visit HubSpot.com today. So, Tina, thank you for being on the show today. We're excited to have you.
- [01:27]
B
Thank you so much for having me.
- [01:28]
A
You have a very interesting background. You are deep into data. You're a former data scientist, but now you're a creator, you are an entrepreneur and you teach the world how to build amazing AI workflows. And you work with companies to build amazing AI workflows. So I guess I want to start with what is the state of the work you can do with AI today? Tell me, like the most popular workflows that you are teaching people, building for people give us kind of the background lay of the land.
- [01:57]
B
Yeah, for sure, for sure. So the most useful workflows are usually not the coolest workflows, so.
- [02:05]
A
True.
- [02:05]
B
That is very true. It's like, what's the coolest kind of workflow? You know, we can make all these robots. What's the most useful workflow we can make reports.
- [02:14]
A
Exactly, right.
- [02:16]
B
Yeah. So I would say there's a lot of report workflows that we work on. These are very custom because companies generally have a very specific way of doing their reports. And a lot of companies have to do a lot reports to their investors, to their stakeholders, you know, a lot of different types of people. So that's a very popular workflow that people try to automate with agent tech AI. Another one is customer service Customer service, like email lists, these kind of things. So for example, we will have people who would be sending emails and then being able to respond to those emails in a more personalized way. That's also a very common and useful workflow because the majority of questions that people ask say are very similar to each other. It's like the same questions over and over again. I don't know how to log into my platform, I don't know what's my password, you know, things like that, which you generally don't need people to do. So AI is also really great at this kind of workflow. I say those two are two of the most common ones that we see and other ones are really just related to operations that are more specific to different companies. Operations and dashboards as well.
- [03:19]
A
Okay, so a lot of reporting, a lot of kind of customer support service and also like transactional elements. Most of the people who are watching this, they love AI, they're proficient in LLMs, but maybe they're not big into the different AI workflow tools. And how do you like, get started here? So maybe you could give us like the overview of the landscape of like, what are the most popular tools out there for building workflows? How do you kind of get started and pick some place to actually build a first workflow?
- [03:48]
B
That is a great question. I really love that and I love how we're framing it in where do we start? As opposed to let me just build something and see where to fit it in. Because that is, I think the number one issue that people have first thing is mapping out your own workflow, like your own human workflow of what you're doing. Generally speaking, there are probably things that you do which are quite repetitive and it involves you having to do it many times. So it's a consistently repetitive task. That is low risk. That's when it's important. If it's a very high risk thing, you definitely don't want to be risking that. And something else I like to keep in mind is what AI is good at. So, for example, AI is very good at being available 24, 7. It's very good at consistency, being able to do the same things more or less, like agentic workflows, you can get them to be quite consistent. And personalization is a really big one. Where is it that personalization can make a big difference?
- [04:40]
A
I think this is a really interesting point. You're basically saying, like, look at the calendar. Find the things you do on a repeated basis that are on your calendar. Right? Like, what are the Reoccurring blocks of like your personal work time and find the things that you do over and over. And I would even take it a step farther. Right, Tina? I think somebody could just do a screen share record of that time and put it into Gemini 3, which can watch up to an hour of video and just ask Gemini, what can I automate off of this work? I have to do this, this hour long task, you know, every other day. What can I automate with AI and how much time could I save? And just like recording the things that you were normally doing, like, you will get really good suggestions from the LLMs of how you could actually be far more efficient.
- [05:28]
B
That is such a good idea. You know what? I'm going to write that down right now. Thank you for that.
- [05:33]
A
No, I'm obsessed with Gemini 3 video consumption. I really am obsessed with it. And it's so good at that and it's so nuanced in understanding what's happening. And I think if somebody was really diligent, there was an awesome creator, AKA Tina, or somebody else, like, you could literally do that for like every hour of your week and really see how much time you could save across an average week with AI. I think it would be pretty amazing the transformation that would come after doing that.
- [06:02]
B
That is so smart. Literally wrote that down. I'm going to implement this. I'll let you know.
- [06:06]
A
I'll reply back. Please share the videos that you do. I want to see. Okay, so that's a little bit of where to get started. I talked to a lot of companies, Tina, and a lot of companies just, they run into a bunch of roadblocks. And I was wondering if maybe you could give us the kind of the common roadblocks and how to get around them when you're getting started going and building these business workflows.
- [06:28]
B
With AI, the most common roadblocks is other people.
