October 21, 2025 • 39 mins
Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX
Unlock the power of AI for personalization at scale in your business. While many focus on AI for efficiency, the real game-changer is creating magical, more human digital experiences for your customers. We're joined by Lee Russell, founder of Minimum Viable Launch and an AI expert with over a decade of experience, to discuss how you can leverage AI to be both "high tech and high touch," transforming your marketing, sales, and product delivery.
Discover the concept of "context engines" and learn why gathering qualitative customer data is your new competitive moat in the age of AI. We explore the tools you need to build this infrastructure, from powerful automation platforms like n8n to the magic of vector databases for chatting with your unstructured data. Lee also shares his strategies for avoiding generic "AI slop" and creating truly unique, high-quality outputs that will make your customer experience remarkable.
Learn more from Lee Russell:
https://www.minimalviablelaunch.io/
Connect with us
https://x.com/ToolUsePodcast
https://x.com/MikeBirdTech
00:00 - Intro
01:27 - "High Tech, High Touch": Personalization at Scale Explained
07:55 - The Best Data for AI Personalization
12:03 - How to Avoid Creating "AI Slop"
21:15 - AI Tools for Building Custom Agents (n8n & Vector DBs)
30:37 - How to Start Learning and Escape "Tutorial Hell"
Subscribe for more insights on AI tools, productivity, and personalization.
Tool Use is a weekly conversation with the top AI experts, brought to you by ToolHive.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Whoever engineers the best context around their customer in
this new age is going to be the winner.
AI allows you to be high tech and high touch, so you can do
personalization at scale and this is the technology that
allows you to build that world class context about your
customers and your data. Using AI for efficiency gains is
(00:20):
something we've talked a lot about on this channel, but
there's so much more we can leverage AI for.
What about making digital experiences more magical?
More human. One of the promises of AI is for
personalization at scale. But how do we do it?
So in episode 62, Tool Use, brought to you by Tool Hive,
we're joined by Lee Russell. Lee is the founder of Minimal
Viable Launch. He's worked in AI for over a
decade and he's worked with someof the fastest growing companies
(00:41):
to integrate AI into their businesses.
So today we're going to discuss how you can use AI for that
personalization at scale, what tools you need, what things you
need to consider, what mindset shifts you need.
So I really hope you enjoy this conversation with Leave Russell.
What's fascinating to me is someof the biggest innovations that
I've been in charge of developing.
The things that give those outsize returns.
(01:02):
It's a small amount of effort and gives you wild results.
Very rarely is the crazy North Star Hail Mary vision.
It's usually a series of small tweaks and small experience
gains that compound on top of each other that say finally,
this is now a consumable experience by by your by your
(01:24):
customers that they love that they tell others about the area
that we are kind of getting people to really focus on.
In terms of where I think the low hanging fruit in with with
AI is, is this moving into this direction of personalization at
scale, this idea that before AI you had a choice.
(01:48):
You could be a highly personalized business which kept
intentionally small charged highticket you, you serve fewer
customers, but you know their names, you know when their dog's
birthday is. It's still the kind of high,
high touch white glove service or you have the choice of let's
be really scalable. And so to be scalable, we're
going to be templated, we're going to be automated.
(02:10):
We're going to lose some of thatpersonalization, but we can
serve many, many more customers in doing so.
And as you move into that kind of more unpersonalized service,
you close less people, you get less loyal customers, you know
that you do lose something there.
AI allows you to be high tech and high touch, so you can do
(02:32):
personalization at scale. And I think a lot of businesses,
if they really thought about what that means, it changes the
game for marketing. It changes their entire sales
pipeline of how you have these customer conversations and then
follow up with them in individualized ways.
It changes your nurture sequencewithin your business massively
(02:53):
and it changes, it changes your product delivery system.
And, and this idea of treating your customers not as the
faceless mass, but as the individual, I think is the big
change that happens in the intelligence era.
People are going to, people are going to look back in 10 to 15
years time and go, do you remember in the 20 twenties when
(03:15):
they had like static landing pages that everyone visited, how
crazy it didn't change to that user's needs at all.
That's going to be a wild thing to look back on.
I love the idea of personalization at scale because
we've all received those emails where we'll be like, hi and then
A brackets first name because something wasn't parsed
correctly. And I feel like like that's an
easy example. Just by addressing someone by
(03:36):
their name in an e-mail, you're have a better likelihood of
getting a success. What other avenues should people
look into in their businesses orin their departments to add some
personalization to the experience?
Well, let's take the e-mail example because I think we've
all had the other form of personalization right where it,
it has your own merge name filled in it, but it's
essentially an e-mail. Like Haley, I saw you're the
(04:00):
owner of minimal viable launch. Great business.
Love what you guys are doing. Anyway, do you want to buy my
stuff? And you're like, what is this
e-mail? Why did you even try and
personalize it like that? It's like the the effort wasn't
worth the juice, wasn't worth the squeeze In the world of AI
marketing. This, this idea of being high
(04:21):
tech, high touch, the information age way of doing
this is we create ACRM of 10,000customers.
We've got their e-mail addresses, we know their first
name. We create one e-mail, we
customize it with the merge fields, one or two points and we
send one e-mail to 10,000 peopleand we hope that one to 2% of
(04:41):
those people click on it, click through to it and make me book a
sales call for that nurture, right?
It's spray and pray. One to many, highly
undifferentiated. The world of high tech, high
touch marketing is very, very different.
You're, you still have the 10,000 people in your CRM, but
you write 10,000 emails and sendthem once.
(05:05):
Everyone gets a pretty individualized experience.
And, and when I talk to marketers about this, like they,
they open their eyes in fear. I, I, I thought they would get
really excited about this, that this terrifies them because they
go, how do we track metrics? How do we track open rates?
Who cares? Who cares about most of that
stuff, right? What we want is to set up an
(05:29):
infrastructure that allows you to send individualized, like
truly individualized touch points to your customers.
And what we often talk about in in our company here at minimal
viable launches, whoever engineers the best context
around their customer in this new age is going to be the
winner. It's not going to be who has the
(05:50):
best AI technology unless you work in open AI, Anthropic,
Google Deep, you know deep mind,whatever.
You don't have access to the true state-of-the-art.
Everyone else has what is the commercial state-of-the-art and
it's one line of code to switch between them.
Who cares if you're loyal to an AI company?
I don't know why because like they're all pretty good.
(06:12):
Just go who has the the least token expense, right?
Don't be loyal. So the models are not your MO.
What is going to be your MO is who can become cleverest at
generating unstructured qualitative data about their
customers. This is your sales calls,
transcripts. It is doing quizzes and other
(06:35):
kind of fun ways of interacting with a customer that allows them
to give you open text entries about their pain points and
their problems so that you can do something in your business.
We call them context engines. You can combine those insights
about your customer and the facts from your CRM into a
template of an e-mail that you want to write and you have 40%
(06:57):
of that e-mail is templated out.You know, it's good high quality
stuff. 60% of that is the AIS hallucination between the
insight and the fact. And they get incredibly
customized personalized outreach.
And then it becomes the game of how do we get more of this?
Some of the experiments we've done with businesses around
(07:18):
this, they get a 20% conversion rate on their emails, 20%.
It's crazy. Now it's new.
That is going to go down over time, I guarantee you.
But right now, if you're brave enough to be an early mover, you
can go and grow your business dramatically by just not doing
(07:39):
terrible e-mail outreach and being clever about how you get
that context of your customer. Could we go into a little bit
deeper because you mentioned thesales calls is a valid form of
collecting this context as well as open-ended answers and in
forms or surveys or however elsedo you see value in things like
setting up Google Analytics so we can see how they interact
(08:01):
with their website or those datapoints kind of outdated in this
AI age? Google Analytics I, I love
Google Analytics and I'm not an SEO expert and this is not my
world. So what I'm going to say is
going to ruffle some feathers and people will probably tell me
why I'm all sorts of wrong, but this is my experience with it.
Google Analytics always shows you the past events so you can
(08:23):
see things have happened, but you very rarely know why they
happened or how they happened. And it's kind of this soup and,
and I, I liken it to a bit of a mental challenge, right?
We got a pool table with a load of, with the, the balls are all
kind of spread all over the place, right?
