How to Implement AI in Your Business or Start Consulting (Advice from AI Agency) [Update 08]

hey it's Max the original flow grma and this is the studio update the show about AI Automation and the people building it in this episode I got a great chat with Alex hona the founder of neurons Labs we zoom out from the usual use case interview and talk about what he's seeing on the ground as an AI consultancy and we also have a think about how to get prepared for a post AGI world I've also got an update on the auto tagging support tickets with AI project that I'm working on with N N's support team let's set the scene with some studio project

updates where are we with the auto tag support tickets with AI project so as a quick recap NN support team uses a it's a help desk system like zendesk new tickets come in and they've got to tag them right now they're doing that manually and it end's growing quickly we're getting hundreds and hundreds of new users a day and so we need to help them save time and we're picking this use case because it's not Mission critical so we roll that out test that out can deploy that and basically set them up so they can start adding their own order on top and solve their own problems so far I've got a proof of concept working and funnily enough the first thing we tried to do is use examples of previous messages now there were quite a few issues with that and the main issues were basically it didn't work as good as just providing descriptions so what we've done instead right now is written descriptions for each tag AI actually helped with that so we created some kind of semi good descriptions AI fleshed those out and so far I've got a little evaluation framework I built out in a workbook that's testing it this is so funny it's like me and parents babies right now out in the car dad's mostly actually get it's Dad Dy to anyways there's a few interesting realizations cuz at first what was happening I was only having one perfect success a few partial successes and a lot of failures but really interesting thing is how you define a failure so the first thing was we don't have all the descriptions of all the tags yet I'm working with a base of 10 right now some of them were failing we were piping in tickets that the system didn't know the tag of yet right so the good news is it's not hallucinating so far in the last days of testing it's not ever coming up with fake tickets but either comes up with no tickets or tickets from the list so that's good because we can add some more descriptions in testing we realized that some of the human tagged tickets can't always be trusted there's subjectivity in those as well we were looking at one case where we were like oh no the AI got it wrong and then we're looking at the actual ticket itself and we decided actually probably the human that did this could have gotten it wrong or been subjective to it so it's actually very interesting because we're going to be onboarding a lot more support people and imagine with each tag that we add someone kind of has to mentally think of 20 30 40 cases to consider each time they answer a ticket and the job is to answer tickets quickly of course cuz there's just like a crapload of tickets coming in right because you guys are all making it happen right so thank you by the way that's obviously all our job is to grow this thing each time it runs it's going to leave a message in the thread in zamad and say hey basically this automation run here's the execution ID of the Automation in case they want to go check that out here's the tags that it added and so we'll let them run that for about a week see how it goes and identify which tags it's not performing very well on and see if we can improve the description for those tags so what I plan to do there is then something like billing might only have a two- sentence description because it's a pretty easy case to classify by but something more proprietary specific Niche to n at n we might have a longer description and two examples so that way we kind of balance not blowing up the context window because one thing that happens one thing I've noticed is if you pipe in far too much context the AI could also get confused so having three long examples and a massive description for each tag doesn't make sense so again one adage I'm seeing all the time in AI is like it's not necessarily that less is more because being verose can be good but it's about start with less and ADD test of to working and then add because if you start with a lot it takes a lot more time to figure out what's working what's actually moving the needle and I found that in my project I probably wasted 3 days to working with these examples where I should have just tried with a description first would have been a lot faster and actually is working better now by the end of this week by the time you're seeing this video this should be up on an instance because I have it just running locally with a local llm model it's very interesting I actually so I'm playing around with some 32b models quen 2. 