October 14, 2025 • 41 mins
Join the Tool Use Discord: https://discord.gg/PnEGyXpjaX
Struggling to cut through the AI hype to find real operational efficiency for your business? In this episode, we're joined by Csongor Barabasi, CEO of Bonsai Labs, an AI value creation partner for private equity firms. Csongor shares his expert framework for deploying AI agents to achieve massive productivity gains.
Learn the practical steps to progress from a simple AI co-pilot to a fully autonomous AI agent , and discover how to create a 'golden data set' and feedback loops to make your systems smarter over time. We cover the critical build vs. buy debate , strategies for process automation , data security considerations , calculating ROI for AI projects , and enabling non-technical teams to build custom workflows. This is your ultimate guide to using AI for tangible business results.
Csongor is the founder/CEO of Bonsai Labs, a specialized AI transformation and value creation firm working with Private Equity portfolio companies. With engineering excellence at it's core, the they deliver AI workflow automations and build custom AI products with tangible impact within weeks.
Bonsai Labs: https://bonsai-labs.com
Csongor: https://linkedin.com/in/barabasicsongor
Connect with us
https://x.com/ToolUsePodcast
https://x.com/MikeBirdTech
https://x.com/barabasicsongor
00:00:00 - Intro
00:02:40 - The Framework: Co-Pilot to Fully Autonomous Agent
00:11:26 - AI Solutions: The Build vs. Buy Debate
00:16:03 - AI for Non-Technical Teams
00:27:40 - How to Document a Process for Automation
00:31:36 - How to Measure the ROI of AI
Subscribe for more insights on AI tools, productivity, and process automation.
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):
Can you figure out a feedback loop to make the system smarter
over time? Maybe you want to get it down to
30 minutes or 10 minutes driven by AI and then scale it across
all the employees. Then you get into the agentic
mode, which is now the AI operates stand alone and then
basically your company can get into a management by exception.
(00:20):
So that's why we need AI which will take away the boring work,
so you can be happy and that we can continue building a
successful business. There is so much AI hype out
there when business leaders wantto know how do they cut through
the noise to find the true efficiency gains?
Do they build? Do they buy?
How do they know what to choose?How do you even know where to
apply AI in the 1st place? Well, in episode 61 of Tool Use
(00:40):
brought by Tool Hive, we're solving all of this and more.
We're joined by Chongo Barabasi,the founder and CEO of Bonsai
Labs. They work with private equity
firms as the AI value creation partner, deploying AI across all
their portfolio companies. So he has a lot of experience on
how to actually use AI for operational efficiency.
So please do this conversation with Chongo I.
Feel that agents has being the talk almost from day one of
(01:03):
ChatGPT everyone was talking about agents for me at some
point. It also felt like this is an
overhyped work, that it's like abuzzword that everyone is using.
Because ultimately if you boil down what an agent is, is just a
sequence of LLM calls. There is nothing very complex or
rocket science about it. Right?
(01:24):
You could be an agent by just copy pasting a prompt, getting
an output, putting it into the next prompt, doing that again.
But nevertheless, even this, this orchestration you could say
of NLM that led to an agentic behaviour.
And what I realized over time isthat really when we call them
agents, it's just a very creative way of how you piece
(01:45):
together NLM calls. And over the last six months we
have seen articles coming out from Anthropic as well on how to
build agents and what are different agentic patterns that
you could be leveraging. For example, a self reflecting
pattern where you generate the output, have another NLM,
reflect on that, give feedback and then generate again.
So ultimately these are all the smart ways of how we piece
(02:08):
together NLM calls so that we get to the desired output that
is an agent for me. And then if it can get to a
level that it can replace a process within a business or a
whole job function, then we start talking about these
digital employees that was also very famous and on the rise a
(02:30):
lot because ultimately an agent could get a whole job function
done. And to get to that point, what
type of safety considerations, reliability, even human in the
loop, what what's required to get it to the point where you
can say, hey, I have a digital employee now versus just a
script that someone will run andmonitor.
Great question. Where could we start?
(02:50):
There are a lot of considerations to be had.
The first consideration is around data security and safety.
So where does this agent live? What data does it interact with
the outputs of the agent? Are they customer facing and if
they are, then are they exposed to some regulatory requirements
(03:13):
or is just internal facing and does it have a significant
business impact if the output ofthe agent is right or wrong?
So you have all these considerations around
compliance, safety, data security.
Then the second question is, of course you want to define what
the agent does. And usually you tie back the
agent to a job function, a business process within your
(03:36):
company. So then you want to make sure
that you outline all the processvery, very clearly and you
understand that. So when you start building your
agent, then the agent has clear objectives and knows the systems
that it has to integrate with. So the second one is just
understanding the process. Afterwards, when you want to
(03:57):
implement an agent, the questionis what is its success criteria?
How do you know that it performs?
And it's the same question as you would be asking about an
employee. How do you know that your team
member performs? How do you do you know that they
are on track? And there it comes about setting
up an evaluation data set, whichcan be maybe as simple as an
input as an expected output, so that you can use this evaluation
(04:20):
data set while you are developing the agent and you can
benchmark yourself against it. And you can see the progress of
starting maybe at a 20% success rate in the beginning and
growing up towards 80% and onwards.
And once you have these mechanisms in place that it's
all about quick iteration, weakly iteration, building out
your agent, testing it against the evaluation set, and once you
(04:43):
feel confident, then putting it in real life.
And of course, businesses have different bars or like when the
agent is ready to go production,if the agent is exposed to your
end customers, has the possibility to damage your
brand, damage your customer relationship, then the bar
ultimate is higher. If it's an internal process,
(05:05):
then the bar often tends to be lower.
And when you want to launch an agent, I also, we, we have this
like framework of how do you getto a truly fully agenting
behaviour where you can just like fire and forget about it
and the work will be done. And that's not like an overnight
process. Usually the way you start is
first you build a copilot experience for your human team
(05:27):
members, which is basically augmenting their work and
helping them do the work better.And then you can see that this
augmentation, is it helpful for the day-to-day work or not?
Like do they approve the suggestion or not?
And this helps you to collect data.
So once you get to a confidence level where the human is
approving enough suggestions, then you get into a human in the
(05:49):
loop where you flip basically the seats.
So before the human was in control, the AI was the copilot.
Now the AI is in control, the human is the copilot or the
reviewer and but the gatekeeper.And then once you feel confident
enough that 99% of the cases areall passed and accepted by the
(06:10):
human, that 1% is successfully flagged for further review.
Then you get into the agentic mode, which is now the AI
operates stand alone. And then basically your company
can get into a management by exception set up where the AI is
doing its work. Whenever there is an exception
or no confidence score, it's surface back to the human team
(06:31):
and it's being reviewed. One thing I'd like to touch on
is I love the idea of progressing from copilot to
agent and just like that, that transition through there.
What do you recommend in terms of strategies or systems set up
so that the collection of the data to know when it performs
well or makes a mistake is stored in such a way that it's
(06:52):
actually useful? We're not just saying arbitrary
audit log of all the thumbs up, thumbs down, but actually saying
when this action occurred, this was incorrect, this was desired
behavior, and then somehow feed that back in the system.
Like do you have any strategies or approaches for that?
I think what you have described is the right mentality to have
in terms of how can you figure out a feedback loop to make the
(07:14):
system smarter over time. But if there is like A1 size
fits all approach, I don't thinkthere is because every single
use case has a different way to evaluate it and has a different
way of data structure and then adifferent way of data
collection. But we can like dig into one
example, which is the most common one around.
Everyone wants a better agent towrite emails, right?
(07:36):
For example, we have agents for writing sales emails.
We have agents like like companies like Fixer or managing
our inbox. And we want that AI agent to
have a better draft for us. Now, how this data loop can look
for you is that the AI agent writes the first draft and then
it's monitoring whether you haveaccepted the draft or you have
(07:56):
made changes. If you have accepted, that's
great. If you have made changes, then
the new changed and improved version is considered the
success version. And then what is possible behind
the scenes over time is that next time you encounter a new
e-mail that you have to draft, you can do a vector search among
(08:17):
all your historical emails that you have written and to see
which are the similar topics, conversation styles that they
meet, the meeting booking exchange.
Find some of the similar best e-mail topics and emails that
have been written. And then choose a couple of
examples and put it in as context into your LLM call.
(08:38):
So that basically you are telling the LLM in the past,
these were some examples which has been approved and which was
successful. So follow the similar pattern to
write the e-mail again. So this way the feedback loop
improves. And then also you can shorten
your search bit based on the time horizon.
Don't look back for three years because likely my writing style
(08:58):
have evolved 3 years ago compared to six months ago and
you can have all these mechanisms in place.