And you've been blindfolded, putout the room.
(08:44):
I whack the, the cue ball. They go everywhere.
You haven't seen it. I bring you in and I say, Mike,
tell me the exact path these balls took to reach this
position they're in. It's an impossible endeavour.
You, you can't do that after thefact.
You have to observe in the moment.
So Google Analytics is useful for so far, but I think the
entire industry of getting obsessed with these kind of
(09:05):
statistics is a little bit, let's just say misguided.
Because as far as what my skill set is concerned, my
understanding of statistics and these trends is like it.
It tells you what happened at scale, but that doesn't matter
for the individual. Where I think marketing and
sales goes in the world of AII kind of touched on it earlier
(09:28):
about this idea of a web page reacting to you.
It's a little bit more complicated on how you do that
for the main company website. Let's let's kind of shorten it
down into a landing page. Let's say you have a landing
page and and this is an idea that I think people are going to
(09:48):
have to get their head around isthis kind of template instance
model. You have a template of a landing
page, 60% of it, you said this is going to be the hero section
kind of a pricing box here. The price is 499.
Whatever it is. This is our privacy Poly that
doesn't need to be hallucinated by AI.
(10:09):
That's that's pretty, you know, standard you you set that
template up and then you have like really advanced merge
fields all over that landing page.
So it uses their name. It's got areas for their big
pain points. It's got it tailors the language
for their business, their business name and their business
industry specifically. OK.
(10:30):
And you're never sending people to a generalized landing page.
You're sending them to a customized template, an instance
of that template. And so then we get engagement
statistics from the instance. So you know, and this is what
personalization at scale really unlocks, you know that that
customer visited the website, they scrolled 3/4 of the way
(10:51):
down. They click that button and
because you've given them a completely unique instance of
that web page that no one else has ever seen.
You know how they got here. You know who they are and you
know what they've done and you have the identifiable object.
So if they hover over a particular FAQ, you, you then
use that insight to send them a follow up e-mail answering that
(11:14):
question. You will become like magic to
this person. That is a bit of a North Star
vision, right? That this, this is kind of what
we're trying to build with this product context engines to
enable you to do these things. I'm not saying this is easy and
not necessarily feasible with a lot of the technology out there
today, but we can create these instances and we can have
(11:36):
personalized tracking on there. And that's what personalization
at scale does for a business. I don't care about the analytics
of 10,000 people. I care about your analytics and
your engagement score, Mike. So then I can use it to deepen
that customer relationship off the back of that.
I agree that focusing on the individual rather than the
aggregate is one of the unlocks with this age.
(12:00):
I love the idea too of it being framed as like a magical
experience. How do you balance the desire
for this hyper personalization with the proliferation of AI
slop? So if I just go and chat you to
him, like write me some copy forthis type of persona, for this
type of business, copy paste on,it's probably bad copy.
So do you recommend people try to like have a fine-tuned model
(12:22):
they run, custom instruction they set up?
How do you try to make that AI quality elevate enough that it
is a magical experience rather than just slop?
I think we have to be mature enough to admit what the
technology actually is. There's a lot of hype around AI.
There's a lot of PhD level intelligence AI and there will
(12:43):
be no jobs and there will be no industries left.
Well, having built a lot of stuff in the last 10 years with
machine learning and AI in in lots and lots of industries at
scale, the statistical pilot trick that is the large language
model, I don't think it's going to get us to true kind of human
level in intelligence, right. So we have a tool here.
(13:07):
We don't have a new being that is arrived on planet Earth with
us that we're now Co evolving with.
We have a tool. It's a continuation of the trend
that has been going on with humanity for for millennia now.
And we need to use this tool in the way that amplifies its
strengths and minimizes its weaknesses.
If anyone's kind of been around the AI space for a long time,
(13:29):
they've heard of more of X paradox, this idea that
computers are really good at things that humans are really
bad at, and humans are really good at things that computers
are really bad at. But humans look at what
computers can do and go, wow, ifit can do, if it can play chess
and go at levels that can't be compete, like of course it can
do my job. Well, no, actually, dear human,
(13:50):
you are quite a miraculous machine yourself.
It's not very easy to compete with you in a lot of spaces,
especially as a computer. So when we understand the
limitations of the tool that we have, this statistical trick, we
have this pilot trick that we'regoing to run again and again and
again and again at scale and in different ways.
(14:10):
And the models are going to get more powerful and we're going to
have different techniques that under the hood that let us do
different slight slight of hand tricks with the tokens that come
out, the words that come out right.
And when we understand this, we,we can understand that at a base
model level, if you didn't modify the output of your AI and
(14:32):
you put the cat sat on there, everyone would get Matt because
that's the top most likely result, right?
And so if everyone has access tothe same models and everyone
gets the same output, you get regression to the mean.
I think this actually has another name because someone's
been moaning at me on social media that that's not the
correct mathematical term. I don't care.
I like regression to the mean, right?
(14:55):
The idea the regression to the mean is as people use more and
more AI and as AI gets more and more kind of it filters into the
training setting, we we get thiskind of muddy training set that
the new models are moving towards.
We get more and more and more ofthe same outputs.
That little M dash that annoys everyone so much is because
(15:16):
there's loads of M dashes in thetraining material that was used
from these ebooks, right? So that's why we get them all
the time. That's why it's a tell you're
the way you fight again against the mean again comes back to
this idea of the context engine,because you can have completely
unique statistical context to put into the model.
(15:40):
Meaning we could have one promptthat says you are a, an SDR.
You're you're a sales Rep for a company that sells windows,
right? Let's say this, this AI is going
to go out into the world and sell windows.
And if you just use the base model, it would have all the
tells and the ways of kind of writing that that everyone can
(16:03):
kind of see because humans are incredible pattern matches,
right? You know, when someone's left
you an AI comment, you don't know why, you know, but your
brain picks up on something subtle.
That experience would kind of suck.
And I think more and more there's going to be a push back
against kind of plain and AI use.
But when you go to the level where you bring that in house,
you break out-of-the-box, so to speak, of ChatGPT and Claude,
(16:27):
and you say, hang on a minute, the underlying technology is
amazing. I want to bake something custom.
I'm going to have a different temperature rating.
I'm going to use different top Psettings, which is how it kind
of switches words in and out based on the statistically the
statistical likelihood of that word appearing.
But I'm also going to have our own vector databases, this store
(16:48):
of insights, and I'm going to feed.
So I'm going to build completelyunique contexts for this AI to
feed off of and I'm going to have my own recipes for how we
transform this data. And we're going to have our own
templates for how we contain thehallucinations of the AI into
high quality outputs. If you go to that level, people
want, people can't tell that your output is AI because it
(17:12):
doesn't match the pattern that they've picked up on.
It is, I'm fascinated when I look at entrepreneurs usage of
AI and they show me their prompts and it says something
like you are a sales developmentrepresentative, you sell
windows, you are going to take the customer's name and business
and write a world class sales letter.
(17:32):
OK, that has a huge variance of success, right?
Sometimes that sales letter is really good.
Sometimes that sales layer is absolutely like mid and
sometimes it's completely awful.But with templating, the way
that I would write that prompt is here is your role.
Here is the template with the double brackets of first name,
(17:54):
biggest pain point, newest deal,amazing insight about the
product, you know, whatever it is.
And then I would give it examples of good and bad few
shot learning examples. And if you are against me in
that market and you're using the, the kind of standard GPT
output versus me with my own infrastructure in the business,
the great context around the customer, the great recipe and
(18:17):
template, like it's like bringing a knife to a gunfight.
I will win because my infrastructure is more
sophisticated than yours. It can output more personalized,
visceral content. And I think that's how you fight
against that. Yep, because it all comes down
to input output. So if you give it more
(18:37):
contextualized input, you're gonna get a better, more
targeted output. If you want to add
personalization to your own interaction with AI agents,
you're going to want to give them access to your data and
systems using MCP. And that can be a little bit
scary, which is why I've been using Toolhive.
Toolhive makes it simple and secure to use MCP.
It includes a registry of trust MCP servers that lets me
containerize any other server with a single command I can
(18:58):
install in a client. Seconds and secret protection
and network isolation are built in.
You can try Toolhive too. It's free and it's open source.