5S working pretty well right now there's also IBM has one called something Granite what's it called Granite Guardian 3 and was okay and then some Sky something 32b but anyways this is all going to be in a long form video on that project and then I'll meet up within the week after next check in on that and see how I can basically assist them and maybe tweak the workflow add some more context but get it in a way to where they can day on day be improving these examples owning the workflow so personally so that I don't get pinged as well but so again so that they can own the solution cuz I think the thing I love about Inn is how much you can hand off to people and again with sub workflows and stuff you don't have to hand off everything to them you can expose what you want to your users the consumers of the automation but they know how to solve their problems a lot better right like if they add a new tag for example and they know that it's failing in this specific case they're going to much faster than me identify what that case is identify how to describe that case into two sentences and slap it in there on average that is right so again the knowledge worker with problem I very much like to have them close to the prompting even if they're not the ones owning it them having a say in it so fine everyone I talk to just leads to better results and boy does it smell like manure out here okay there's a whole big seaming pile of it right there what's next party people so there was

some pains around self-hosting this particularly around the LM models so I used inin self-hosted AI starter kit it basically makes it super easy to stand up in Docker o llama and it and quadrant and postgress all the basics you need to stand up a lot of these AI use cases now I was first starting even with a 3 or 7B model it was painfully slow and I was a bit embarrassed I was like Hey and team like I have a 64 GB Ram Mac Pro whatever what gives and it turns out if you're going to run on CPU which is the case of you're running on Apple silicon you got to run AMA natively not in Docker and then have your Docker stuff like NN be able to reach out to your computer outside of its Docker container to communicate I did that and it was like 10 15 20 times faster still a bit slow right not something we're going to have for production use case but not taking Tak me 10 minutes to click a run go get a coffee come back if you're looking at the self Jose I start kit in the read me there should be a screen right now there is like a little note about doing that I use chpt to help myself but just the key word to remember is that you want ol Lama to run on your computer and your Docker to be able to reach outside of your docking container to communicate with it and then it was good to go I was able to download any kind of model I want there actually run it pretty quickly pretty well other than for the German internet of course took like 20 minutes to download a model so yeah on that project next week we're going to take a pause on that let the support team run that for a bit and then do a follow-up and then week afterwards and hopefully it'll be the point where it's a great working solution lots of learnings to share and I'll release a mini documentary on that whole project and release some templates that are very similar to what we did so you can basically deploy that for yourself and again it's for support ticket tagging but since it's classify incoming message based on a description list Etc as you guys know with a little bit of creativity you can apply that to lots of different kind of stuff of course and on that project so far fine tuning doesn't seem like it's going to be a good fit because the team does want to be able to change those tags more frequently than we were expecting but nonetheless I'm very excited about fine tuning again I've been chatting to some folks that have been teaching me some stuff on it so once that's all wrapped up we're going to get into fine-tuning and I think I'm going to try and like fine-tune a Gmail model it's basically trained on my previous messages something I wouldn't want to send to Big AI but something I'm happy to train locally and deploy a small local model that I could of generating email drafts that I can still as a human go in and tweak and I think that's really relatable cuz we all got inboxes probably most of us don't want to send our personal emails to Big Ai and if you're watching this you're probably Big Nerd so probably pretty interesting to you all right what's next and before this next bit this is a quick public service announcement for my own mental health I've done one of these before but guys my DMs is not tax support okay if you want to buy an Enterprise license go to our website contact sales I don't have a fast pass there you're going to have to talk to our sales team if you're stuck on a bug I'm probably the worst person to help you my inbox is hell right now because of all you lovely people thank you but then what I'm doing most of the time is copying and pasting that message and asking the team which obviously is super inefficient and we don't like that as an automation company if you have questions feature ideas you're looking for an n and freelancer any of that kind of stuff go to community. nn. we also got a bunch of the NN team in there and like hundreds of NN community members who are developers and professional automation Consultants who are all in there helping you for free basically because that's just the kind of community we are I don't know it's a kind of thrill when you see someone ask a nice thoughtful question respectfully that has good context if you write my credential no work with no other context like imagine if you got a ticket like that you probably snooze that one too go what the so don't do that all right PSA over