Nice. And I feel like there's even a
more simplistic version, maybe not for something as broad as
e-mail, but just storing the golden data set like an ideal
example and even just feeding that into the LLM call as here
is some examples of successful ones like Mimic.
(09:20):
Take his inspiration and go fromthere.
Absolutely, Absolutely. Do you, besides writing emails,
do you have any other common usecases where businesses should
start trying to create this golden data set?
And they can say, this is how the the best human at our
company does a certain task. Let's store this and then we can
feed into an AI system when it'swhen it's ready.
I would say every AI system should have a golden data set in
(09:44):
place. The reason is that building
systems with large language models is not a deterministic
process. The answer is stochastic is even
just like. Even just like.
One extra space in the prompt will yield to a slightly
different outcome than it was before.
So it's really critical to have an evaluation data set in place
(10:06):
because this then enables you asan engineering team, as a
product team, to iterate faster.If you have confidence in your
train, in your testing set, thenyou can quickly iterate, run it
against the benchmark and see ifit works.
I can give you some other examples which we have been
setting up in terms of evaluation.
So we have done a lot of work inlegal tech, for example,
(10:27):
building AI agents for the legalservice industry.
And there is extremely hard to set up evaluation sets
especially for the answer generation.
Like how do you judge this and if an answer is is right?
We have had more success around the document retrieval side of
setting up evaluation because weknow that for one question what
(10:48):
are the relevant documents that should be retrieved.
So at least we can benchmark that with a simple precision and
and recall the metric. Other examples we have been for
example parsing invoices and parsing documents for
accounting. Again, there is straightforward
how to set UA data set because you have line items that you
want to extract. Then you want to make sure that
(11:10):
all those line items are found after your LLM call or OCR call
has happened. And one thing I want people to
consider too is that you don't just need to build this golden
data set when you're building your own systems.
Like even if you buy an off the shelf solution, having these
ideal use cases can still be fedin.
Could we touch on that the buildversus buy debate?
What, what's your usual approachwhen you recommend people like,
is there everything custom or ordo you recommend certain tools
(11:33):
that people can just, you know, go online and pick up
themselves? A great question.
I definitely wouldn't recommend that build everything custom and
there are a couple of considerations to be made.
The first consideration is whether is there anyone who can
maintain a custom software in the future.
Do you have an engineering team?Are you planning to hire an
engineer, even just one engineer, but are you planning
(11:53):
to invest in that? That's the first question.
And if, if the answer is a category no, then you could just
look at buy the second part where you have the optionality
of like potentially having an engineering team in the future.
And let's say you, you want to buy a product for operational
efficiency. And what we have seen for legacy
businesses, they have some processes which are really
(12:16):
difficult to change. And in that case, often when you
buy a, an off the shelf solution, then the off the shelf
solution already has a premade concept in terms of how to use
it and the process. And then it often becomes really
hard to fit together the bot product and the company
processes. So in that scenario, also it's
worth considering them to build your solution, especially that
(12:39):
right now building software is becoming faster than ever, which
means if it's faster than it's also cheaper than ever.
And then there will be some, let's say, integrations around
the AI system, which will not bea core differentiation for you
in the future or will not be your mold.
And then sometimes it's just notworth to build out all that
(13:00):
complexity when it's not relatedto your core differentiation.
One example is voice AI agents. You can implement voice AI
agents in your business, but nowis it worth it to implement the
whole infrastructure end to end by yourself, meaning text to
speech, speech to text, turn taking detection, A
normalization background, sound removal and all that is it takes
(13:24):
years of effort and millions of dollars of R&D investment.
So there are already companies invested, 10s of millions even
and they only do this exceptionally well.
In that case, it just makes sense to buy it off the shelf
and then you could integrate this off the shelf product with
some other custom elements that you have built.
So these are roughly the the various scenarios that we
(13:46):
consider. Fundamentally it's, it's one
question is, is this core differentiator for you in the
future, is it worth investing money into that and all the IP
in house And the second one is have an engineering team, an R&D
team or you would rather just get something off the shelf.
Now, there is actually another aspect which is coming to play
(14:06):
now that I remember is that withsome software, especially in the
enterprise setting, it might sound like a great idea to buy a
software off the shelf. Let's say for a performance
review management software, you would buy Lattice and you will
have to pay 200,000 per year forlicenses due to the number of
(14:27):
employees you have in this scenario is also a question.
How about you just vibe called it internally, then you make it
production ready and then you build it once you invest once,
maybe 50K or 100K even. But then you own this forever.
And beyond that you can also tailor it and customize it to
your own processes. And also what we observe in this
(14:50):
situation that this that that companies most of the time they
only use a very small subset of the software, just a feature
which they over index on and allthe rest is ignored.
So why would you pay so much money for those licenses where
you could just extract this feature maybe in two weeks with
a small internal team and and and this is a new evolving model
(15:13):
that we see. And I'm glad you brought that up
because one thing that I find very exciting about this new
space is just with the importance of data.
The ability to vibe code up datapipelines and integrations has
really like lowered the bar where even if you get certain
types of software, one example, I use granola during meetings,
but then you can pipe it into Zapier and then that can go
(15:34):
everywhere. And you can put into my Obsidian
Vault. So I can take information and
then run cloud code from my Obsidian Vault, actually get
intelligent insights on my notes, make connections.
There's so much ways that you can amplify and enhance existing
software, which is like small little additions to to the
systems that we have available to us.
Also at Box 1 Ventures, I've been teaching the team how to
build up their own dashboard. So like we have an internal
(15:54):
system, but I've exposed internal API and now they don't
have to ask me, hey, can you build a dashboard for X?
They can build themselves interface with the real data.
It's so exciting. Do you see any other trends with
how small non-technical teams can start leveraging AI to get
more customized in their workflows?
What you have outlined actually sounds like the best practice
that I have seen as well. So the way we think about this
(16:17):
is that within the company, if employees all get access to
ChatGPT to Claude's desktop, whatever it would be, and they
only have access to that, that almost like a single player mode
way of working with AI. But our work is often
multiplayer and multiplayer could mean interacting with
various people, interacting withvarious tools and softwares and
(16:38):
pulling data together and then running the insight.
So the system that you have set up at Vox one is is incredible
and, and it really aligns with with what we see as well on one
end for the multiplayer modes, you need an orchestration layer.
You mentioned Zapier, we are a big fan of NA10.
We actually just became a, an accredited partner of NA10 as
well. So we recommend to set up NA10,
(16:59):
give your team access, let them play around with the workflows.
And the other one is also, we have seen the AI app, be it your
10 GPT desktop, your cloud desktop almost becoming like a
control panel because you can connect through MCP, various
external applications like you are connecting your Obsidian
(17:20):
probably via an MCP and then youcan quickly query it and then do
more things on top of the data. These are the two patterns we
have been seeing. Nice.
And just to double click on MCP,do you see any risks or
considerations people should take into account before they
spin up their own, where they might want to expose their own
internal data via MCP server? So for example, you don't have
(17:42):
to write an API client to interface with the database, you
can just have your MCP so you can chat with Claude or our
ChatGPT. Do you have any best practices
there? Probably not yet.
Best practices. We haven't built enough MCP
servers to get to a best practice level where we can
define this. But what we have seen is that
the Entropic team is already making improvements to the MCP
(18:03):
library in the sense of improvedsecurity.
You have authentication as well that you can have at hand and
also improving the way that you can roll out MCP server updates
if you have it across the organization and you have
multiple cloud desktop users. That's one.
And the other one is we also seesome start-ups which are rising
(18:24):
to fame, both backed by YC and also some in Europe which give
you this like secure MCP container only once you just
like push your code and you can also host the MCP server locally
and make sure that it only workswithin your VPC.
These are some of the best practices that we have been
seeing. When you use MCP in your
(18:44):
business processes, you got to give it access to real data and
systems and that can be scary and that's why I recommend
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 install it in a client in
seconds and secret protection and network isolation are built
in. You can try Toolhive as well.
It's free and it's open source. You can learn more at Toolhive
(19:06):
dot dev and they're back to the conversation, which I'll learn.
So on top of MCP exposing the data, one thing I think people
need to pay attention to is how they actually store their data
because not everything is structured.
Sometimes things are just in a in a Word document or in APDF.
Do you have any recommendations for how companies, whether like
brand new startups or more established enterprises, can
(19:27):
start thinking about how they store and structure their data
so that it's optimized as they start implementing more AI
systems? Yeah, great question.
And there have been also some statistics around, I don't
remember the exact number, but just a sheer amount of data
sitting in unstructured format within enterprises is gigantic
compared to the structured data.Ultimately it's not very
(19:50):
different from conducting a simple data transformation work
where you you have your data source, you define the desired
format that you want to date your data source to be in, and
then you do the transformation in between.