- [06:32]
A
It's data. And other people are the most common. Two, I have failed.
- [06:35]
B
True. I absolutely concur with that statement. Other people and data. Oh, yeah, absolutely. I think, yeah. Oftentimes, like when we do work with clients, literally, like the first blocker that it comes across is other people. Because they're like, oh, first you have to convince stakeholders, Right. That you want to do this. And it's oftentimes, like with AI solutions, it's not actually as plug and play as people would like it to be. Like, you can have a great solution. Right. But the problem is not the solution. Solution is great. The problem is that you can't use the solution. Based upon your current architecture, it doesn't necessarily support that specific solution. So what we find is, like, a lot of the work that we do isn't necessarily coming up with, like, this fancy new agent. It's like, how do I incorporate these solutions in a custom way that makes sense for that specific company in also a way that has enough transparency and trust so that people are able to accept it.
- [07:30]
A
Got it. Okay. And if other people are the problem, a lot of times, like, what are your tips? What are your tricks for getting other people on board and not being blockers anymore?
- [07:40]
B
Yeah, so I actually find that it does depend a lot on the company culture. So we work with companies from all over the world. We have North America, Europe, Asia, like all of these different places, and their culture is very, very different. Very so different for different places. And I think the way that stakeholders are managed as well are also really different as well. Say, like the North American way. We find that what generally works was that you have, you know, your primary stakeholder, you build something, and then if you say, here, I built this thing, here's like a demo, and the stakeholders can see the potential of what it is that you're building. You know, cost analysis, return investment, things like that. Generally they would be more amenable to, hey, like, you know, we can start this as a trial, at least very roughly. That seems to work decently from a, like, North American perspective or like, maybe a more like Western world on the Asian side of things. What we find is you can't really do that. If you just do that, it's a little disruptive is how I would put it. Yeah. At least from. From what we've seen, it's a little disruptive. It's more about kind of actually just convincing a stakeholder to be able to portion out a experiment or a project. A lot of times this is driven from stakeholders where, like, investors saying, hey, you guys need to have AI in this company. And them going like, I don't know how to put AI in my company. So they would come and find us and we would help them, like, come up with a strategy, and then they would allocate a certain area to experiment with using AI that seems to work better from the Asian side of things. I'm being super general right now because.
- [09:13]
A
Of course, general painting broad regions, but there are cultural nuances. Here is one of your big points.
- [09:18]
B
Exactly. Because, yeah, most of our demographics, I would say, is blend to like North America versus like, East Asia. So that's why I tend to know the best.
- [09:25]
A
Got It. I think what's interesting about that is that there's different ways to kind of get past that initial inertia of getting started within an organization. But one thing that you said that there, especially around like the kind of culture in East Asia is sometimes there is a solution in search of a problem. It's like, we want AI, we want to go and do AI and then like, figure it out versus, like being very problem obsessed. I think being problem obsessed and figuring out the value agnostic of any technology is always the best way to do something. So, like, if somebody's out there and their boss is like yelling in their ear, like, hey, we gotta do more AI, what would you tell them to get out of that? Like, I'm just looking at random AI technology and getting to a world where they're like, actually understanding a problem and building something useful against that. Do you have any, like, heuristics or frameworks for like, how people should think about that?
- [10:22]
B
Yeah, yeah. It really just goes back to what I said earlier. Mapping out that workflow, potentially using Gemini perhaps, and then just seeing where it actually lies so that companies don't know where their problem is. But hiring an external consultant is actually a really good approach. Not because they don't know like, what their problem is. It's because they need someone to point it out to them and then offer like a viable solution. And then they'll be like, oh, okay. And then they would go and actually like, implement these experiments within the organization. But yeah, it really comes down to that. Just identifying where the issues are, where it is that things can be automated, where it is that using agents could be helpful, agentic workflows could be helpful, and then integrating that into their systems. So this goes back to identifying the problem itself.
- [11:09]
A
Okay, got it. Look, building AI agents is cool, but if your prompts suck, your agents suck. That's just reality. It's all in how you prompt it. Tina, put together a killer AI prompt engineering guide. It's got the exact techniques you need to craft prompts that make your AI agents actually work. Get it right now, scan the QR code or click the link in the description. Now let's get back to the show. Well, speaking of identifying problems, and you also had a tip around. Build like a basic version of something and use that to kind of get the process started. I know you're building a lot. I think you've got an agent in a workflow that you want to show everybody today. Maybe we could kick off by like having you walk us through that Talk about what it is, how you built it, the problem you're solving. Let's model for everybody what they would do if this was their company and they were trying to basically think through solving this type of problem.