version 1. 77 of NN hit this week and it's one of these versions there was just like a bunch of cool little and mediumsized things in there for example the Deep seek native node is out but do watch out everyone cuz I just heard today that deep seek League like a million API keys of their users and something on an unsecured hard drive so like if you use it perhaps self host it a few moments next I had a really

insightful chat with Alex honchar as you know if you watched the show I really like talking with active builders in AI because the thing is when you got a lot of interest in the space a lot of kind of pantic influences so to speak being honest the more I learn about AI like people hit me up as an expert guys I'm a wide-eyed learner I think compared to the average do know quite a bit now but I'm still learning there a lot of BS out there I don't know I'll read someone's suggestion I'm like from the last 50 hours of LM prompt writing that I did that makes no sense but yeah anyway so on that note had a chat with Alex honard and in this one we tried something a little different we zoomed out from a specific nadn use case and we talked about the kinds of things that he's seeing from clients the kinds of requests he's seeing some suggestions if you're looking to get into automation or AI Consulting I think there's a lot of people thinking of going out on their own ESP when you have the likes of Sam Melton saying hey there's going to be a billion dollar unicorn perhaps you won't but you can probably do a lot more with a lot less these days if you know what you're doing with AI I'm so I'm sure a lot of us are thinking about it so we talk about how to approach AI in 2025 ideas on how to bring it into your organization and what the hell to do once AGI comes if you're a knowledge worker so on that note watch this next bit carefully hey Alex welcome to the show hey Max good to see you Alex could you give us a little intro sure I'm co-founder of AI consultancy basically my whole career I work in AI I did that think everything from teaching researching papers coding building companies so I'm here continueing this work that's very cool I think there's a lot of people joining the AI space now which is great obviously the industry is growing but it's always nice to talk with some veterans how did you get into Ai and what did that look like just to kind of set the stage it's a funny story because I was studing at University and I was studying mathematics and not because I was a good in mathematics because I thought okay it's kind of abstract enough field to study so I can study anything else that wasn't really good at it as it turned out studying mathematics is kind of hard so the find motivation was looking what actually applied mathematicians are doing basically applied fields and that's how found computer vision and that's how I found machine learning and obviously it's SP my interest because okay this is where the matri calculus the optimization that's where it's working that was the very beginning I got barely out alive with a C+ in discrete math so I definitely know what you're talking about but it's when you see it being applied to things that you can do with that for a lot of people get increase that motivation I think we're seeing that with AI today I think a lot of people learning Concepts they wouldn't have thought be getting into a few years ago because of that and then most recently so you're Consulting and Advising uh customers on building AI today it's one of the reasons I want to speak with you today again I think folks in the trenches have a lot of good insights can you give a little context on what you're doing right now just to frame the conversation so basically as most probably you know and listeners maybe have a guess that AI was evolving over last years and if for example at the very beginning of the company you were saying all look we are the smartest guys in eii machine learning we have phds Etc obviously right now it's much more about pench man management integration and let's say less about research mathematics So currently our focus on is at actually integrating sometimes ready sometime custom a solutions to the customer businesses natur would help them to do the change management both from the strategy perspective so the leadership is aligned and also bottom up teaching empowering individual employees to actually use the tools regardless if it's hgp or some custom C tool or to use it properly correctly and get the benefits if you look at sort of I guess the last 12 months for your Consulting work what are the use case themes that you see folks coming to you for or that you ultimately arrive at from Discovery again it might be really biased because every company has their own Market access we see a lot of interest kind of like on the two distinct sides of the road one is anything related to the customer interactions basically it can be as simple as customer support catbot call center can be some kind of personalization in the e-commerce or any kind of marketplace basically digital experience personalization calls and really deep back basically back office processes especially in documents processing again that's where what we see when there's a lot of interest in the last maybe couple of months also people see it as some kind of antic experience it's not just about even co-pilot okay I have a tool that helps me to do something the expectations now actually switching from augmentation which may was the worth of the last year to Automation and are you seeing you know folks deploying full agents you see a lot of folks talking about AI agents and I'm probably guilty of