And you can learn more at Toolhive dot dev.
Now, back to the conversation with Lee.
I love that you said input output.
I've been watching a ton of Sam Ovens videos recently.
I don't know if you know who whoSam Ovens is.
He runs school.com with Alexa Mosey, incredible entrepreneur,
(19:19):
like grew grew a consulting business, I think to like $18
million on his own, like as a one like crazy guy.
And he talks about input output all of the time.
And there's a great saying in software, garbage in, garbage
out, Most people are putting garbage into their AI and they
(19:40):
go, why is this giving me AI slot?
Why is this terrible? Well, because like what you did
was terrible. And, and you know, let's take AI
video creation, for example. I've been, my Instagram is
flooded with AI videos at the minute.
But because I look at content creators who teach you how to be
a better content creator, there are a few people using AI.
(20:03):
Tools to enhance their constant content creation process.
They've done the base video filming themselves and then
they're laying AI and AI techniques on top.
Those videos are incredible. They are unbelievable.
And then there's those who just lazily say, hey GPT, give me a
video of Stephen Hawking on 1/2 pipe.
(20:26):
OK, like it's funny for five seconds I guess, but like you
get out what you put in. And This is why I'm not
concerned about AI taking jobs because there's always going to
be humans who Co create with theAI who put more in than those
who just say I'm vibe coding. Yeah, yeah, cool.
You're going to be out compete by the engineers who are who
(20:47):
know how to write software and use AI to be even better at
that. The trick I find is, is use AI
as like a conversational partnerto learn a topic deeply and then
apply that knowledge with AI assistance to build something
cool. Yeah cuz you can't just like you
said, if you just everyone has access to the same tools.
So if you just go and ask a basic question, everyone has
access to the exact same response.
(21:08):
Before we move on, you mentioneda laundry list, a few ideas of
things that can help build that advantage and build that
infrastructure. Would you mind going over a
couple tools that either you have experience with or that
you'd recommend people explore so that they can start building
up these systems before they go fully custom?
You don't need to go all the waydown to the level of building
something can like land graph, land chain, you know, a fully
(21:30):
custom coded solution. We, we at minimal viable launch,
we build stuff in that technology when it, when it is
required, but we also try and get away with the least amount
of technology in a process that we can, which is funniest
technologies, right? I love tools.
NAN is, is like one of my favorite things in the world.
(21:50):
It, it has a learning curve, butwhen you get over that and
there's a great YouTube ecosystem around NAN, which is
why I recommend it a lot. Like you can find a tutorial,
start really simple and build some basic agents that will send
you an e-mail and you will learnso much through that process.
And then you'll think, well, what else can I, what can I put
(22:11):
together? Remember one of my first one of
my first agents, I have this 10 year vision document that I work
on with, with my coach, right? I put that into an agent, it has
16 questions it asks me and thenit outputs a system that I
should build in my business to achieve that goal faster.
I thought this is a fun experiment.
Every Monday morning I sit, I call this thing system 0I sit
(22:33):
down with it and I talk to it and it says, what do you want to
talk about your business, your health, your relationship?
What there's a lot in that document, right?
That thing has made me more money than I know what to do
with because it's made me set upinvestment systems.
It's made me set up saving systems.
It's made me set up whole new lead magnets in the business.
(22:54):
It hasn't really helped on the the relationship, if I'm honest
with you. But like that simple experiment
has paid dividends. And that is a very, very, very,
very, very simple agent to make.Anyone could do that with a
YouTube tutorial and and six hours work banging the head
against the keyboard. Official first ever agent.
(23:16):
OK, explore that tool. Explore those kind of drag and
drop bills. Don't build something called
make.com or Zapier. Those are those are pretty
awesome tools as well. If you want to go deeper and
this is more technically complexand it's a little harder to
explain, but a technology that Ithink everyone needs to get very
(23:38):
aware of very quickly is something called a vector
database. And a vector database, you can
imagine a normal database of theinformation age, it's rows and
columns. It looks like an Excel
spreadsheet, right? And the reason we have things
like SharePoint and Google Driveis that unstructured data, your
(24:02):
call transcript, your video, your social media posts, your
non-smoking policy at work, What?
All that stuff lives in those abysses in your business.
You put something in SharePoint,where did it go?
Who knows? I've never seen that file again.
That stuff doesn't fit in a relational database.
You would have a very. It's hard to find all of the
(24:23):
rows and columns that would workfor that.
So a vector database solves a very particular problem for AI.
And bear with me because this isa little complicated to to
explain, but you can imagine that a vector database instead
of rows and columns is a 3D space.
It's like a cube of, of space that sits in, in your computer.
(24:44):
And if you imagine inside that cube, there's loads and little
dots. It kind of looks like the, the,
the math of the Milky Way, something like that, right?
There's points in space further back, closer towards you, up,
down, left, right, you know, all, all of this stuff.
And each of those dots represents a data point.
And what you do with a vector database is you feed your
documents, your video, your audio, whatever it is into AAI
(25:08):
model that creates what's calledan embedding.
It uses very clever algorithms to say, this is what this chunk
of a paragraph of this document,this is what it means, right?
It's about a sunrise. It's a, it's a picture of a
sunrise with a beautiful golden Meadow in front of it.
(25:29):
And it gives it a really long string of numbers, which is like
a location on a map. OK.
And that relates to a space within your vector database.
Why the hell would you want to do this?
Because it allows you to use your language to chat with your
data. So if you had thousands of
pictures in your vector database, some of them are like
(25:49):
you, your friends, you're out drinking, you're on a boat,
there's a sunrise, there's one of the filters in the forest,
there's loads of photos in there.
And you have an AI agent that says it's connected to this
vector database and you put intothe chat, pull me back all of my
photos of sunsets or sunrises. That sentence has a string, a
(26:13):
location that's very similar to the location of the photo with
the sunrise in it. They're very close in space
because they mean the same thing.
So the AI goes to that point, searches for the the data points
around it, pulls them all back and then can chat with you about
this. This is magic.
When you know what to do with this, this is how you ingest
(26:35):
every policy document in your business that you've ever had.
And you could have your HR representative chat with that
and, and create an answer for a policy at any moment in time,
within seconds. We've worked with a
sustainability consultancy here in the UK.
They have to look through thousands of documents,
thousands of pages of documents to answer questions like what's
(26:57):
the flow rate? They do skyscrapers, right?
What's the flow rate of the tapsin the bathroom on floor 102?
I don't know. OK, we go find the document.
No, now the AI has ingested every single one of those
documents and they open up a chat window and say what is that
flow rate? It goes on page 62 of document
(27:17):
ABC. It states the flow rate is bang.
This gives you super powers if you are content creator.
Imagine every video for yourself, Mike.
Imagine every podcast transcriptyou've got, you stick into a
vector database and you then have an AI agent that helps you
generate ideas for content. And you say you give it a
(27:37):
particular prompt and you say goand find me.
Every time I've spoken about personalization at scale and
this video comes up and five other videos comes up and then
it comes up with new ideas basedon that and connects them for
you. This is magic for a business and
if you it is confusing and it istechnical, but if you can get
(27:58):
your head around it, this is thetechnology that allows you to
build that world class context about your customers and your
data. I feel like more or more people
are getting exposed to this semantic searching and trying to
find the similarity between different concepts, so I'm glad
you brought it up because it is a very intimidating aspect to
get into. Just because you mentioned the
content angle, one thing that I found in the past was I would go
(28:21):
on to Andrew Huberman's site, type in a certain thing that I
want and then you just like timestamp a video and you could
watch the segment, but it lackedthe ability to chat where you
can actually say my objective, my goal is to sleep better.
And they can reference all of those specifically.
And this is what a vector database would unlock.
Also one thing just quickly before I forget on N8 N they
just recently announced the ability to do prompt to initial
(28:43):
template. So you can actually go in use
natural language to be like I want an agent to connect to my
Gmail whenever I get like a a pitch deck, start this workflow
and then it'll do the nodes. You can still drag and drop an
interface with it, but you are the 3rd guest.
I think back-to-back who's mentioned N 8 N, So you're
really pushing me towards tryingto give it a try.
It's a great platform. Like there is a, there is a
learning curve and I think the interface is a little confusing.