- [12:07]
B
Yeah, for sure, for sure. Let me briefly explain the problem first.
- [12:11]
A
Yeah, please.
- [12:12]
B
So whenever we have people buying a product and a subscription based product, there's this company in which they're getting a lot of people sending them emails wanting to unsubscribe from their product. And once ends up happening is that they would have this generic flow where they would just try offering them discounts as a way for them to not unsubscribe. But the problem with this is they're very generic. Like, they'll be like, I have a very specific problem. And the solution would be, here's 10% off. It doesn't feel very personalized. And the unsubscription rate was still very, very high. And as we all know, customer acquisition is already really difficult. So once you actually get those customers, you really do not want them to leave. Same with like email marketing as well. It's like actually a similar problem. It's like once you actually acqu. You don't want them to leave. So the general problem that we're trying to solve here is like, how do we prevent this churn from happening in a way that doesn't involve coming with an extensive plan, getting a lot of people to work on it in order to solve this problem? AI is a really great solution because going back to those three different things that AI is good at 24, 7 availability, consistency, as well as personalization. Personalization is massive here.
- [13:18]
A
Huge.
- [13:18]
B
Exactly. So it's like people, they like to feel as if they're being heard, right? It's very important for people to feel like they're being heard. So the solution that we came up with is creating an AI agent that's able to identify the problem based upon the email that you would get as a company. And then they would have a predetermined set of discount codes and discounts that are applying for specific types of features addressing certain types of problems that they can identify. Like common problems that people would have. And the AI agent would be able to reply to that email, making the customer feel as if they are being heard and then offering the correct upsell. Not upsell, but like, you know, the.
- [13:58]
A
Coupon code retention discount or yeah, whatever. To try to hold on to the customer.
- [14:01]
B
Sure, exactly. Exactly. And then otherwise, of course, we need to have an out on this. If the agent Feels like he doesn't know how to proceed with this. They would escalate that to an actual human to be able to deal with.
- [14:12]
A
Sweet. That little intro was littered with helpful tips around, you know, having contextual discounting the value of AI personalization. It's not just like solving this repetitive problem 24, 7, which is really important, but it's doing so in a way that is as good or in some cases better than a human could do because of like the speed and scale of which it has to happen.
- [14:34]
B
Absolutely. Yep. So there's like speed, availability and personalization. That's something that humans would not be able to achieve.
- [14:41]
A
So the platform you're building this in is N8N, which we've talked about on the show before. N8N. Gunloop are a couple of the, you know, early very popular AI workflow tools, right?
- [14:52]
B
Yeah. So here is the NA10 workflow that we have. And here we have some examples of emails and cancellation reasons. So we have dates over here and a bunch of different emails. You know, they're saying why it is that they're canceling stuff, like their team size has decreased, onboarding process was too confusing, customer support, et cetera. And here are examples of some of the offers that we predetermined. So we're executing this workflow. So normally this would be automated. So whenever there's an email that comes in, it would actually trigger this entire workflow. But just to show for demo purposes. Okay, so workflow executed successfully here. So we go over here to the audit logs. We need to make sure that we're auditing things. And we have the email draft here and it says, hi, Cather, we understand that the onboarding process was confusing and getting started was difficult for you. So to help improve your experience, make it easier to get up and running, we'd like to offer you a special free two week trial of our engineering toolkit add on. So this would be very specific to Kather, who would be emailing because she's having problems with the onboarding. And this is very specific to engineering as well. So you can use the end trial at checkout to activate this trial. We encourage you to take advantage of this opportunity to explore the added benefits and see how it can assist you better. So that would be an example of a very targeted email back to Katharia, who hopefully will then not want to cancel it anymore.
- [16:08]
A
Exactly. That's the goal. Right. But how long does it take to build something like this? Tina, we've shown Nadan on this show before and it's like, oh, I think some people are like, oh, that's hard. That's complicated. How do I even like, think about doing this?
- [16:21]
B
I want to take a guess. Let's talk about like the actual. Just the building of the agent itself. Right. After determining what it is that we need to build, you want to take a guess.
- [16:28]
A
Well, you built this, right? How much practice did you have before building this? Because that's important. A lot. A lot.
- [16:35]
B
Great question. I love how you're asking these questions. Like first must. I've got to get the understanding of the landscape perfect. Yes, a lot, A lot.