The only thing that changed is that previous to L and Ms. We
couldn't do a transformation of unstructured data reliably,
(20:11):
right? Right now we can.
We can process, call transcripts, word documents.
So the question is define your problem.
For your problem, what is the data that you need?
And then look for where can you get the data from and build out
the transformation in between. Now for this transformation,
anonyms can be very powerful. And for example, you could take
(20:33):
APDF or a Word document, just get the text, pass it into an
LLM and then parse it if your PDF is scanned.
We have seen that Mistral OCR seems to be the best performing
OCR model currently. So you can pass your PDF through
this. Or you could also pass each page
of the PDF as an image into an LLM like maybe Google Gemini 2.5
(20:56):
Flash. You can also process Excel files
with Excel files. Still a bit tricky with LLMS.
Again, what you can do, take a screenshot of the one tab in the
excel, send it as an image. NLM is pretty good at describing
to the content or getting back to you with insights and then
based on this you transform the data which is in a relational
database and then the rest is all the same as so far.
(21:18):
Nice. I'd like to pivot slightly
because one blind spot I have isI've been in tech for a decade.
A lot of the companies we talkedto are tech.
But there's so many non-technical companies out
there or just like slightly technical that I worry the gap
between productivity of us technical companies that are
leveraging AI for as many systems as possible and the non
tech who just aren't even aware of what's available is just
(21:41):
going to continue to expand. And we need to make sure
everyone kind of benefits from this technology.
So I believe you've recently hadsome experience deploying AI
agents in the fire safety vertical.
Would you mind talking about some of the experience there?
How the process went? Things to look out for?
Just what helpful tips could someone who's not a techie
person or not in a techie industry get to benefit from AI?
Absolutely, absolutely. I would say indeed companies who
(22:05):
are more technical have an edge,but it's easy to close that gap
for non-technical teams as well.The simplest way how you can get
started is just to start using AI and start using the tools
that you have out there. Start using ChatGPT on a daily
basis and they're the most common blocker that we see is
(22:26):
that Despite that, people have access to change GPT, let's say
they don't know how to use it. And then it leads to
frustration. And we hear the common complaint
of like, until I could have figured this out, I would have
done it by 10 times already. And I would typically agree,
yes, you could have. But have you thought about all
the other thousand times that you have to do the same thing in
(22:47):
the future? Have you accounted for that?
Usually the answer is is no. So the first step where a
company should start is just give change PT to your users or
cloud desktop, whichever it would be.
And either offer them a way to like AI training session, a
prompt engineering training session and like get everyone to
a base level to feel comfortablewith this tool.
(23:10):
I would say this is where where you start.
And this also ties into a framework which we see for
companies who want to adopt AI that it's not not all about like
technology only as we have been talking so far, like we have
been super technical, but it's also about people and the
process transformation. And this is where we believe
(23:31):
that the company really succeedsin becoming AI native or AI
driven as an organization when they pay equal attention to all
the three buckets of people, process and product.
So with people, it's simple. You just give them the tools and
let them to learn, learn it, to use it.
And also one thing which we not account for enough is give them
(23:56):
psychological safety to have themotivation and the will to try
the AI tools and to use it. Because if the attitude towards
the AI system is at all, this will come and take my job.
Well, I better resist as much asI can.
That will not lead very far downthe line.
So this has to come from the leadership team to set the
(24:18):
psychological safety within the company of like, hey guys, we
are not going to let you go. This is not the reason why we
want you to use AI. Our ambition for example, maybe
is to grow the company and we want to grow, but we do not want
to grow the team as well. We just want to grow the
business. Or it's like, guys, we saw that
you are burnt out. We just want to somehow find a
way to take away the boring, thethe busy work from your
(24:41):
day-to-day so that you can really do your happiest work in
this company. So that's why we need AI which
will take away the boring work so you can be happy and that we
can continue building a successful business.
And then you have the other side, which is the process where
it's about like you identify theprocess where you want to
implement AI. Of course, you have to have very
(25:03):
clear understanding of that process, map it out in a process
map. Once you have automated it, it,
it's not enough to just automateit and tell everybody, hey, this
is ready. This is the new process.
It requires constant reinforcement or we call it like
change management where you worktogether with the teams to train
them on the new process. Offer weekly office hours where
(25:25):
they can come and they can ask questions.
Offer them an A direct and an easy gateway where they can
report if things are not workingor they are not happy with
certain parts of the process or maybe they have ideas for
improvement. Because ultimately these people
will be your end users for the deployed AI system and after all
of these type technologies thereas underpinning which just
(25:48):
empowers it all. But it's not the top of mind
that you should be talking aboutwith your people everyday.
You should only be talking aboutthe people and the process.
Technology will empower it all. Yeah, that's a wonderful
perspective. So many people need to remember
that change is scary for many people and especially with the
doom and gloom you hear on the news as to AI is going to wipe
(26:08):
out jobs and display so many people.
The the thought of it coming in being like, hey, how do you do
your job? We're going to have AI work
alongside you is definitely intimidating.
But if you can reassure them that this is going to take away
the drudger of your job, allow you to be more creative, allow
you to focus on other things. You can grow, expand, learn into
a new role, or even just become much more efficient at your
current role where instead of being able to deal with X number
(26:30):
of tasks, you can deal with double the amount of tasks
without getting the extra burnout.
To to build on that one thing that I found really interesting,
just use my dad as an example. Told him a ChatGPT way back in
the day, the rush should often really think about it.
But then he introduced him to Suno, the music generator.
He took all of his old songs, made them into like new songs.
It was really cool. And then he introduced him to
perplexity as just like the nextstep being like, instead of
(26:51):
using Google just to try perplexity, now he's just
chatting to me on the regular because it was an incremental
progress of like play and enjoyment rather than just hey,
this will make you more productive.
Have at it it. It's a lot less appealing that
way. So I encourage people as I've
done a bunch of other episodes, as mentioned, a bunch of other
episodes just play like, explore, like there's so many
cool things out there that if you can just kind of get past
(27:12):
that fear of the uncertain that we're, you know, we're entering
a new age when all of a sudden, like we put electricity through
sand and then we get intelligence out of it.
It's pretty miraculous. It's gonna be a net positive.
It's just there are growing pains and we just have to try to
work on change management so that we bring everyone along for
the ride. Absolutely, Absolutely.
I would love to go into process a little bit more cuz that's one
(27:34):
area that really interests me. Do you have any strategies for
the initial collection? Like the documentation of a
process, should someone just like stand over someone's
shoulder and write down the workthat they do?
Should we try to implant some type of tracking software or
what is a low friction way that we can start documenting these
processes? So we have step 0 for process
(27:54):
automation. The simplest way how we start is
just interviewing the team members or watching over their
shoulder and spending a day or two with them and seeing how
they work is as simple as that. We have been doing process
mapping and like consulting has been around for many, many 10s
of years, right? So they have been doing process
mapping, looking at the process,reworking it.
(28:15):
This time is, is no different atall.
The only critical part is just to have full clarity and really
deeply understand the process. Many people that we see just
like quickly jump to a solution like, oh, we understand the high
process, let's now build AI agents and let's now build the
tools and you are missing the new ones and the detail.
(28:35):
And now, for example, with external parties, like external
parties like ourselves, the client really feels unhurt.
If you just come to them like they speak for 5 minutes and
you're like, Oh yeah, got it. That's now build AI.
So it always with also the part of lecture, understanding the
process, but also showing the client that you really get their
business and their context and their domain.
(28:57):
And also do not think that if you automated the process once
is the same across every company.
It's actually different for every single company because
there are cultural values, cultural analysis, ways that
people work differently, and allthis has to be taken into
account. Yeah, and that kind of rings
true with I guess not that recently more, but there's an
article about how 90% of AI projects failed.
(29:19):
And what's really unclear is were these people just being
like, hey, let's try this one product and just like through it
without really auditing if it was the right tool for the job.
What do you recommend for peoplefor selecting the right tool for
the job when there are 1,000,000out there?
They seem to really overlap, like distract.
You do everything like. How do people figure out what
tool they actually need? I think the article that you
(29:39):
have mentioned is actually the reason why those failed is
because they weren't clear around what they are automating.
I have a great friend in in San Francisco and he told me was
that in San Francisco the thinking is that technology is
never a problem. The only problem is your
business model. And now if we translate this to
the process, technology is nevera problem.
(30:02):
The only problem is you are unclear about what process you
are alternating and what is the ROI that you are that you want
to achieve and and you want to drive.
So I would say that 95% of thosepilots, they were all unclear
and they were just building sexyAI agents just for the sake of
technology. But actually they had 0 clarity
around what is the business value that will be delivered.
(30:24):
So then ultimately when the POC phase is done, they will
realize, oh, there is no business case.