(29:09):
But when it clicks, because you've got these, you for the
audience to like trying to visualize this, you've got these
nodes. And when you click into the
node, you get almost like a, a split view of these two panels.
And what it is, is information in, do something with the agent
or whatever the node is and theninformation out and you're
passing information left to right.
(29:29):
And that takes a little while toget your head around.
But once you get your head around that, you go, oh, OK, I
can transform data through thesepipelines.
I can download a spreadsheet andI can run an agent on every one
of them. We, we did this for a
sustainability consultancy. They have to analyze the spend
data of their customers line by line for the entire year.
(29:51):
They said, give us a big, big Excel spreadsheet and we'll go
through and we'll give it a carbon category and we'll give
you like what is your carbon output based on how much you
spent in that category. Used to take them days.
We've now built an agent that like to 85% standard and some of
it is just doesn't have the information to do it.
And they have that they're humanwould have to go and ask the the
(30:13):
company for that. It runs through, they click it
once, it runs through line by line by line by line line and it
fills it in and they go, yeah, you know what, like that's doing
an incredible job, but it doesn't fire the sustainability
consultant. Now the sustainability
consultant spends more time prepping the data and reviewing
the data and consulting with thecustomers, which unlocks new
(30:33):
revenue opportunities. They've now become high tech,
high touch. What do you recommend as
resources people can go to online to try to get exposure to
this? How they're trying to implement
AI to help their business be more successful?
Where? Where can they go?
I had to challenge a couple of weeks back I did a presentation
to 800 businesses where they're all coaches, consultants, like
(30:55):
very non-technical people and, and, and on the older side of
life. So they weren't digital natives.
Let's say they haven't grown up with like an iPad in their in
their hand. How do you teach them to build
AI agents? Because there is there is
complexities to this stuff. What I would say if you are in
the world of ChatGPT or clawed and nanthropic, like just start
(31:18):
by creating the custom GPT startit.
That's pretty that's actually pretty simple and it can give
you some really good results. It gives you the ability to have
a system prompt and fine tuning data that you can play around
with. OK.
And then if you want to break out-of-the-box, a platform like
NAN is very, very powerful, but you are going to have more
(31:42):
confusion there. So go onto YouTube.
My, my most profitable habit in my life is the fact that I'm a
complete YouTube nerd and I willsit late on a Saturday or Sunday
night or most weeknights watching a 2 hour long tutorial
by some guy who's got like 10 subscribers on how to build
something in NAN or here's how you code something in pipe.
(32:04):
Like, all of my skills haven't really been acquired through
traditional I studied drawing for God's sake.
Like, it's not that useful, right?
But I've become a developer. I've become an engineer who's
built products that scale to millions of users and it will
shock people maybe to know that all of that information has come
from YouTube and books. But YouTube in particular I
(32:27):
think is a goldmine if you can find the right tutorial for your
problem. Someone somewhere has
experienced the exact problem you have and they probably
recorded a video about it. And you can use AI to articulate
the problem you are facing. And then you can start to
search. Because I think that the thing I
(32:47):
will say if you're going to go down this journey as a
non-technical person of how you learn to build AI or anything
technical, to brace yourself. Because part of the process and
part of the joy is getting stuckto the point where you want to
throw your laptop out the windowand you will then fix something
and it will work. It will run and you will get the
(33:11):
biggest dopamine hit you've everhad in your life because you've
overcome a very hard problem. But the, the issue in the start
of learning this is you don't know what you don't know.
So you don't know that there's avery easy method or tool that
you could use to solve the problem.
And so you struggle trying to build something custom and you
could have just pulled somethingin that someone else has done
(33:32):
and solve it in five seconds, right?
And this unknown unknowns is really difficult to overcome.
So AI is actually your friend here because you need the
language to be able to describe your problem.
Open up advanced mode ChatGPT. So this is what I'm going to
like. What how can I describe this
problem? What am I facing here?
What's the terminology? And I love using AI to
(33:54):
accelerate, accelerate my business.
I love using AI to accelerate myability to articulate my
problems because that's how it actually speeds up my learning.
Yeah, and I will always vouch for YouTube as a learning tool.
I've learned the vast majority of my career skills through it
as well. And I, I want it like like
emphasize the point they made because I find it very valid.
(34:15):
Have a, a conversation with the AI about what you want to do.
Just brain dump any type of additional thoughts, context,
questions. You can put it all in there and
then it'll help structure your thinking.
You can ask for a little learning plan.
I actually find Gemini is decentin this regard because it has
such a tight integration with YouTube that it can say, here's
my recommendation, here's some videos to support it.
I mean, once you know the properquery from Anthropic or or
(34:38):
ChatGPT, you can just go on YouTube either way.
But it's finding a workflow thatworks for you and going back to
the framing of what way do you learn best?
Do you want it to generate some images like diagrams help you
out? Do you want to explain like
you're 5 till you get the very core fundamentals?
But the curiosity and experimentation and willingness
to play and just explore, discover is going to be a
(34:59):
massive differentiator for how people benefit from this
technology. That is the key skill.
I don't know if you know who George Hotz is, the guy who like
first hacked the iPhone. He does this like he does the
streams every like week, every couple of weeks on whenever he's
what I find fascinating about him is when he is curious about
(35:19):
something, he goes deep on it, right.
And you know, maybe a slightly controversial individual in at
some points in time, but I find him fascinating to watch because
he's someone who just relentlessly follows his
curiosity and it has massively impacted his life.
I'm kind of the same thing. I this is a really silly thing
(35:41):
when I first started to learn tocode.
You get into this world called tutorial hell, you're going to
watch a tutorial, you're going to be able to build what they
build and then you're going to go great.
But it's, it doesn't actually teach you how to develop.
It teaches you how to follow tutorials really quickly and
really well. It doesn't actually teach you
how to think like a developer ortechnologist.
(36:01):
And so to escape tutorial hell, you must break out of the mold
and create something that is truly yours and go through the
struggle. You, you actually, there is a,
you have to earn this right. And I remember the thing I
wanted at the time was, was to check the value of my Tesla
investments, right? I'd got into I'd, I'd be reading
(36:24):
books about personal finance, read The richest Man in Babylon
was like investing, you know, very small amounts of money.
And I wrote a very silly little script that I think I still have
on my, in my Apple notes file somewhere right where at the
bottom that basically knew the static value of my shares.
And it did a very simple API call and it looked up the value
of Tesla shares right now and did a calculation.
(36:46):
And I remember when I wrote thatand it run and I had this
experience of like, Oh my God, Ican now build anything in the
world. What's cool about that?
I didn't have AI at the time. That was really hard to do.
What's cool about AI now is you can kind of shortcut the route
out of tutorial help, follow thetutorial, get so far.
(37:10):
Then when you are curious, let'ssay you're making a tutorial
where it's like a, hey, I'm going to build an AI agent that
sends a very simple Slack message.
I want something that when I click this button, it sends us
like that's a good project to start with and you do a tutorial
on that. But then you open up AI and you
say, hey, I've got this system, It sends a Slack message.
(37:32):
I want to be able to send a file.
Help me extend this and it will help you get out of the tutorial
and get into your own lane. Then you are actually learning.
The trap is when you then start trying to rely too much on AI to
do the thinking for you and you become lazy and this is hard.
And then AI has done the learning you haven't.
Now you are someone who can control AI and do some stuff.
(37:56):
Maybe it's valid and try to retain control, but use AI to
break out those tutorials to customize the tutorial for you.
And ask it questions, ask your definitions, get the why out of
it. Like try try to go deep and then
you never know where else in your journey it'll apply.
Lee, I had a ton of fun with this.
(38:18):
Really appreciate you sharing the insights.
Before we let you go, is there anything you want the audience
to know? I've loved this conversation.
Mike has been great as always toto talk about these subjects.
Yeah, we help people build thesesystems into their businesses
that allow them to do high tech,high touch personalization at
scale. These context engines for these
businesses. We run a monthly master class
(38:40):
where we tell people, we show people how to disrupt their
industry with AI agents in that way.
Not to fire your team, not to become more efficient, but to
blow your customers socks off bycreating remarkable customer
experiences. I would love to see you there.