this sometimes too and it's more of an agentic workflow with a lot of human in the loop and a lot of sort of checks and balances are you seeing sort of fully autonomous agents being deployed or is it sort of more so of the hybrid at this stage actually for example even I'm kind of like technical guy by my background I don't mind mixing terms because again I think when technical guys trying to creating all the right taxonomy of terms that actually it makes things even harder actually I don't care if it's true AI agent or agentic workflow or in how FS role play system if it does automate work and reduces human involvement I don't care how you call it and you know many technical guys might not like me for understand this but which will also over complicate things D to here right often it is a gentic workflow so it's not really some orchestrating agents that autonomously defines everything we still need to basically Define even talking about the customers the for it works much better when thetic human a designer to design the stages to design almost like a set of states where this agent can be yeah you can say it's not fully automated it involves the human okay if it does speed up the work that's good the same about actual invol in the operations it's always good to have a switch even like a one state classifier that already can predict is this the question the cold that request that an agent agentic workl whatever can answer by itself or this is something unseen uncommon maybe emotional that you better give to the human you can reduce humans like 70 80 90% And that's just a r I think it's a really like sober way to look at this cuz you see a lot of discourse about people talking about some Frontier Model and if whether it's achieved AGI yet or not but that's not really like tied to do we have technology we can leverage today that can solve you know thousands and tens of thousands of hours of of work and I'm curious if you were talking to a small business owner or someone let's say in the Ops Team who's looking to upscale on AI in 2025 so they want to learn some of their own skills what advice would you give them where would you tell them to focus so I want to bring couple of mental models one maybe YouNow and the readers who of the Venture company blog called nfx I think a while ago they found like a five level e Spectrum basically how you can make yourself an a company starting from a basement that every employee of the company uses hgp or some kind of a tool to the kind end of the spectrum you become an open AI in your Niche right so basically when we talk about how to use AI to small companies spectrum and it's okay to choose your place so I would say my answer would be if you talk about a company say small medium business who actually doesn't do anything with AI I think it's good and okay approach to start with the level one basically get one Department 5 10 people the old use one AI tool of choice can be cbots it can be a marketing tool can be video generation right depending of Department can be like basically small Artic automation of some like emails stents flow SDR work if it's engineering actually make sure that all of the engineers use any kind of compilot not every engineer still uses co-pilot csor no what whatever it is so say this would be the very first step at a beer basis to level up it a little bit what we doing at our company we running a small hackathon both for Technical and non technical teams and this tools like n8n and other automation tools actually even like guys in sales HR operations they can take a Constructor and build a small automation themselves I find also really important where everyone has the game like so it's not like the management is coming buying new subscriptions and now you use hgp it's like we can get together on hakon find the tool choose a tool buy this easy subscription and actually try to do it yourself not because your manager told you so but because you actually said okay I'm working with the emails I did the research I found this tool I want to buy because it makes my work better so it sounds like if you are at an SMB or you know a small organization and you're competent with GPT and you want to take the next step that perhaps organizing a hackathon and trying to have the knowledge workers who have the problem solve their own problems is a good first step to kind of explore that and see what's possible yeah and also want to give a credit it's not you know my idea I learn it from actually one of our clients it's a Spanish startup penot and they run it at their company every half a year I think it's a great idea great initiative to keep people up to speed and give them like a freed to experiment to choose and I like this approach a lot yeah it's actually something that we do at NM as well with our team it's also a model I'm seeing be very effective can you share any tips on getting Buy in from decision makers at an organization when you're trying to enact that change so let's take a situation maybe you're someone in Ops you really see there's some pain points that you could solve and you're trying to get your head of Ops let's say to say yes to a proof of Concepts honestly I think proof of concepts are not going to make it anymore because Again by definition PC It's s basically just to prove that it's possible right that you already know it's possible there are a lot of tutorials lot of YouTubers actually many clients already said you don't need a p actually I think not what I think what I see is to get a Buy in need to demonstrate integration and result I think talking about PC right now it's a bit like outdated what they want to prove is like it's obvious that AI can read the emails they need to prove it okay actually show me that on your personal email is a personal account okay don't connect to their company that there some agent that can take the emails and maybe put them to hopspot or Salesforce whatever Ser we use it's like that's what can get the buy and of course the AR but think first of all not PCS on the MVPs with integration and second think about the right okay so you have to be able to demonstrate May here recording your screen maybe some other kind of analytics maybe some companies they use the time tracking with but need to demonstrate by numbers our employees spend on average 2. 