So that's just shelf. What we do is we have often told
our clients that I'm sorry, we just cannot work on this product
because there is no ROI. We just don't understand what is
the ROI and usually ties back toimproving your margins, driving
(30:46):
top line somehow back to the EBITDA and PNL impact or in the
overall EV of the the the enterprise value of the company.
And if we cannot see that we don't even get started on that
work because it's a failed project and the failed
initiative from the get go. By the time you could have the
(31:09):
sexiest implementation, but there is no business case then
the client will be unhappy. If the client is unhappy, we
will also I'll be happy because it means we didn't do our job
well and they consider that we did a poor job.
Yeah. So that's what I would
recommend. It's just really be critical
around what is the ROI, drive itback to the dollar value.
(31:30):
And if the dollar value outweighs of the ROI outweighs
the investment, just go for it. That is a no brain.
Do you have any advice or techniques around measuring ROI
for AI? Because I'm just thinking, if we
give everyone a ChatGPT subscription, how can we?
And I know there's, there's sucha wide range.
So I'm thinking more generally, but how can we make sure that
the, whatever, 30 bucks a month that you commit to a
(31:52):
subscription pays for itself? Do we try to say employee
satisfaction output? Are there other things we should
try to take into account? Because you mentioned burnout
earlier, If we can have people deliver the same amount of
output, but there's less burnout, that seems something
cool. But how?
How do we measure that? Great question.
I think justifying ROI for just the ChatGPT subscription, it's
(32:14):
probably a bit harder and not sotangible.
Maybe you can measure it back tothe employee satisfaction for
example. But as I said, that's a single
player mode and so with ChatGPT access, you are not yet playing
the game of becoming an AI native company.
(32:35):
You are just doing the table stakes, which is absolutely
mandatory for everyone to do. It's just the foundations.
So you are not better or more successful just because you are
judging the subscription. This is not table stakes.
I think the unlock comes afterwards where you get into
the multiplayer processes and the ways to all to really
(32:58):
automate the process end to end.That's when you can dig into
Rois for example, just clocking of how long does it take to
complete this piece of work. Let's say if you are in the
legal industry and you are doinglegal research, you can clock
that same research takes 2 hoursfor you.
Now the challenge is that you will do the ROI based on the
(33:19):
hourly rate of the of the person.
And then maybe you want to get it down to 30 minutes or to 10
minutes driven by AI and then scale it across all the
employees and like how many hours the organization is
spending on that. And then you can start waiting
ROI versus the investment that is required.
So this is how I would tie it back.
(33:39):
Yeah, that's, that's mostly whatwe.
See what about for you personally?
What? What AI tools do you leverage or
have you implemented to help make yourself more productive as
you go about your day? That changes constantly because
this AI tools just get better and better every single week.
So my go to LLMS at the moment are Grok.
(34:00):
I'm testing clods about with clod, mainly clod code and open
AI ChatGPT. These are my go to LLMS.
Then in terms of orchestration, I have a personal NA10 instance,
so I can automate as much as possible there.
Whenever there is a repetitive flow that I find that I can
automate that with NA10 and I have been just testing actually.
(34:24):
There is another one, Granola. I'm a big fan of granola, using
it end to end. They just released a super
exciting feature which is aroundchatting with all your notes and
then getting weekly summaries. So I think the team is just
doing an absolutely amazing workin improving their product.
I'm a big no trend user now theyhave no trend AI so that also
(34:44):
makes life easier and I think that is retty much all I use and
Whiser actually yeah, I use Whiser for dictation sometimes
just writing the long prompts. It just takes ages and more hay
to just speak it out loud. Yeah, as someone who works from
home, I I talk to my computer far more than I type.
(35:05):
I yeah, especially now that I'veheavily integrated cursor and
cloud code into my workflow, I will just tell the prompt, read
the output, make a couple littlechanges, get back to prompt with
my voice again, and it's just the bandwidth difference is
phenomenal. You are not the first person to
bring up NA10, and as someone who's been developed for so
long, I have an aversion to it just because I want to write my
own stuff. But I've heard a lot of good
(35:25):
things. Could you talk about some of
your use case for NA10 and when it is a benefit to that instead
trying to write a Python script?Yeah, of course.
Well, you can write a Python script and have it in NA 10 as a
note. So you can.
That's the beauty of NA 10. So I think Zapier, I find Zapier
to be very just end user, user experience focused and not very
(35:46):
developer friendly or low level.Why with NA10 you could go into
the low levels as an engineer ifyou want to write custom
scripts. So that's an advantage.
You can also build your own custom nodes in NA10.
I think a big one for companies and especially larger companies
and enterprises is self hosting.So you can host it on your old
servers. Make sure that the data never
(36:07):
leaves. Then there is an advantage of
NA10, and this is similar to Zapier and all the rest of the
tools that they just have like thousands of integrations
out-of-the-box. So why bother with the Gmail
integration and Outlook integration, A Google Drive
integration? I wouldn't want to write that
code that's from scratch. I'm more than happy for NA10 to
handle it. So I think these are some of the
(36:29):
reasons. And also we are very excited
about the impact of AI in the professional services industry.
And the professional services companies just don't have the
engineering bandwidth or the engineering teams to maintain
code bases. So anytime for them it's a great
fit because they can automate the work flows maintained with a
(36:49):
very low engineering effort and still they get the benefit of a
fully built. Software.
This was a lot of fun. Before we let you go, is there
anything you want the audience to?
Know I'm curious from your end what AI tools you are using
beside your measured cursor cloud codes you are talking
about with your computer, What else?
Yeah, so I have a cloud desktop running for my main chat
(37:11):
interface. I still use open AI from time to
time, but more so for deep research only.
I actually do perplexity for guest research.
So I'll do a quick tell me aboutthe certain guest.
I use Aqua Voice as my transcriber.
The I used to use Super Whisper,but I migrate to Aqua Voice when
it allows you to have the optionof doing a like a screen input
at the same time so it can actually get context from your
(37:33):
screen. So if you're talking to someone
with an unusual name or you're doing code, it'll actually do
the function name the way it's written in your code rather than
just verbatim what you say. I use Obsidian for my notes.
I'm a big fan of the file over app principle.
And because it's just a collection of markdown files, I
can have cloud code running in that directory to create custom
dashboards for me, as well as things like the very robust plug
(37:55):
in ecosystem. So I can do a data view plug in
and actually have code operate across my notes where I can say
like get all of the unmarked to do's that are due in the next
week and put them in a list. That's been a lot of fun.
Cursor is my main interface for coding.
I've tried the different tools Ijust found.
I keep kind of migrating back tocursor.
It's not perfect but it's it's good enough for me.
(38:17):
I was big into Neo vim for a while, but now that I just typed
less because I'm speaking so much, I still use my my vim
bindings but it's just less of aa requirement for me to have
that like ultra efficient control over my computer.
Around the podcast. I use descript for editing.
I use Canva for the thumbnails. I feel like that's it.
(38:38):
I'll let you know if there's more.
But yeah, it's I tried to audit every tool that I'm using and
seeing if there's an AI alternative, test them out, some
of them. It's AI for AI's sake.
It doesn't add any actual value,just increase the price of it.
But some of them you actually get a minor bit of efficiency
here. And like I was emphasizing
earlier, the ability to pipe data from one system into a core
repository I think is under indexed in this day and age.
(39:01):
And I think creating your secondbrain, having well structured,
well formatted data that you canfeed into an AI system is going
to pay massive dividends sooner than most people realize.
So I really hope more people getinto the, the data hoarding,
data harvesting stage and just start saying, I, I write emails
on the regular. How can I start storing the gems
or, or you put out a bang or tweet.
How can you start building your data set of things that go viral
(39:23):
for the model that will help yougo viral more often?
I like that. I like that storing and building
up your own data set that's that's a really good one.
And actually, and one other thing too, like I, for example,
have a moon Lander keyboard thatI've cooked to connect a hammer
spoon to control my Mac. I don't really know Lua.
It's a very easy language, but just one other thing to learn.
But just by having a chat with Claude, I can be like, I want to
(39:45):
accomplish this. I have Hammerspoon.
I have like this config for my moon Lander.
And then it just like outputs the Lua I need.
And all of a sudden I have more efficiencies for my keyboard,
which aren't technically an AI tool, but it was AI derived.
Super cool. So char, where can people keep
up with you? Tell me what Bonsai labs?
Yeah, absolutely. So you can find me on LinkedIn
and similarly is Bonsai Dash, la-bs.com is where you can find
(40:10):
our company. And as as you mentioned in the
beginning, we work with private equity firms, work with some of
the top 10 firms in the world, but also with smaller and medium
ones. And we are the nimble and fast
moving AI value creation partnerfor this firms where we go
through the portfolio companies and we just help them get to AI
(40:33):
value as quickly as possible. So you can find me on LinkedIn
to reach out anytime. Thank you for listening to my
conversation with Chong or I hada blast talking with him.