You can find links for that master class at our website
Minimal Viable launch dot IO. Thank you for listening to
(39:02):
conversation with Lee Russell. I'm curious if you've tried any
techniques to do personalizationat scale.
I really believe it is going to be a differentiator as we move
forward into this AIH. How can you make your feel like
a one of one, very special, veryfocused and bring that magic to
the experience? I want to give a quick shout out
to Tool High for supporting the show so I can have conversations
like this and I'll see you next week.
Whoever engineers the best context around their customer in
this new age is going to be the winner.
AI allows you to be high tech and high touch, so you can do
personalization at scale and this is the technology that
allows you to build that world class context about your
customers and your data. Using AI for efficiency gains is
(00:20):
something we've talked a lot about on this channel, but
there's so much more we can leverage AI for.
What about making digital experiences more magical?
More human. One of the promises of AI is for
personalization at scale. But how do we do it?
So in episode 62, Tool Use, brought to you by Tool Hive,
we're joined by Lee Russell. Lee is the founder of Minimal
Viable Launch. He's worked in AI for over a
decade and he's worked with someof the fastest growing companies
(00:41):
to integrate AI into their businesses.
So today we're going to discuss how you can use AI for that
personalization at scale, what tools you need, what things you
need to consider, what mindset shifts you need.
So I really hope you enjoy this conversation with Leave Russell.
What's fascinating to me is someof the biggest innovations that
I've been in charge of developing.
The things that give those outsize returns.
(01:02):
It's a small amount of effort and gives you wild results.
Very rarely is the crazy North Star Hail Mary vision.
It's usually a series of small tweaks and small experience
gains that compound on top of each other that say finally,
this is now a consumable experience by by your by your
(01:24):
customers that they love that they tell others about the area
that we are kind of getting people to really focus on.
In terms of where I think the low hanging fruit in with with
AI is, is this moving into this direction of personalization at
scale, this idea that before AI you had a choice.
(01:48):
You could be a highly personalized business which kept
intentionally small charged highticket you, you serve fewer
customers, but you know their names, you know when their dog's
birthday is. It's still the kind of high,
high touch white glove service or you have the choice of let's
be really scalable. And so to be scalable, we're
going to be templated, we're going to be automated.
(02:10):
We're going to lose some of thatpersonalization, but we can
serve many, many more customers in doing so.
And as you move into that kind of more unpersonalized service,
you close less people, you get less loyal customers, you know
that you do lose something there.
AI allows you to be high tech and high touch, so you can do
(02:32):
personalization at scale. And I think a lot of businesses,
if they really thought about what that means, it changes the
game for marketing. It changes their entire sales
pipeline of how you have these customer conversations and then
follow up with them in individualized ways.
It changes your nurture sequencewithin your business massively
(02:53):
and it changes, it changes your product delivery system.
And, and this idea of treating your customers not as the
faceless mass, but as the individual, I think is the big
change that happens in the intelligence era.
People are going to, people are going to look back in 10 to 15
years time and go, do you remember in the 20 twenties when
(03:15):
they had like static landing pages that everyone visited, how
crazy it didn't change to that user's needs at all.
That's going to be a wild thing to look back on.
I love the idea of personalization at scale because
we've all received those emails where we'll be like, hi and then
A brackets first name because something wasn't parsed
correctly. And I feel like like that's an
easy example. Just by addressing someone by
(03:36):
their name in an e-mail, you're have a better likelihood of
getting a success. What other avenues should people
look into in their businesses orin their departments to add some
personalization to the experience?
Well, let's take the e-mail example because I think we've
all had the other form of personalization right where it,
it has your own merge name filled in it, but it's
essentially an e-mail. Like Haley, I saw you're the
(04:00):
owner of minimal viable launch. Great business.
Love what you guys are doing. Anyway, do you want to buy my
stuff? And you're like, what is this
e-mail? Why did you even try and
personalize it like that? It's like the the effort wasn't
worth the juice, wasn't worth the squeeze In the world of AI
marketing. This, this idea of being high
(04:21):
tech, high touch, the information age way of doing
this is we create ACRM of 10,000customers.
We've got their e-mail addresses, we know their first
name. We create one e-mail, we
customize it with the merge fields, one or two points and we
send one e-mail to 10,000 peopleand we hope that one to 2% of
(04:41):
those people click on it, click through to it and make me book a
sales call for that nurture, right?
It's spray and pray. One to many, highly
undifferentiated. The world of high tech, high
touch marketing is very, very different.
You're, you still have the 10,000 people in your CRM, but
you write 10,000 emails and sendthem once.
(05:05):
Everyone gets a pretty individualized experience.
And, and when I talk to marketers about this, like they,
they open their eyes in fear. I, I, I thought they would get
really excited about this, that this terrifies them because they
go, how do we track metrics? How do we track open rates?
Who cares? Who cares about most of that
stuff, right? What we want is to set up an
(05:29):
infrastructure that allows you to send individualized, like
truly individualized touch points to your customers.
And what we often talk about in in our company here at minimal
viable launches, whoever engineers the best context
around their customer in this new age is going to be the
winner. It's not going to be who has the
(05:50):
best AI technology unless you work in open AI, Anthropic,
Google Deep, you know deep mind,whatever.
You don't have access to the true state-of-the-art.
Everyone else has what is the commercial state-of-the-art and
it's one line of code to switch between them.
Who cares if you're loyal to an AI company?
I don't know why because like they're all pretty good.
(06:12):
Just go who has the the least token expense, right?
Don't be loyal. So the models are not your MO.
What is going to be your MO is who can become cleverest at
generating unstructured qualitative data about their
customers. This is your sales calls,
transcripts. It is doing quizzes and other
(06:35):
kind of fun ways of interacting with a customer that allows them
to give you open text entries about their pain points and
their problems so that you can do something in your business.
We call them context engines. You can combine those insights
about your customer and the facts from your CRM into a
template of an e-mail that you want to write and you have 40%
(06:57):
of that e-mail is templated out.You know, it's good high quality
stuff. 60% of that is the AIS hallucination between the
insight and the fact. And they get incredibly
customized personalized outreach.
And then it becomes the game of how do we get more of this?
Some of the experiments we've done with businesses around
(07:18):
this, they get a 20% conversion rate on their emails, 20%.
It's crazy. Now it's new.
That is going to go down over time, I guarantee you.
But right now, if you're brave enough to be an early mover, you
can go and grow your business dramatically by just not doing
(07:39):
terrible e-mail outreach and being clever about how you get
that context of your customer. Could we go into a little bit
deeper because you mentioned thesales calls is a valid form of
collecting this context as well as open-ended answers and in
forms or surveys or however elsedo you see value in things like
setting up Google Analytics so we can see how they interact
(08:01):
with their website or those datapoints kind of outdated in this
AI age? Google Analytics I, I love
Google Analytics and I'm not an SEO expert and this is not my
world. So what I'm going to say is
going to ruffle some feathers and people will probably tell me
why I'm all sorts of wrong, but this is my experience with it.
Google Analytics always shows you the past events so you can
(08:23):
see things have happened, but you very rarely know why they
happened or how they happened. And it's kind of this soup and,
and I, I liken it to a bit of a mental challenge, right?
We got a pool table with a load of, with the, the balls are all
kind of spread all over the place, right?
And you've been blindfolded, putout the room.
(08:44):
I whack the, the cue ball. They go everywhere.
You haven't seen it. I bring you in and I say, Mike,
tell me the exact path these balls took to reach this
position they're in. It's an impossible endeavour.
You, you can't do that after thefact.
You have to observe in the moment.
So Google Analytics is useful for so far, but I think the
entire industry of getting obsessed with these kind of
(09:05):
statistics is a little bit, let's just say misguided.
Because as far as what my skill set is concerned, my
understanding of statistics and these trends is like it.
It tells you what happened at scale, but that doesn't matter
for the individual. Where I think marketing and
sales goes in the world of AII kind of touched on it earlier
(09:28):
about this idea of a web page reacting to you.
It's a little bit more complicated on how you do that
for the main company website. Let's let's kind of shorten it
down into a landing page. Let's say you have a landing
page and and this is an idea that I think people are going to
(09:48):
have to get their head around isthis kind of template instance
model. You have a template of a landing
page, 60% of it, you said this is going to be the hero section
kind of a pricing box here. The price is 499.