- [16:43]
A
But I mean a lot is. If somebody dedicated a few hours a week over several, several weeks, a few months, you could probably be proficient in the core skills. I imagine once you have the core skills down, this is probably a couple hours.
- [16:58]
B
That's actually correct. We do run an agent's bootcamp as well, which is like a 28 day program. And by the end of the 28 day program, people are able to build this workflow like where whatever workout is very similar to this in about a couple of hours. Yep, absolutely. That was.
- [17:13]
A
And how much time, really good 28 days do people need to spend?
- [17:16]
B
So averaging around like four to six hours per week.
- [17:20]
A
Okay, so what you're, you're really saying is you need to do about an hour a day a month in the current ease of use and complexity of in any AI workflow app to really kind of have a baseline proficiency to build real business workflows. And a lot of that is probably like your own data, your own systems and everything internally in addition to the actual workflow building itself. Is that right?
- [17:45]
B
Yes. This is an example in which we're using N8. I happen to be a kind of person. Like what we tend to teach is I don't think people should be so into the tools themselves because if you understand the fundamentals of what you're building, I could use really like different types of tools. In this case, it's a no code tool. If I want to make a code tool, I can use a code tool. I can use a lot of different things. I tend to not focus so much on that. It's more like understanding. Here are the major components of building an agent. Right. You need to have certain components that are there. You need to have evaluations and you need to have testing. So these are like fundamental things that you must have. After you do that, you simply apply that to any tool it is that you're using. In this case, it would be NA10.
- [18:21]
A
Before we keep going, can you talk about the fundamental steps a little bit more? Because I don't think most people know things like evaluations and the core steps in doing this are really, really important. And you're basically giving people the roadmap of no matter what tool they use or what they're trying to build. These are the building blocks that have to be a part of any of this process, right?
- [18:40]
B
So what is important when we're thinking about building an agent before you actually just start and going and building it the way I describe it. An analogy I like to use is like a burger, right? It's like a burger. You can have different kinds of ingredients in the burger. You need to have a bun, you need to have vegetables, hopefully you need to have a patty, you need to have condiments. However, the type of patty, vegetable bun can be different. You can have like whole wheat bun, you can have different types of vegetables, you can have ketchup as opposed to mustard. But if you don't have all of these components, you don't really have a burger. You have like a piece of bread or like some weird sandwich situation. So it's very similar to agents in this sense because agents, there are certain characteristics that it does need to have to be considered an agentic workflow or else you're going to have problems when you actually do implement it. But you can switch out a lot of these different things depending on what it is that you're trying to build and what constraints that you may have. I actually really like the OpenAI framework for how they categorize different components of an agent. For example, the first thing that you must have an agent, you need to have a large language model. Can't not have a large language model. But the type of large language model can be different, right? Depending on what it is that you're looking for. Another one would be tools. So what tools are you going to give your agent in order to accomplish the task that it needs to accomplish? There would be knowledge and memory. So what kind of databases does it need to have short term memory, does it need to have long term memory, does it need to have privacy concerns? Or choosing what type of memory is also important. Then there's the audio and speech component. Many agents these days, because audio and speech is so good now having that component can be very helpful depending on the type of workflow that you have. And then finally there's the guardrails. So things that you should make sure that your agent doesn't do. And then you also have on the side, deployment, testing, and evaluation. The last two are things that people tend to neglect, which is the guardrail side. And also testing, evaluations and deployment. Building your actual agent is actually just the beginning, because how are you supposed to know how good your agent is, right, if you don't have a way to test it? So that's why evaluations is like the other 50%. You need to build out a set of evaluations to make sure that your agent is behaving the way that it's supposed to. You don't want your customer service agent to go rogue, right. And start saying things that it's not supposed to do.
- [20:53]
A
Okay, that is super helpful. And I think the hamburger analogy is a great one for people to kind of get their heads around. Oh, these are the core elements that you actually need to build an agent or an AI kind of native workflow.
- [21:07]
B
Exactly, exactly. Like, I really encourage everybody, before you go and start building anything, think through this carefully first. Think about your problem first and figure out what the problem is. Then make sure that you're addressing each of these components, and especially on the evaluation sides. Think about what does success actually look like for your agent and how do you translate that to evaluations? The great thing is that many platforms like NA10, for example, they do have simple evaluations, as you can see over here. But, you know, you can also use evaluations for different frameworks as well. Most of them have evaluation frameworks, and it's really important to take time to think through some of these evaluations. Even if you just start off with like five evals, that's actually quite good. Five evals is much better than zero evals, for example. And then you will go back and tweak your prompt. It's usually the prompt that you're tweaking in order to make sure that it's behaving.