He clearly knows his stuff and has done this a lot of times to
figure out the best ways to implement AI to solve these
operational difficulties. But I'm curious, from your end,
what processes are you struggling to automate?
Have you tried the tools you mentioned?
(40:53):
Are there any others that you like?
I'd love to hear from you. And I'd also love it if you like
and subscribe this video so we can get more of these
conversations going, so we can actually start relaying more
value to people so they can learn how to use AI to become
more productive and efficient. I'd like to give a quick shout
out to Tool Hive for supporting the show so I can have
conversations like this and I'llsee you next week.
Can you figure out a feedback loop to make the system smarter
over time? Maybe you want to get it down to
30 minutes or 10 minutes driven by AI and then scale it across
all the employees. Then you get into the agentic
mode, which is now the AI operates stand alone and then
basically your company can get into a management by exception.
(00:20):
So that's why we need AI which will take away the boring work,
so you can be happy and that we can continue building a
successful business. There is so much AI hype out
there when business leaders wantto know how do they cut through
the noise to find the true efficiency gains?
Do they build? Do they buy?
How do they know what to choose?How do you even know where to
apply AI in the 1st place? Well, in episode 61 of Tool Use
(00:40):
brought by Tool Hive, we're solving all of this and more.
We're joined by Chongo Barabasi,the founder and CEO of Bonsai
Labs. They work with private equity
firms as the AI value creation partner, deploying AI across all
their portfolio companies. So he has a lot of experience on
how to actually use AI for operational efficiency.
So please do this conversation with Chongo I.
Feel that agents has being the talk almost from day one of
(01:03):
ChatGPT everyone was talking about agents for me at some
point. It also felt like this is an
overhyped work, that it's like abuzzword that everyone is using.
Because ultimately if you boil down what an agent is, is just a
sequence of LLM calls. There is nothing very complex or
rocket science about it. Right?
(01:24):
You could be an agent by just copy pasting a prompt, getting
an output, putting it into the next prompt, doing that again.
But nevertheless, even this, this orchestration you could say
of NLM that led to an agentic behaviour.
And what I realized over time isthat really when we call them
agents, it's just a very creative way of how you piece
(01:45):
together NLM calls. And over the last six months we
have seen articles coming out from Anthropic as well on how to
build agents and what are different agentic patterns that
you could be leveraging. For example, a self reflecting
pattern where you generate the output, have another NLM,
reflect on that, give feedback and then generate again.
So ultimately these are all the smart ways of how we piece
(02:08):
together NLM calls so that we get to the desired output that
is an agent for me. And then if it can get to a
level that it can replace a process within a business or a
whole job function, then we start talking about these
digital employees that was also very famous and on the rise a
(02:30):
lot because ultimately an agent could get a whole job function
done. And to get to that point, what
type of safety considerations, reliability, even human in the
loop, what what's required to get it to the point where you
can say, hey, I have a digital employee now versus just a
script that someone will run andmonitor.
Great question. Where could we start?
(02:50):
There are a lot of considerations to be had.
The first consideration is around data security and safety.
So where does this agent live? What data does it interact with
the outputs of the agent? Are they customer facing and if
they are, then are they exposed to some regulatory requirements
(03:13):
or is just internal facing and does it have a significant
business impact if the output ofthe agent is right or wrong?
So you have all these considerations around
compliance, safety, data security.
Then the second question is, of course you want to define what
the agent does. And usually you tie back the
agent to a job function, a business process within your
(03:36):
company. So then you want to make sure
that you outline all the processvery, very clearly and you
understand that. So when you start building your
agent, then the agent has clear objectives and knows the systems
that it has to integrate with. So the second one is just
understanding the process. Afterwards, when you want to
(03:57):
implement an agent, the questionis what is its success criteria?
How do you know that it performs?
And it's the same question as you would be asking about an
employee. How do you know that your team
member performs? How do you do you know that they
are on track? And there it comes about setting
up an evaluation data set, whichcan be maybe as simple as an
input as an expected output, so that you can use this evaluation
(04:20):
data set while you are developing the agent and you can
benchmark yourself against it. And you can see the progress of
starting maybe at a 20% success rate in the beginning and
growing up towards 80% and onwards.
And once you have these mechanisms in place that it's
all about quick iteration, weakly iteration, building out
your agent, testing it against the evaluation set, and once you
(04:43):
feel confident, then putting it in real life.
And of course, businesses have different bars or like when the
agent is ready to go production,if the agent is exposed to your
end customers, has the possibility to damage your
brand, damage your customer relationship, then the bar
ultimate is higher. If it's an internal process,
(05:05):
then the bar often tends to be lower.
And when you want to launch an agent, I also, we, we have this
like framework of how do you getto a truly fully agenting
behaviour where you can just like fire and forget about it
and the work will be done. And that's not like an overnight
process. Usually the way you start is
first you build a copilot experience for your human team
(05:27):
members, which is basically augmenting their work and
helping them do the work better.And then you can see that this
augmentation, is it helpful for the day-to-day work or not?
Like do they approve the suggestion or not?
And this helps you to collect data.
So once you get to a confidence level where the human is
approving enough suggestions, then you get into a human in the
(05:49):
loop where you flip basically the seats.
So before the human was in control, the AI was the copilot.
Now the AI is in control, the human is the copilot or the
reviewer and but the gatekeeper.And then once you feel confident
enough that 99% of the cases areall passed and accepted by the
(06:10):
human, that 1% is successfully flagged for further review.
Then you get into the agentic mode, which is now the AI
operates stand alone. And then basically your company
can get into a management by exception set up where the AI is
doing its work. Whenever there is an exception
or no confidence score, it's surface back to the human team
(06:31):
and it's being reviewed. One thing I'd like to touch on
is I love the idea of progressing from copilot to
agent and just like that, that transition through there.
What do you recommend in terms of strategies or systems set up
so that the collection of the data to know when it performs
well or makes a mistake is stored in such a way that it's
(06:52):
actually useful? We're not just saying arbitrary
audit log of all the thumbs up, thumbs down, but actually saying
when this action occurred, this was incorrect, this was desired
behavior, and then somehow feed that back in the system.
Like do you have any strategies or approaches for that?
I think what you have described is the right mentality to have
in terms of how can you figure out a feedback loop to make the
(07:14):
system smarter over time. But if there is like A1 size
fits all approach, I don't thinkthere is because every single
use case has a different way to evaluate it and has a different
way of data structure and then adifferent way of data
collection. But we can like dig into one
example, which is the most common one around.
Everyone wants a better agent towrite emails, right?
(07:36):
For example, we have agents for writing sales emails.
We have agents like like companies like Fixer or managing
our inbox. And we want that AI agent to
have a better draft for us. Now, how this data loop can look
for you is that the AI agent writes the first draft and then
it's monitoring whether you haveaccepted the draft or you have
(07:56):
made changes. If you have accepted, that's
great. If you have made changes, then
the new changed and improved version is considered the
success version. And then what is possible behind
the scenes over time is that next time you encounter a new
e-mail that you have to draft, you can do a vector search among
(08:17):
all your historical emails that you have written and to see
which are the similar topics, conversation styles that they
meet, the meeting booking exchange.
Find some of the similar best e-mail topics and emails that
have been written. And then choose a couple of
examples and put it in as context into your LLM call.
(08:38):
So that basically you are telling the LLM in the past,
these were some examples which has been approved and which was
successful. So follow the similar pattern to
write the e-mail again. So this way the feedback loop
improves. And then also you can shorten
your search bit based on the time horizon.
Don't look back for three years because likely my writing style
(08:58):
have evolved 3 years ago compared to six months ago and
you can have all these mechanisms in place.
Nice. And I feel like there's even a
more simplistic version, maybe not for something as broad as
e-mail, but just storing the golden data set like an ideal
example and even just feeding that into the LLM call as here
is some examples of successful ones like Mimic.
(09:20):
Take his inspiration and go fromthere.
Absolutely, Absolutely. Do you, besides writing emails,
do you have any other common usecases where businesses should
start trying to create this golden data set?
And they can say, this is how the the best human at our
company does a certain task. Let's store this and then we can
feed into an AI system when it'swhen it's ready.
I would say every AI system should have a golden data set in
(09:44):
place. The reason is that building
systems with large language models is not a deterministic
process. The answer is stochastic is even
just like. Even just like.
One extra space in the prompt will yield to a slightly
different outcome than it was before.