Whatever it is. This is our privacy Poly that
doesn't need to be hallucinated by AI.
(10:09):
That's that's pretty, you know, standard you you set that
template up and then you have like really advanced merge
fields all over that landing page.
So it uses their name. It's got areas for their big
pain points. It's got it tailors the language
for their business, their business name and their business
industry specifically. OK.
(10:30):
And you're never sending people to a generalized landing page.
You're sending them to a customized template, an instance
of that template. And so then we get engagement
statistics from the instance. So you know, and this is what
personalization at scale really unlocks, you know that that
customer visited the website, they scrolled 3/4 of the way
(10:51):
down. They click that button and
because you've given them a completely unique instance of
that web page that no one else has ever seen.
You know how they got here. You know who they are and you
know what they've done and you have the identifiable object.
So if they hover over a particular FAQ, you, you then
use that insight to send them a follow up e-mail answering that
(11:14):
question. You will become like magic to
this person. That is a bit of a North Star
vision, right? That this, this is kind of what
we're trying to build with this product context engines to
enable you to do these things. I'm not saying this is easy and
not necessarily feasible with a lot of the technology out there
today, but we can create these instances and we can have
(11:36):
personalized tracking on there. And that's what personalization
at scale does for a business. I don't care about the analytics
of 10,000 people. I care about your analytics and
your engagement score, Mike. So then I can use it to deepen
that customer relationship off the back of that.
I agree that focusing on the individual rather than the
aggregate is one of the unlocks with this age.
(12:00):
I love the idea too of it being framed as like a magical
experience. How do you balance the desire
for this hyper personalization with the proliferation of AI
slop? So if I just go and chat you to
him, like write me some copy forthis type of persona, for this
type of business, copy paste on,it's probably bad copy.
So do you recommend people try to like have a fine-tuned model
(12:22):
they run, custom instruction they set up?
How do you try to make that AI quality elevate enough that it
is a magical experience rather than just slop?
I think we have to be mature enough to admit what the
technology actually is. There's a lot of hype around AI.
There's a lot of PhD level intelligence AI and there will
(12:43):
be no jobs and there will be no industries left.
Well, having built a lot of stuff in the last 10 years with
machine learning and AI in in lots and lots of industries at
scale, the statistical pilot trick that is the large language
model, I don't think it's going to get us to true kind of human
level in intelligence, right. So we have a tool here.
(13:07):
We don't have a new being that is arrived on planet Earth with
us that we're now Co evolving with.
We have a tool. It's a continuation of the trend
that has been going on with humanity for for millennia now.
And we need to use this tool in the way that amplifies its
strengths and minimizes its weaknesses.
If anyone's kind of been around the AI space for a long time,
(13:29):
they've heard of more of X paradox, this idea that
computers are really good at things that humans are really
bad at, and humans are really good at things that computers
are really bad at. But humans look at what
computers can do and go, wow, ifit can do, if it can play chess
and go at levels that can't be compete, like of course it can
do my job. Well, no, actually, dear human,
(13:50):
you are quite a miraculous machine yourself.
It's not very easy to compete with you in a lot of spaces,
especially as a computer. So when we understand the
limitations of the tool that we have, this statistical trick, we
have this pilot trick that we'regoing to run again and again and
again and again at scale and in different ways.
(14:10):
And the models are going to get more powerful and we're going to
have different techniques that under the hood that let us do
different slight slight of hand tricks with the tokens that come
out, the words that come out right.
And when we understand this, we,we can understand that at a base
model level, if you didn't modify the output of your AI and
(14:32):
you put the cat sat on there, everyone would get Matt because
that's the top most likely result, right?
And so if everyone has access tothe same models and everyone
gets the same output, you get regression to the mean.
I think this actually has another name because someone's
been moaning at me on social media that that's not the
correct mathematical term. I don't care.
I like regression to the mean, right?
(14:55):
The idea the regression to the mean is as people use more and
more AI and as AI gets more and more kind of it filters into the
training setting, we we get thiskind of muddy training set that
the new models are moving towards.
We get more and more and more ofthe same outputs.
That little M dash that annoys everyone so much is because
(15:16):
there's loads of M dashes in thetraining material that was used
from these ebooks, right? So that's why we get them all
the time. That's why it's a tell you're
the way you fight again against the mean again comes back to
this idea of the context engine,because you can have completely
unique statistical context to put into the model.
(15:40):
Meaning we could have one promptthat says you are a, an SDR.
You're you're a sales Rep for a company that sells windows,
right? Let's say this, this AI is going
to go out into the world and sell windows.
And if you just use the base model, it would have all the
tells and the ways of kind of writing that that everyone can
(16:03):
kind of see because humans are incredible pattern matches,
right? You know, when someone's left
you an AI comment, you don't know why, you know, but your
brain picks up on something subtle.
That experience would kind of suck.
And I think more and more there's going to be a push back
against kind of plain and AI use.
But when you go to the level where you bring that in house,
you break out-of-the-box, so to speak, of ChatGPT and Claude,
(16:27):
and you say, hang on a minute, the underlying technology is
amazing. I want to bake something custom.
I'm going to have a different temperature rating.
I'm going to use different top Psettings, which is how it kind
of switches words in and out based on the statistically the
statistical likelihood of that word appearing.
But I'm also going to have our own vector databases, this store
(16:48):
of insights, and I'm going to feed.
So I'm going to build completelyunique contexts for this AI to
feed off of and I'm going to have my own recipes for how we
transform this data. And we're going to have our own
templates for how we contain thehallucinations of the AI into
high quality outputs. If you go to that level, people
want, people can't tell that your output is AI because it
(17:12):
doesn't match the pattern that they've picked up on.
It is, I'm fascinated when I look at entrepreneurs usage of
AI and they show me their prompts and it says something
like you are a sales developmentrepresentative, you sell
windows, you are going to take the customer's name and business
and write a world class sales letter.
(17:32):
OK, that has a huge variance of success, right?
Sometimes that sales letter is really good.
Sometimes that sales layer is absolutely like mid and
sometimes it's completely awful.But with templating, the way
that I would write that prompt is here is your role.
Here is the template with the double brackets of first name,
(17:54):
biggest pain point, newest deal,amazing insight about the
product, you know, whatever it is.
And then I would give it examples of good and bad few
shot learning examples. And if you are against me in
that market and you're using the, the kind of standard GPT
output versus me with my own infrastructure in the business,
the great context around the customer, the great recipe and
(18:17):
template, like it's like bringing a knife to a gunfight.
I will win because my infrastructure is more
sophisticated than yours. It can output more personalized,
visceral content. And I think that's how you fight
against that. Yep, because it all comes down
to input output. So if you give it more
(18:37):
contextualized input, you're gonna get a better, more
targeted output. If you want to add
personalization to your own interaction with AI agents,
you're going to want to give them access to your data and
systems using MCP. And that can be a little bit
scary, which is why I've been using Toolhive.
Toolhive makes it simple and secure to use MCP.
It includes a registry of trust MCP servers that lets me
containerize any other server with a single command I can
(18:58):
install in a client. Seconds and secret protection
and network isolation are built in.
You can try Toolhive too. It's free and it's open source.
And you can learn more at Toolhive dot dev.
Now, back to the conversation with Lee.
I love that you said input output.
I've been watching a ton of Sam Ovens videos recently.
I don't know if you know who whoSam Ovens is.
He runs school.com with Alexa Mosey, incredible entrepreneur,
(19:19):
like grew grew a consulting business, I think to like $18
million on his own, like as a one like crazy guy.
And he talks about input output all of the time.
And there's a great saying in software, garbage in, garbage
out, Most people are putting garbage into their AI and they
(19:40):
go, why is this giving me AI slot?
Why is this terrible? Well, because like what you did
was terrible. And, and you know, let's take AI
video creation, for example. I've been, my Instagram is
flooded with AI videos at the minute.
But because I look at content creators who teach you how to be
a better content creator, there are a few people using AI.
(20:03):
Tools to enhance their constant content creation process.
They've done the base video filming themselves and then
they're laying AI and AI techniques on top.
Those videos are incredible. They are unbelievable.