3 hours in their email client every day you want to reduce it to 5 minutes and this is my MVP OB this my personal account how I implement this in workflow that takes my emails and pushes them to the CRM that's basically how we get the buy it's like it's similar in the framing of what it is but the little Nuance I think there really makes an impact and many people say oh it's like I cannot do it with my company data because I have this policy create like go to Google Gmail create a free account make some synthetic emails on the CRM just run it it's like you know sometimes the blockers are like coming I'm sorry to say from the close mindedness like people a bit like used to you know like in this organizational limits okay you can use your email create a new email create synthetic five synthetic emails just to demonstrate it's working on now one can for Breo to put this as someone you know I'm in Dev I'm creating use cases all the time guys if you ever need data AI is good at synthetic data you don't need a fancy prompt even a half house prompt is creating some pretty good results 100% Alex I'd be remissed if I didn't ask this because you run a successful AI consultancy I can imagine a lot of folks be if they're in an agency today or looking to get into this kind of Consulting what would you want them to hear as advice to know when you were getting into this and what should they be thinking about in 2025 if they're trying to sort of move in this direction I have to ask one question to specify are we talking about the folks who want to go solo as it to be like solo prur and public Builders or they want to build the organization because I did both and my work as a solar freelancer and building the company it has similarities but those that's very different let's start off with as entrepreneur because I think again you know you have Sam Alman saying we're going to see a un solo company I think with the tooling the idea of a Sol entrepreneur doing a lot more value is possible so let's talk about that case first what I see and again also basically most of my career I was never formally employed so I never had a like real employment experience and what really helps in this number one what many people who already used to work in corporate or for even for sta by organization is you need to be able really comfortable with taking risks and Haven seen the game you cannot go and sell to a client as say solar print or one person company and asking for the salary or you spending time you have to be able to sell the result of your work and for example what I did when I was starting I was saying to the clients so if I don't deliver you don't pay no it's kind of like a risk I mean I'm not saying I'm keep doing the same now when I also build a company but this also gives you completely different risk perception and also the clients are eager to work with you because they have no risk so basically I would say there are many guys who are obviously much really good in technology maybe even good in sales much better than me or whatever but there might people who are not comfortable with taking the risk the skills is different from the attitude and you just need to practice taking risk can be a small you don't have to quit your job but if you for example if you can in the evenings get such contracts and deliver them for the fixed price which they're basically risk free for the client I think you will be good to go I think it's some really solid advice and I think also highlights that it's not just in the technology there's a lot of soft skills there actually many people who are really talented both in technology and sales but they're not comfortable with the risk interesting they can be great Executives and startups and be companies but they just and go solo or set their own company again and I know person I'm not there are many people and always around me there people who better me in technology they always surround me people who are better me in communication and Sals I'm okay I'm not bad but I always see people better than me but I think what they do differently have much more comfortable with takeup risks thank you for that Insight I really want to thank you so much for your time before we wrap up is there anything you wish you I'd ask you actually I think what you should have asked me and this I think this is the question in the air is when AGI is coming and what we all are going to do in the AGI era all us smart guys who do with the automations technology when there is the agent who does Agents that's I think the question in the air that is a very good question I'm sure it's in the subconscious of a lot of us technical folks so Alex what's your take