So it's really critical to have an evaluation data set in place
(10:06):
because this then enables you asan engineering team, as a
product team, to iterate faster.If you have confidence in your
train, in your testing set, thenyou can quickly iterate, run it
against the benchmark and see ifit works.
I can give you some other examples which we have been
setting up in terms of evaluation.
So we have done a lot of work inlegal tech, for example,
(10:27):
building AI agents for the legalservice industry.
And there is extremely hard to set up evaluation sets
especially for the answer generation.
Like how do you judge this and if an answer is is right?
We have had more success around the document retrieval side of
setting up evaluation because weknow that for one question what
(10:48):
are the relevant documents that should be retrieved.
So at least we can benchmark that with a simple precision and
and recall the metric. Other examples we have been for
example parsing invoices and parsing documents for
accounting. Again, there is straightforward
how to set UA data set because you have line items that you
want to extract. Then you want to make sure that
(11:10):
all those line items are found after your LLM call or OCR call
has happened. And one thing I want people to
consider too is that you don't just need to build this golden
data set when you're building your own systems.
Like even if you buy an off the shelf solution, having these
ideal use cases can still be fedin.
Could we touch on that the buildversus buy debate?
What, what's your usual approachwhen you recommend people like,
is there everything custom or ordo you recommend certain tools
(11:33):
that people can just, you know, go online and pick up
themselves? A great question.
I definitely wouldn't recommend that build everything custom and
there are a couple of considerations to be made.
The first consideration is whether is there anyone who can
maintain a custom software in the future.
Do you have an engineering team?Are you planning to hire an
engineer, even just one engineer, but are you planning
(11:53):
to invest in that? That's the first question.
And if, if the answer is a category no, then you could just
look at buy the second part where you have the optionality
of like potentially having an engineering team in the future.
And let's say you, you want to buy a product for operational
efficiency. And what we have seen for legacy
businesses, they have some processes which are really
(12:16):
difficult to change. And in that case, often when you
buy a, an off the shelf solution, then the off the shelf
solution already has a premade concept in terms of how to use
it and the process. And then it often becomes really
hard to fit together the bot product and the company
processes. So in that scenario, also it's
worth considering them to build your solution, especially that
(12:39):
right now building software is becoming faster than ever, which
means if it's faster than it's also cheaper than ever.
And then there will be some, let's say, integrations around
the AI system, which will not bea core differentiation for you
in the future or will not be your mold.
And then sometimes it's just notworth to build out all that
(13:00):
complexity when it's not relatedto your core differentiation.
One example is voice AI agents. You can implement voice AI
agents in your business, but nowis it worth it to implement the
whole infrastructure end to end by yourself, meaning text to
speech, speech to text, turn taking detection, A
normalization background, sound removal and all that is it takes
(13:24):
years of effort and millions of dollars of R&D investment.
So there are already companies invested, 10s of millions even
and they only do this exceptionally well.
In that case, it just makes sense to buy it off the shelf
and then you could integrate this off the shelf product with
some other custom elements that you have built.
So these are roughly the the various scenarios that we
(13:46):
consider. Fundamentally it's, it's one
question is, is this core differentiator for you in the
future, is it worth investing money into that and all the IP
in house And the second one is have an engineering team, an R&D
team or you would rather just get something off the shelf.
Now, there is actually another aspect which is coming to play
(14:06):
now that I remember is that withsome software, especially in the
enterprise setting, it might sound like a great idea to buy a
software off the shelf. Let's say for a performance
review management software, you would buy Lattice and you will
have to pay 200,000 per year forlicenses due to the number of
(14:27):
employees you have in this scenario is also a question.
How about you just vibe called it internally, then you make it
production ready and then you build it once you invest once,
maybe 50K or 100K even. But then you own this forever.
And beyond that you can also tailor it and customize it to
your own processes. And also what we observe in this
(14:50):
situation that this that that companies most of the time they
only use a very small subset of the software, just a feature
which they over index on and allthe rest is ignored.
So why would you pay so much money for those licenses where
you could just extract this feature maybe in two weeks with
a small internal team and and and this is a new evolving model
(15:13):
that we see. And I'm glad you brought that up
because one thing that I find very exciting about this new
space is just with the importance of data.
The ability to vibe code up datapipelines and integrations has
really like lowered the bar where even if you get certain
types of software, one example, I use granola during meetings,
but then you can pipe it into Zapier and then that can go
(15:34):
everywhere. And you can put into my Obsidian
Vault. So I can take information and
then run cloud code from my Obsidian Vault, actually get
intelligent insights on my notes, make connections.
There's so much ways that you can amplify and enhance existing
software, which is like small little additions to to the
systems that we have available to us.
Also at Box 1 Ventures, I've been teaching the team how to
build up their own dashboard. So like we have an internal
(15:54):
system, but I've exposed internal API and now they don't
have to ask me, hey, can you build a dashboard for X?
They can build themselves interface with the real data.
It's so exciting. Do you see any other trends with
how small non-technical teams can start leveraging AI to get
more customized in their workflows?
What you have outlined actually sounds like the best practice
that I have seen as well. So the way we think about this
(16:17):
is that within the company, if employees all get access to
ChatGPT to Claude's desktop, whatever it would be, and they
only have access to that, that almost like a single player mode
way of working with AI. But our work is often
multiplayer and multiplayer could mean interacting with
various people, interacting withvarious tools and softwares and
(16:38):
pulling data together and then running the insight.
So the system that you have set up at Vox one is is incredible
and, and it really aligns with with what we see as well on one
end for the multiplayer modes, you need an orchestration layer.
You mentioned Zapier, we are a big fan of NA10.
We actually just became a, an accredited partner of NA10 as
well. So we recommend to set up NA10,
(16:59):
give your team access, let them play around with the workflows.
And the other one is also, we have seen the AI app, be it your
10 GPT desktop, your cloud desktop almost becoming like a
control panel because you can connect through MCP, various
external applications like you are connecting your Obsidian
(17:20):
probably via an MCP and then youcan quickly query it and then do
more things on top of the data. These are the two patterns we
have been seeing. Nice.
And just to double click on MCP,do you see any risks or
considerations people should take into account before they
spin up their own, where they might want to expose their own
internal data via MCP server? So for example, you don't have
(17:42):
to write an API client to interface with the database, you
can just have your MCP so you can chat with Claude or our
ChatGPT. Do you have any best practices
there? Probably not yet.
Best practices. We haven't built enough MCP
servers to get to a best practice level where we can
define this. But what we have seen is that
the Entropic team is already making improvements to the MCP
(18:03):
library in the sense of improvedsecurity.
You have authentication as well that you can have at hand and
also improving the way that you can roll out MCP server updates
if you have it across the organization and you have
multiple cloud desktop users. That's one.
And the other one is we also seesome start-ups which are rising
(18:24):
to fame, both backed by YC and also some in Europe which give
you this like secure MCP container only once you just
like push your code and you can also host the MCP server locally
and make sure that it only workswithin your VPC.
These are some of the best practices that we have been
seeing. When you use MCP in your
(18:44):
business processes, you got to give it access to real data and
systems and that can be scary and that's why I recommend
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 install it in a client in
seconds and secret protection and network isolation are built
in. You can try Toolhive as well.
It's free and it's open source. You can learn more at Toolhive
(19:06):
dot dev and they're back to the conversation, which I'll learn.
So on top of MCP exposing the data, one thing I think people
need to pay attention to is how they actually store their data
because not everything is structured.
Sometimes things are just in a in a Word document or in APDF.
Do you have any recommendations for how companies, whether like
brand new startups or more established enterprises, can
(19:27):
start thinking about how they store and structure their data
so that it's optimized as they start implementing more AI
systems? Yeah, great question.
And there have been also some statistics around, I don't
remember the exact number, but just a sheer amount of data
sitting in unstructured format within enterprises is gigantic
compared to the structured data.Ultimately it's not very
(19:50):
different from conducting a simple data transformation work
where you you have your data source, you define the desired
format that you want to date your data source to be in, and
then you do the transformation in between.
The only thing that changed is that previous to L and Ms. We
couldn't do a transformation of unstructured data reliably,
(20:11):
right? Right now we can.
We can process, call transcripts, word documents.
So the question is define your problem.
For your problem, what is the data that you need?
And then look for where can you get the data from and build out
the transformation in between. Now for this transformation,
anonyms can be very powerful. And for example, you could take
(20:33):
APDF or a Word document, just get the text, pass it into an
LLM and then parse it if your PDF is scanned.
We have seen that Mistral OCR seems to be the best performing
OCR model currently. So you can pass your PDF through
this. Or you could also pass each page
of the PDF as an image into an LLM like maybe Google Gemini 2.5
(20:56):
Flash. You can also process Excel files
with Excel files. Still a bit tricky with LLMS.