And then there's those who just lazily say, hey GPT, give me a
video of Stephen Hawking on 1/2 pipe.
(20:26):
OK, like it's funny for five seconds I guess, but like you
get out what you put in. And This is why I'm not
concerned about AI taking jobs because there's always going to
be humans who Co create with theAI who put more in than those
who just say I'm vibe coding. Yeah, yeah, cool.
You're going to be out compete by the engineers who are who
(20:47):
know how to write software and use AI to be even better at
that. The trick I find is, is use AI
as like a conversational partnerto learn a topic deeply and then
apply that knowledge with AI assistance to build something
cool. Yeah cuz you can't just like you
said, if you just everyone has access to the same tools.
So if you just go and ask a basic question, everyone has
access to the exact same response.
(21:08):
Before we move on, you mentioneda laundry list, a few ideas of
things that can help build that advantage and build that
infrastructure. Would you mind going over a
couple tools that either you have experience with or that
you'd recommend people explore so that they can start building
up these systems before they go fully custom?
You don't need to go all the waydown to the level of building
something can like land graph, land chain, you know, a fully
(21:30):
custom coded solution. We, we at minimal viable launch,
we build stuff in that technology when it, when it is
required, but we also try and get away with the least amount
of technology in a process that we can, which is funniest
technologies, right? I love tools.
NAN is, is like one of my favorite things in the world.
(21:50):
It, it has a learning curve, butwhen you get over that and
there's a great YouTube ecosystem around NAN, which is
why I recommend it a lot. Like you can find a tutorial,
start really simple and build some basic agents that will send
you an e-mail and you will learnso much through that process.
And then you'll think, well, what else can I, what can I put
(22:11):
together? Remember one of my first one of
my first agents, I have this 10 year vision document that I work
on with, with my coach, right? I put that into an agent, it has
16 questions it asks me and thenit outputs a system that I
should build in my business to achieve that goal faster.
I thought this is a fun experiment.
Every Monday morning I sit, I call this thing system 0I sit
(22:33):
down with it and I talk to it and it says, what do you want to
talk about your business, your health, your relationship?
What there's a lot in that document, right?
That thing has made me more money than I know what to do
with because it's made me set upinvestment systems.
It's made me set up saving systems.
It's made me set up whole new lead magnets in the business.
(22:54):
It hasn't really helped on the the relationship, if I'm honest
with you. But like that simple experiment
has paid dividends. And that is a very, very, very,
very, very simple agent to make.Anyone could do that with a
YouTube tutorial and and six hours work banging the head
against the keyboard. Official first ever agent.
(23:16):
OK, explore that tool. Explore those kind of drag and
drop bills. Don't build something called
make.com or Zapier. Those are those are pretty
awesome tools as well. If you want to go deeper and
this is more technically complexand it's a little harder to
explain, but a technology that Ithink everyone needs to get very
(23:38):
aware of very quickly is something called a vector
database. And a vector database, you can
imagine a normal database of theinformation age, it's rows and
columns. It looks like an Excel
spreadsheet, right? And the reason we have things
like SharePoint and Google Driveis that unstructured data, your
(24:02):
call transcript, your video, your social media posts, your
non-smoking policy at work, What?
All that stuff lives in those abysses in your business.
You put something in SharePoint,where did it go?
Who knows? I've never seen that file again.
That stuff doesn't fit in a relational database.
You would have a very. It's hard to find all of the
(24:23):
rows and columns that would workfor that.
So a vector database solves a very particular problem for AI.
And bear with me because this isa little complicated to to
explain, but you can imagine that a vector database instead
of rows and columns is a 3D space.
It's like a cube of, of space that sits in, in your computer.
(24:44):
And if you imagine inside that cube, there's loads and little
dots. It kind of looks like the, the,
the math of the Milky Way, something like that, right?
There's points in space further back, closer towards you, up,
down, left, right, you know, all, all of this stuff.
And each of those dots represents a data point.
And what you do with a vector database is you feed your
documents, your video, your audio, whatever it is into AAI
(25:08):
model that creates what's calledan embedding.
It uses very clever algorithms to say, this is what this chunk
of a paragraph of this document,this is what it means, right?
It's about a sunrise. It's a, it's a picture of a
sunrise with a beautiful golden Meadow in front of it.
(25:29):
And it gives it a really long string of numbers, which is like
a location on a map. OK.
And that relates to a space within your vector database.
Why the hell would you want to do this?
Because it allows you to use your language to chat with your
data. So if you had thousands of
pictures in your vector database, some of them are like
(25:49):
you, your friends, you're out drinking, you're on a boat,
there's a sunrise, there's one of the filters in the forest,
there's loads of photos in there.
And you have an AI agent that says it's connected to this
vector database and you put intothe chat, pull me back all of my
photos of sunsets or sunrises. That sentence has a string, a
(26:13):
location that's very similar to the location of the photo with
the sunrise in it. They're very close in space
because they mean the same thing.
So the AI goes to that point, searches for the the data points
around it, pulls them all back and then can chat with you about
this. This is magic.
When you know what to do with this, this is how you ingest
(26:35):
every policy document in your business that you've ever had.
And you could have your HR representative chat with that
and, and create an answer for a policy at any moment in time,
within seconds. We've worked with a
sustainability consultancy here in the UK.
They have to look through thousands of documents,
thousands of pages of documents to answer questions like what's
(26:57):
the flow rate? They do skyscrapers, right?
What's the flow rate of the tapsin the bathroom on floor 102?
I don't know. OK, we go find the document.
No, now the AI has ingested every single one of those
documents and they open up a chat window and say what is that
flow rate? It goes on page 62 of document
(27:17):
ABC. It states the flow rate is bang.
This gives you super powers if you are content creator.
Imagine every video for yourself, Mike.
Imagine every podcast transcriptyou've got, you stick into a
vector database and you then have an AI agent that helps you
generate ideas for content. And you say you give it a
(27:37):
particular prompt and you say goand find me.
Every time I've spoken about personalization at scale and
this video comes up and five other videos comes up and then
it comes up with new ideas basedon that and connects them for
you. This is magic for a business and
if you it is confusing and it istechnical, but if you can get
(27:58):
your head around it, this is thetechnology that allows you to
build that world class context about your customers and your
data. I feel like more or more people
are getting exposed to this semantic searching and trying to
find the similarity between different concepts, so I'm glad
you brought it up because it is a very intimidating aspect to
get into. Just because you mentioned the
content angle, one thing that I found in the past was I would go
(28:21):
on to Andrew Huberman's site, type in a certain thing that I
want and then you just like timestamp a video and you could
watch the segment, but it lackedthe ability to chat where you
can actually say my objective, my goal is to sleep better.
And they can reference all of those specifically.
And this is what a vector database would unlock.
Also one thing just quickly before I forget on N8 N they
just recently announced the ability to do prompt to initial
(28:43):
template. So you can actually go in use
natural language to be like I want an agent to connect to my
Gmail whenever I get like a a pitch deck, start this workflow
and then it'll do the nodes. You can still drag and drop an
interface with it, but you are the 3rd guest.
I think back-to-back who's mentioned N 8 N, So you're
really pushing me towards tryingto give it a try.
It's a great platform. Like there is a, there is a
learning curve and I think the interface is a little confusing.
(29:09):
But when it clicks, because you've got these, you for the
audience to like trying to visualize this, you've got these
nodes. And when you click into the
node, you get almost like a, a split view of these two panels.
And what it is, is information in, do something with the agent
or whatever the node is and theninformation out and you're
passing information left to right.
(29:29):
And that takes a little while toget your head around.
But once you get your head around that, you go, oh, OK, I
can transform data through thesepipelines.
I can download a spreadsheet andI can run an agent on every one
of them. We, we did this for a
sustainability consultancy. They have to analyze the spend
data of their customers line by line for the entire year.
(29:51):
They said, give us a big, big Excel spreadsheet and we'll go
through and we'll give it a carbon category and we'll give
you like what is your carbon output based on how much you
spent in that category. Used to take them days.
We've now built an agent that like to 85% standard and some of
it is just doesn't have the information to do it.