on that again I think one of my kind of like metas and mindsets is hope for the best but prepare for the worst it's also Associated to the risk taken and I think what can be the worst case scenario that you should prepare for is that the all the work that can be done on the computer will be automated not augmented but ultimate this is the worst case scenario so which means that the work that is done outside of computer basically meeting people selling building relationships also by the way taking risk because taking risk is also like a bit more like a meta skill political or quasi political things that anyway it's like you can boil it down to their relationships this is what states outside of computer this is what I think we still will be able to do in one or another way I'm worried about social Mobility because right now we rely on our talent cre skills knowledge to move and down the social Lea when this will be automated this will be weird because the way fact you know right now you know like last couple of hundred years Bally from the Renaissance till now the concept of social ability exists our freedoms are high and higher high I'm afraid that this can just go away but still basically the work repeating myself work outside of computer basically negotiation sales relationships some kind of Politics the risk taking strategy most probably I hope with still can do it and I hope that I'll be able to do it as well so it sounds like those soft skills might become more and more valuable as the computers can handle things like it's not just knock out a fod like a code function right worst case scenario again I'm not confident this will be that the case but I think if you can bet on this it's kind of aernal advice to live among the humans this is what you would just have to do again I'm not talking about the world when you even cannot do anything for humans right I think that then every since we discuss made a sense yeah makes a lot of sense well thanks for that hot take Alex I see that you're posting online Sometimes some of your sort of learnings and insights as you go if people want to hear more about your thoughts and

thinkings and experience where can they do that I'm mostly active on X and Linkin they're partially intersecting but on X I post a bit more things just bit more like random stuff some like links that I read LinkedIn I just post a bit more like longer things there's going to be a link somewhere here above or below as you watching great well Alex thanks so much and look forward to chatting with you soon bye yeah thanks for inviting with a do likewise hey thanks so much Alex best part of this job is talking with all these really intelligent people who happen to give me time of day because of inid end so thanks for the email really appreciate it Lis and yeah let me know if you like these kinds of high level interviews different from the use cases I've been doing I'm going to always try new stuff so yeah when you see something new just let me know if you like it if you say nah we don't do it again if you say yeah we do it more I'm not going to say it's a democracy cuz I definitely decide it's more like a benevolent dictatorship where if I don't keep you guys happy you decapitate Me by not watching my videos anymore so hopefully that's incentive enough for me to make the right call where are we all right gang that is a wrap for this week but I'm cooking up something extra special for next week the hint is it's 6,200 km away and I'm going to solve a big problem that I have with another flow gramar who's already been on the show this is only the eighth episode there's only so many people could be if anyone figures this out be the slowmo is going to be awesome and then this is a segue that on an operational note we paid for a YouTube audit right so some smart people looked at our YouTube and was like oh no good right you're doing this wrong you're doing that right but you're mostly doing this wrong and being honest as much as I hate to say it they basically said the studio update is no good what I mean by that they said YouTube videos need to be single theme right like the title of this video should be man walks in park and what happens next you get what I mean so it's Fair feedback it's something I've been thinking about maybe it's me being 30 years old and having a Nostalgia for TV shows um yeah so next week as part of that special little project that I just hinted we're going to try releasing the segments that I do on their own so it'll be more segments coming out more frequently I'll still do a walk and talk intro and outro of them for the longer videos like this special project next week it'll be more of a mini documentary so yeah so we're going to try that for a few weeks but do know I'm monitoring your feedback and unfortunately guys I've really don't have the bandwidth to reply to every comment right now I'm not going to reply with AI I think that's stupid being honest like the whole point is to connect right and to share information but I am reading them and so if you guys really hate it screw the algorithms we'll go back to the weekly episode and if you love it we'll continue doing it and if you hate it and the algorithms love it probably Lis will decide yeah whatever you do keep watching like any good product I'm always iterating I'm always taking feedback and selectively taking the feedback that I like to improve I'm Max Max Max Max I'm Max this is the studio update build something wild this weekend and happy flow graming