Again, what you can do, take a screenshot of the one tab in the
excel, send it as an image. NLM is pretty good at describing
to the content or getting back to you with insights and then
based on this you transform the data which is in a relational
database and then the rest is all the same as so far.
(21:18):
Nice. I'd like to pivot slightly
because one blind spot I have isI've been in tech for a decade.
A lot of the companies we talkedto are tech.
But there's so many non-technical companies out
there or just like slightly technical that I worry the gap
between productivity of us technical companies that are
leveraging AI for as many systems as possible and the non
tech who just aren't even aware of what's available is just
(21:41):
going to continue to expand. And we need to make sure
everyone kind of benefits from this technology.
So I believe you've recently hadsome experience deploying AI
agents in the fire safety vertical.
Would you mind talking about some of the experience there?
How the process went? Things to look out for?
Just what helpful tips could someone who's not a techie
person or not in a techie industry get to benefit from AI?
Absolutely, absolutely. I would say indeed companies who
(22:05):
are more technical have an edge,but it's easy to close that gap
for non-technical teams as well.The simplest way how you can get
started is just to start using AI and start using the tools
that you have out there. Start using ChatGPT on a daily
basis and they're the most common blocker that we see is
(22:26):
that Despite that, people have access to change GPT, let's say
they don't know how to use it. And then it leads to
frustration. And we hear the common complaint
of like, until I could have figured this out, I would have
done it by 10 times already. And I would typically agree,
yes, you could have. But have you thought about all
the other thousand times that you have to do the same thing in
(22:47):
the future? Have you accounted for that?
Usually the answer is is no. So the first step where a
company should start is just give change PT to your users or
cloud desktop, whichever it would be.
And either offer them a way to like AI training session, a
prompt engineering training session and like get everyone to
a base level to feel comfortablewith this tool.
(23:10):
I would say this is where where you start.
And this also ties into a framework which we see for
companies who want to adopt AI that it's not not all about like
technology only as we have been talking so far, like we have
been super technical, but it's also about people and the
process transformation. And this is where we believe
(23:31):
that the company really succeedsin becoming AI native or AI
driven as an organization when they pay equal attention to all
the three buckets of people, process and product.
So with people, it's simple. You just give them the tools and
let them to learn, learn it, to use it.
And also one thing which we not account for enough is give them
(23:56):
psychological safety to have themotivation and the will to try
the AI tools and to use it. Because if the attitude towards
the AI system is at all, this will come and take my job.
Well, I better resist as much asI can.
That will not lead very far downthe line.
So this has to come from the leadership team to set the
(24:18):
psychological safety within the company of like, hey guys, we
are not going to let you go. This is not the reason why we
want you to use AI. Our ambition for example, maybe
is to grow the company and we want to grow, but we do not want
to grow the team as well. We just want to grow the
business. Or it's like, guys, we saw that
you are burnt out. We just want to somehow find a
way to take away the boring, thethe busy work from your
(24:41):
day-to-day so that you can really do your happiest work in
this company. So that's why we need AI which
will take away the boring work so you can be happy and that we
can continue building a successful business.
And then you have the other side, which is the process where
it's about like you identify theprocess where you want to
implement AI. Of course, you have to have very
(25:03):
clear understanding of that process, map it out in a process
map. Once you have automated it, it,
it's not enough to just automateit and tell everybody, hey, this
is ready. This is the new process.
It requires constant reinforcement or we call it like
change management where you worktogether with the teams to train
them on the new process. Offer weekly office hours where
(25:25):
they can come and they can ask questions.
Offer them an A direct and an easy gateway where they can
report if things are not workingor they are not happy with
certain parts of the process or maybe they have ideas for
improvement. Because ultimately these people
will be your end users for the deployed AI system and after all
of these type technologies thereas underpinning which just
(25:48):
empowers it all. But it's not the top of mind
that you should be talking aboutwith your people everyday.
You should only be talking aboutthe people and the process.
Technology will empower it all. Yeah, that's a wonderful
perspective. So many people need to remember
that change is scary for many people and especially with the
doom and gloom you hear on the news as to AI is going to wipe
(26:08):
out jobs and display so many people.
The the thought of it coming in being like, hey, how do you do
your job? We're going to have AI work
alongside you is definitely intimidating.
But if you can reassure them that this is going to take away
the drudger of your job, allow you to be more creative, allow
you to focus on other things. You can grow, expand, learn into
a new role, or even just become much more efficient at your
current role where instead of being able to deal with X number
(26:30):
of tasks, you can deal with double the amount of tasks
without getting the extra burnout.
To to build on that one thing that I found really interesting,
just use my dad as an example. Told him a ChatGPT way back in
the day, the rush should often really think about it.
But then he introduced him to Suno, the music generator.
He took all of his old songs, made them into like new songs.
It was really cool. And then he introduced him to
perplexity as just like the nextstep being like, instead of
(26:51):
using Google just to try perplexity, now he's just
chatting to me on the regular because it was an incremental
progress of like play and enjoyment rather than just hey,
this will make you more productive.
Have at it it. It's a lot less appealing that
way. So I encourage people as I've
done a bunch of other episodes, as mentioned, a bunch of other
episodes just play like, explore, like there's so many
cool things out there that if you can just kind of get past
(27:12):
that fear of the uncertain that we're, you know, we're entering
a new age when all of a sudden, like we put electricity through
sand and then we get intelligence out of it.
It's pretty miraculous. It's gonna be a net positive.
It's just there are growing pains and we just have to try to
work on change management so that we bring everyone along for
the ride. Absolutely, Absolutely.
I would love to go into process a little bit more cuz that's one
(27:34):
area that really interests me. Do you have any strategies for
the initial collection? Like the documentation of a
process, should someone just like stand over someone's
shoulder and write down the workthat they do?
Should we try to implant some type of tracking software or
what is a low friction way that we can start documenting these
processes? So we have step 0 for process
(27:54):
automation. The simplest way how we start is
just interviewing the team members or watching over their
shoulder and spending a day or two with them and seeing how
they work is as simple as that. We have been doing process
mapping and like consulting has been around for many, many 10s
of years, right? So they have been doing process
mapping, looking at the process,reworking it.
(28:15):
This time is, is no different atall.
The only critical part is just to have full clarity and really
deeply understand the process. Many people that we see just
like quickly jump to a solution like, oh, we understand the high
process, let's now build AI agents and let's now build the
tools and you are missing the new ones and the detail.
(28:35):
And now, for example, with external parties, like external
parties like ourselves, the client really feels unhurt.
If you just come to them like they speak for 5 minutes and
you're like, Oh yeah, got it. That's now build AI.
So it always with also the part of lecture, understanding the
process, but also showing the client that you really get their
business and their context and their domain.
(28:57):
And also do not think that if you automated the process once
is the same across every company.
It's actually different for every single company because
there are cultural values, cultural analysis, ways that
people work differently, and allthis has to be taken into
account. Yeah, and that kind of rings
true with I guess not that recently more, but there's an
article about how 90% of AI projects failed.
(29:19):
And what's really unclear is were these people just being
like, hey, let's try this one product and just like through it
without really auditing if it was the right tool for the job.
What do you recommend for peoplefor selecting the right tool for
the job when there are 1,000,000out there?
They seem to really overlap, like distract.
You do everything like. How do people figure out what
tool they actually need? I think the article that you
(29:39):
have mentioned is actually the reason why those failed is
because they weren't clear around what they are automating.
I have a great friend in in San Francisco and he told me was
that in San Francisco the thinking is that technology is
never a problem. The only problem is your
business model. And now if we translate this to
the process, technology is nevera problem.
(30:02):
The only problem is you are unclear about what process you
are alternating and what is the ROI that you are that you want
to achieve and and you want to drive.
So I would say that 95% of thosepilots, they were all unclear
and they were just building sexyAI agents just for the sake of
technology. But actually they had 0 clarity
around what is the business value that will be delivered.
(30:24):
So then ultimately when the POC phase is done, they will
realize, oh, there is no business case.
So that's just shelf. What we do is we have often told
our clients that I'm sorry, we just cannot work on this product
because there is no ROI. We just don't understand what is
the ROI and usually ties back toimproving your margins, driving
(30:46):
top line somehow back to the EBITDA and PNL impact or in the
overall EV of the the the enterprise value of the company.
And if we cannot see that we don't even get started on that
work because it's a failed project and the failed
initiative from the get go. By the time you could have the
(31:09):
sexiest implementation, but there is no business case then
the client will be unhappy. If the client is unhappy, we
will also I'll be happy because it means we didn't do our job
well and they consider that we did a poor job.
Yeah. So that's what I would
recommend. It's just really be critical
around what is the ROI, drive itback to the dollar value.