And they have that they're humanwould have to go and ask the the
(30:13):
company for that. It runs through, they click it
once, it runs through line by line by line by line line and it
fills it in and they go, yeah, you know what, like that's doing
an incredible job, but it doesn't fire the sustainability
consultant. Now the sustainability
consultant spends more time prepping the data and reviewing
the data and consulting with thecustomers, which unlocks new
(30:33):
revenue opportunities. They've now become high tech,
high touch. What do you recommend as
resources people can go to online to try to get exposure to
this? How they're trying to implement
AI to help their business be more successful?
Where? Where can they go?
I had to challenge a couple of weeks back I did a presentation
to 800 businesses where they're all coaches, consultants, like
(30:55):
very non-technical people and, and, and on the older side of
life. So they weren't digital natives.
Let's say they haven't grown up with like an iPad in their in
their hand. How do you teach them to build
AI agents? Because there is there is
complexities to this stuff. What I would say if you are in
the world of ChatGPT or clawed and nanthropic, like just start
(31:18):
by creating the custom GPT startit.
That's pretty that's actually pretty simple and it can give
you some really good results. It gives you the ability to have
a system prompt and fine tuning data that you can play around
with. OK.
And then if you want to break out-of-the-box, a platform like
NAN is very, very powerful, but you are going to have more
(31:42):
confusion there. So go onto YouTube.
My, my most profitable habit in my life is the fact that I'm a
complete YouTube nerd and I willsit late on a Saturday or Sunday
night or most weeknights watching a 2 hour long tutorial
by some guy who's got like 10 subscribers on how to build
something in NAN or here's how you code something in pipe.
(32:04):
Like, all of my skills haven't really been acquired through
traditional I studied drawing for God's sake.
Like, it's not that useful, right?
But I've become a developer. I've become an engineer who's
built products that scale to millions of users and it will
shock people maybe to know that all of that information has come
from YouTube and books. But YouTube in particular I
(32:27):
think is a goldmine if you can find the right tutorial for your
problem. Someone somewhere has
experienced the exact problem you have and they probably
recorded a video about it. And you can use AI to articulate
the problem you are facing. And then you can start to
search. Because I think that the thing I
(32:47):
will say if you're going to go down this journey as a
non-technical person of how you learn to build AI or anything
technical, to brace yourself. Because part of the process and
part of the joy is getting stuckto the point where you want to
throw your laptop out the windowand you will then fix something
and it will work. It will run and you will get the
(33:11):
biggest dopamine hit you've everhad in your life because you've
overcome a very hard problem. But the, the issue in the start
of learning this is you don't know what you don't know.
So you don't know that there's avery easy method or tool that
you could use to solve the problem.
And so you struggle trying to build something custom and you
could have just pulled somethingin that someone else has done
(33:32):
and solve it in five seconds, right?
And this unknown unknowns is really difficult to overcome.
So AI is actually your friend here because you need the
language to be able to describe your problem.
Open up advanced mode ChatGPT. So this is what I'm going to
like. What how can I describe this
problem? What am I facing here?
What's the terminology? And I love using AI to
(33:54):
accelerate, accelerate my business.
I love using AI to accelerate myability to articulate my
problems because that's how it actually speeds up my learning.
Yeah, and I will always vouch for YouTube as a learning tool.
I've learned the vast majority of my career skills through it
as well. And I, I want it like like
emphasize the point they made because I find it very valid.
(34:15):
Have a, a conversation with the AI about what you want to do.
Just brain dump any type of additional thoughts, context,
questions. You can put it all in there and
then it'll help structure your thinking.
You can ask for a little learning plan.
I actually find Gemini is decentin this regard because it has
such a tight integration with YouTube that it can say, here's
my recommendation, here's some videos to support it.
I mean, once you know the properquery from Anthropic or or
(34:38):
ChatGPT, you can just go on YouTube either way.
But it's finding a workflow thatworks for you and going back to
the framing of what way do you learn best?
Do you want it to generate some images like diagrams help you
out? Do you want to explain like
you're 5 till you get the very core fundamentals?
But the curiosity and experimentation and willingness
to play and just explore, discover is going to be a
(34:59):
massive differentiator for how people benefit from this
technology. That is the key skill.
I don't know if you know who George Hotz is, the guy who like
first hacked the iPhone. He does this like he does the
streams every like week, every couple of weeks on whenever he's
what I find fascinating about him is when he is curious about
(35:19):
something, he goes deep on it, right.
And you know, maybe a slightly controversial individual in at
some points in time, but I find him fascinating to watch because
he's someone who just relentlessly follows his
curiosity and it has massively impacted his life.
I'm kind of the same thing. I this is a really silly thing
(35:41):
when I first started to learn tocode.
You get into this world called tutorial hell, you're going to
watch a tutorial, you're going to be able to build what they
build and then you're going to go great.
But it's, it doesn't actually teach you how to develop.
It teaches you how to follow tutorials really quickly and
really well. It doesn't actually teach you
how to think like a developer ortechnologist.
(36:01):
And so to escape tutorial hell, you must break out of the mold
and create something that is truly yours and go through the
struggle. You, you actually, there is a,
you have to earn this right. And I remember the thing I
wanted at the time was, was to check the value of my Tesla
investments, right? I'd got into I'd, I'd be reading
(36:24):
books about personal finance, read The richest Man in Babylon
was like investing, you know, very small amounts of money.
And I wrote a very silly little script that I think I still have
on my, in my Apple notes file somewhere right where at the
bottom that basically knew the static value of my shares.
And it did a very simple API call and it looked up the value
of Tesla shares right now and did a calculation.
(36:46):
And I remember when I wrote thatand it run and I had this
experience of like, Oh my God, Ican now build anything in the
world. What's cool about that?
I didn't have AI at the time. That was really hard to do.
What's cool about AI now is you can kind of shortcut the route
out of tutorial help, follow thetutorial, get so far.
(37:10):
Then when you are curious, let'ssay you're making a tutorial
where it's like a, hey, I'm going to build an AI agent that
sends a very simple Slack message.
I want something that when I click this button, it sends us
like that's a good project to start with and you do a tutorial
on that. But then you open up AI and you
say, hey, I've got this system, It sends a Slack message.
(37:32):
I want to be able to send a file.
Help me extend this and it will help you get out of the tutorial
and get into your own lane. Then you are actually learning.
The trap is when you then start trying to rely too much on AI to
do the thinking for you and you become lazy and this is hard.
And then AI has done the learning you haven't.
Now you are someone who can control AI and do some stuff.
(37:56):
Maybe it's valid and try to retain control, but use AI to
break out those tutorials to customize the tutorial for you.
And ask it questions, ask your definitions, get the why out of
it. Like try try to go deep and then
you never know where else in your journey it'll apply.
Lee, I had a ton of fun with this.
(38:18):
Really appreciate you sharing the insights.
Before we let you go, is there anything you want the audience
to know? I've loved this conversation.
Mike has been great as always toto talk about these subjects.
Yeah, we help people build thesesystems into their businesses
that allow them to do high tech,high touch personalization at
scale. These context engines for these
businesses. We run a monthly master class
(38:40):
where we tell people, we show people how to disrupt their
industry with AI agents in that way.
Not to fire your team, not to become more efficient, but to
blow your customers socks off bycreating remarkable customer
experiences. I would love to see you there.
You can find links for that master class at our website
Minimal Viable launch dot IO. Thank you for listening to
(39:02):
conversation with Lee Russell. I'm curious if you've tried any
techniques to do personalizationat scale.
I really believe it is going to be a differentiator as we move
forward into this AIH. How can you make your feel like
a one of one, very special, veryfocused and bring that magic to
the experience? I want to give a quick shout out
to Tool High for supporting the show so I can have conversations
like this and I'll see you next week.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Paper Ghosts: The Texas Teen Murders
Paper Ghosts: The Texas Teen Murders takes you back to 1983, when two teenagers were found murdered, execution-style, on a quiet Texas hill. What followed was decades of rumors, false leads, and a case that law enforcement could never seem to close. Now, veteran investigative journalist M. William Phelps reopens the file — uncovering new witnesses, hidden evidence, and a shocking web of deaths that may all be connected. Over nine gripping episodes, Paper Ghosts: The Texas Teen Murders unravels a story 42 years in the making… and asks the question: who’s really been hiding the truth?
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com