- [21:54]
A
And evals is like a standard for what the output looks like. And is the output consistently meeting your standard? Is that the right way to think about it?
- [22:01]
B
That's. That's a great way to think about it. So evaluations are when you are having a certain input, the expected output is it doing the thing that it's supposed to be doing. Like, for example, if you have a math AI agent, very simple, and it's you like two plus two, right? The answer has to be four, and it happens to be five, then you would fail that eval. So you would actually be able to quantify when it does fail, and you can quantify what it's failing at. And also the Percentage of failures. And this is how you can go back and iteratively improve the results of your agent. Otherwise you're kind of just guessing.
- [22:33]
A
I love that. Okay, that is super helpful. I think we've covered a lot of ground in our conversation today. So we started with using AI and agents and workflows to solve problems, not just for the sake of technology. And we talked about taking the repetitive work, documenting it in some way, whether you're just writing notes or taking a screen capture, then asking an LLM for feedback, like, what should I think about documenting? What should be the places to start? And then we kind of went into actually, well, how do you think about agentic workflows? How do you build agents? What's the hamburger need to look like? What are the core components to that? You walked us through a great demo of how to build a basic agent to help retain customers and stop customers from canceling to kind of close out the show here. Is there anything else that you would tell people is a must know, must understand about this space. You're teaching workflow building on a regular basis. Like, what else do people really need to consider when they're getting started here?
- [23:34]
B
Domain expertise. This is not actually AI specific, because in my opinion, the people who end up building the most valuable agentic systems that we've seen this like over and over again are actually, surprisingly not the engineers. They tend to be people who have very deep domain expertise in a field. Like for example, we had some very interesting ones, like people in like pest control. You know, it's like they just spend like 10 years doing pest control or people who work in a law firm, accounting firm, in like privacy for very specific niches like drug development, pharmaceuticals. So deep expertise is actually the most valuable thing of an agentic workflow because that translates into your ability to actually evaluate the agent. If I, as a technical person who doesn't know anything about pest control, if I go build a pest control agent, I don't know like what I should even be evaluating, right? It's like, is this, is it to do it correctly? I have no idea. I don't even know if I'm even building the right thing to begin with because I don't even know what the workflow of a pest control person could be. But someone with deep domain expertise and actually just the touch on the agent, like the tool side, like being able to understand how agents work, that's a very powerful combination. Again, like we have people going through the bootcamp and it's like a 28 day bootcamp where, you know, whatever. They're just like watching videos and learning about the technical side. Because tools are actually really accessible now. You don't even need to know how to code in a lot of times. Some people are going to disagree with me that one, so. But many workflows don't even require you initially to know how to code, to be able to build something out. But what they do require is deep expertise in what it is that you're building.
- [25:12]
A
So I think this is one of the most important things. All of the builders I talk to and product managers and product leaders that I talk to are like, oh, the product manager of the future is really just a deep domain expert who wants to build technology around it. It's not somebody who has deep understanding of the technology and then is trying to learn the domain enough to build the technology, which is kind of what it has been for the last, you know, 30, 40, 50 years. And so that's a major shift. And so I think what we're really saying is, if you're watching this and you're a domain expert and you're part of marketing or company building, then you're already have the most important skill necessary to be successful in this AI age, which is the deep understanding of the problem and the solutions for those problems, so that you can then go build the right agent or workflow to address it. Is that right?
- [26:02]
B
Absolutely. I love the way you put it. And isn't that so interesting, because previously it was the engineers, right? There's the technical people that had all that power, and now it's really so much so. It's like all that time that you spend in a specific domain, just translating that with a touch of technology. You have so much potential there.
- [26:21]
A
Could not agree more. We're going to leave and close out the show on this bit of optimism because I do think it is how everybody should be realizing that the best practitioners and craftspeople are going to be the builders of the future. And we just gave you, I think, the kind of core framework. You can obviously check out Tina's courses and bootcamp and content if you want to go way deeper in the building side. This is just kind of the introductory here. And Tina, thank you so much for joining us on Market against the Grain today. We really appreciate it.
- [26:49]
B
Thank you so much for having me.
- [26:50]
A
Thank you. See you all next week. Sam.