(31:30):
And if the dollar value outweighs of the ROI outweighs
the investment, just go for it. That is a no brain.
Do you have any advice or techniques around measuring ROI
for AI? Because I'm just thinking, if we
give everyone a ChatGPT subscription, how can we?
And I know there's, there's sucha wide range.
So I'm thinking more generally, but how can we make sure that
the, whatever, 30 bucks a month that you commit to a
(31:52):
subscription pays for itself? Do we try to say employee
satisfaction output? Are there other things we should
try to take into account? Because you mentioned burnout
earlier, If we can have people deliver the same amount of
output, but there's less burnout, that seems something
cool. But how?
How do we measure that? Great question.
I think justifying ROI for just the ChatGPT subscription, it's
(32:14):
probably a bit harder and not sotangible.
Maybe you can measure it back tothe employee satisfaction for
example. But as I said, that's a single
player mode and so with ChatGPT access, you are not yet playing
the game of becoming an AI native company.
(32:35):
You are just doing the table stakes, which is absolutely
mandatory for everyone to do. It's just the foundations.
So you are not better or more successful just because you are
judging the subscription. This is not table stakes.
I think the unlock comes afterwards where you get into
the multiplayer processes and the ways to all to really
(32:58):
automate the process end to end.That's when you can dig into
Rois for example, just clocking of how long does it take to
complete this piece of work. Let's say if you are in the
legal industry and you are doinglegal research, you can clock
that same research takes 2 hoursfor you.
Now the challenge is that you will do the ROI based on the
(33:19):
hourly rate of the of the person.
And then maybe you want to get it down to 30 minutes or to 10
minutes driven by AI and then scale it across all the
employees and like how many hours the organization is
spending on that. And then you can start waiting
ROI versus the investment that is required.
So this is how I would tie it back.
(33:39):
Yeah, that's, that's mostly whatwe.
See what about for you personally?
What? What AI tools do you leverage or
have you implemented to help make yourself more productive as
you go about your day? That changes constantly because
this AI tools just get better and better every single week.
So my go to LLMS at the moment are Grok.
(34:00):
I'm testing clods about with clod, mainly clod code and open
AI ChatGPT. These are my go to LLMS.
Then in terms of orchestration, I have a personal NA10 instance,
so I can automate as much as possible there.
Whenever there is a repetitive flow that I find that I can
automate that with NA10 and I have been just testing actually.
(34:24):
There is another one, Granola. I'm a big fan of granola, using
it end to end. They just released a super
exciting feature which is aroundchatting with all your notes and
then getting weekly summaries. So I think the team is just
doing an absolutely amazing workin improving their product.
I'm a big no trend user now theyhave no trend AI so that also
(34:44):
makes life easier and I think that is retty much all I use and
Whiser actually yeah, I use Whiser for dictation sometimes
just writing the long prompts. It just takes ages and more hay
to just speak it out loud. Yeah, as someone who works from
home, I I talk to my computer far more than I type.
(35:05):
I yeah, especially now that I'veheavily integrated cursor and
cloud code into my workflow, I will just tell the prompt, read
the output, make a couple littlechanges, get back to prompt with
my voice again, and it's just the bandwidth difference is
phenomenal. You are not the first person to
bring up NA10, and as someone who's been developed for so
long, I have an aversion to it just because I want to write my
own stuff. But I've heard a lot of good
(35:25):
things. Could you talk about some of
your use case for NA10 and when it is a benefit to that instead
trying to write a Python script?Yeah, of course.
Well, you can write a Python script and have it in NA 10 as a
note. So you can.
That's the beauty of NA 10. So I think Zapier, I find Zapier
to be very just end user, user experience focused and not very
(35:46):
developer friendly or low level.Why with NA10 you could go into
the low levels as an engineer ifyou want to write custom
scripts. So that's an advantage.
You can also build your own custom nodes in NA10.
I think a big one for companies and especially larger companies
and enterprises is self hosting.So you can host it on your old
servers. Make sure that the data never
(36:07):
leaves. Then there is an advantage of
NA10, and this is similar to Zapier and all the rest of the
tools that they just have like thousands of integrations
out-of-the-box. So why bother with the Gmail
integration and Outlook integration, A Google Drive
integration? I wouldn't want to write that
code that's from scratch. I'm more than happy for NA10 to
handle it. So I think these are some of the
(36:29):
reasons. And also we are very excited
about the impact of AI in the professional services industry.
And the professional services companies just don't have the
engineering bandwidth or the engineering teams to maintain
code bases. So anytime for them it's a great
fit because they can automate the work flows maintained with a
(36:49):
very low engineering effort and still they get the benefit of a
fully built. Software.
This was a lot of fun. Before we let you go, is there
anything you want the audience to?
Know I'm curious from your end what AI tools you are using
beside your measured cursor cloud codes you are talking
about with your computer, What else?
Yeah, so I have a cloud desktop running for my main chat
(37:11):
interface. I still use open AI from time to
time, but more so for deep research only.
I actually do perplexity for guest research.
So I'll do a quick tell me aboutthe certain guest.
I use Aqua Voice as my transcriber.
The I used to use Super Whisper,but I migrate to Aqua Voice when
it allows you to have the optionof doing a like a screen input
at the same time so it can actually get context from your
(37:33):
screen. So if you're talking to someone
with an unusual name or you're doing code, it'll actually do
the function name the way it's written in your code rather than
just verbatim what you say. I use Obsidian for my notes.
I'm a big fan of the file over app principle.
And because it's just a collection of markdown files, I
can have cloud code running in that directory to create custom
dashboards for me, as well as things like the very robust plug
(37:55):
in ecosystem. So I can do a data view plug in
and actually have code operate across my notes where I can say
like get all of the unmarked to do's that are due in the next
week and put them in a list. That's been a lot of fun.
Cursor is my main interface for coding.
I've tried the different tools Ijust found.
I keep kind of migrating back tocursor.
It's not perfect but it's it's good enough for me.
(38:17):
I was big into Neo vim for a while, but now that I just typed
less because I'm speaking so much, I still use my my vim
bindings but it's just less of aa requirement for me to have
that like ultra efficient control over my computer.
Around the podcast. I use descript for editing.
I use Canva for the thumbnails. I feel like that's it.
(38:38):
I'll let you know if there's more.
But yeah, it's I tried to audit every tool that I'm using and
seeing if there's an AI alternative, test them out, some
of them. It's AI for AI's sake.
It doesn't add any actual value,just increase the price of it.
But some of them you actually get a minor bit of efficiency
here. And like I was emphasizing
earlier, the ability to pipe data from one system into a core
repository I think is under indexed in this day and age.
(39:01):
And I think creating your secondbrain, having well structured,
well formatted data that you canfeed into an AI system is going
to pay massive dividends sooner than most people realize.
So I really hope more people getinto the, the data hoarding,
data harvesting stage and just start saying, I, I write emails
on the regular. How can I start storing the gems
or, or you put out a bang or tweet.
How can you start building your data set of things that go viral
(39:23):
for the model that will help yougo viral more often?
I like that. I like that storing and building
up your own data set that's that's a really good one.
And actually, and one other thing too, like I, for example,
have a moon Lander keyboard thatI've cooked to connect a hammer
spoon to control my Mac. I don't really know Lua.
It's a very easy language, but just one other thing to learn.
But just by having a chat with Claude, I can be like, I want to
(39:45):
accomplish this. I have Hammerspoon.
I have like this config for my moon Lander.
And then it just like outputs the Lua I need.
And all of a sudden I have more efficiencies for my keyboard,
which aren't technically an AI tool, but it was AI derived.
Super cool. So char, where can people keep
up with you? Tell me what Bonsai labs?
Yeah, absolutely. So you can find me on LinkedIn
and similarly is Bonsai Dash, la-bs.com is where you can find
(40:10):
our company. And as as you mentioned in the
beginning, we work with private equity firms, work with some of
the top 10 firms in the world, but also with smaller and medium
ones. And we are the nimble and fast
moving AI value creation partnerfor this firms where we go
through the portfolio companies and we just help them get to AI
(40:33):
value as quickly as possible. So you can find me on LinkedIn
to reach out anytime. Thank you for listening to my
conversation with Chong or I hada blast talking with him.
He clearly knows his stuff and has done this a lot of times to
figure out the best ways to implement AI to solve these
operational difficulties. But I'm curious, from your end,
what processes are you struggling to automate?
Have you tried the tools you mentioned?
(40:53):
Are there any others that you like?
I'd love to hear from you. And I'd also love it if you like
and subscribe this video so we can get more of these
conversations going, so we can actually start relaying more
value to people so they can learn how to use AI to become
more productive and efficient. I'd like to give a quick shout
out to Tool Hive for supporting the show so I can have
conversations like this and I'llsee 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