Introduction

I wish some of you were here for my first presentation. I talked about AI as a thinking partner. It was way more fun. This is a little bit more boring.

I tried to make it more interesting, but this one is going to be a little bit of theory and then some practice. But we'll try, we'll try.

Why AI Projects Fail: The Big Picture

So why AI projects fail? I'm going to only cover three key reasons, and then I'm going to give you some practical points.

Because there's many reasons, I think there are three main ones, and then we'll go to the Q &A. So, but we're going to cover the core point.

There's eight core points that I think are important. And then I'm going to show you some of the case studies that are really cool. Right. So this is my favorite part.

AI Adoption in Real Companies

AI adoption. You're working for a company and they want you to adopt AI. What usually happens is they go from the top to the bottom. We have to do something with AI.

We have to be good with AI. So what does it look like? Well, it kind of looks something like that.

What Employees Actually Feel

This is, I did a Reddit. research. I think I put like 3 ,000 posts or something like that. These are recent.

My favorite one is this one, the one in the bottom half. My boss expects everything to be done twice as fast because the AI is doing the work. This is just great.

I love how we have to appease the god named ChatGPT until we get an 8 out of 10 on a landing page. I mean, that happens so often and a a lot, a lot in so many companies.

This is what a lot of people feel like when they say, oh, I have to use AI, like you have to figure it out. Not very good, not a great start.

The Adoption Gap (Benedict Evans) and What It Implies

But they actually did, let's go through that quote and then I have a point around a survey that S &P did with companies.

So Benedict Evans, he's a very famous tech analyst and he said this, there are people in tech who are running agents 24 seven, there are small numbers of people outside tech who are automating parts of their workflow with

Claude. Then there's everyone else who's opening this every couple of days at most.

Most people and most companies are not tool builders who readily, quickly rebuild their workflows. This is something he posted just a couple of days ago.

He's a super, super smart guy right at the very front of tech. And this is one of the really key things why there's a massive gap between people who are doing these insane things and normal users.

And it's so easy to get lost when we spend a lot of time with AI or we just use it regularly.

And the fact that his last point around people are not being tool builders, this is one of the core things why the projects fail. And there's a few more that are really important.

Evidence from the S&P Survey

So, S &P did a survey, and I think they, this is 2025, 42 % of companies, they failed with their AI initiatives and they scrapped it. This is up from 17 % from 2024. Just didn't work for a bunch of reasons.

So why does that happen?

Failure Mode #1: No Clear Problem Definition

Well, the main one, the main one, there's no definition of what the problem is. Usually, like I said, technology gets pushed down. We have to figure out AI. We have to make sure we're not missing out.

So there's no clear problem. What are we trying to do? If

there's no clear problem that has been defined, there's no metric. So nobody really knows how to succeed. Nobody knows what the number we are chasing.

Are we getting the hours? Are we getting the something? It's not very clear.

And the project itself is not clear. Usually gets pushed to the CTO. We have to figure it out.

They hire an outside consultant or consultancy. They get the project done. It works in a demo.

Then they leave and then they realized this doesn't work at all.

This is one of the really key reasons why this stuff doesn't work.

Case Study: The AI SDR That Backfired

There was a B2B sales organization, a few companies bought AISDR to run their sales. Failed miserably.

AISDR sends 3 ,000 emails a day chasing all the people. I think they had even a stat. They got like 42 meetings booked. Only 11 people showed up. Zero calls booked.

booked, and then they got their real SDR, $75 ,000 a year, not $180 ,000, $75 ,000, book in like 12 meetings a week because it's a real person.

Because the AI SDR congratulated one of the prospects because they raised the round, which was complete bullshit. Not true. And then, of course, you run in B2B, you have a high ACV, average contract value, you cannot afford that.

But why did they start? was it a problem with a real human SDR? It wasn't a problem. It was like, oh, AI SDR is the future. We should do that. No clear problem definition.

Nobody clearly said like, oh, what is the problem with this person? Are they good? Of course they are. And there's quite a few case studies that really talk about that.

Failure Mode #2: Nobody Owns the System

Then another very common problem, nobody really owns this stuff. of. They brought this tech, they did the demo, and then they disappeared.

Now, this not only happens with AI. I worked in marketing operations with Salesforce, with automation before AI came out, Marketo, Parda, these things like that.

A lot of times they don't have an owner. There's a lot of people, nobody owns this operation part, and so you come in, they get help and it's a mess. That was the deterministic system.

Here, with the black box, with what Michael talked about, it's far more important to have somebody who can maintain, run the system, improve the system, and teach people how to do it, how to work through it, especially when people are not very interested in using it as much as the management wants to push it down.

Failure Mode #3: Verification Becomes Extra Work

The other part is it didn't really remove the problem. It only added another one.

There was a great quote. I loved it. Somebody said it.

Now I'm a secret layer in the company translating the AI output to make sure it makes sense. And that is a real problem.

There's no verification or the verification is not good enough because people haven't thought about that. And that creates a lot of issues.

Now, are we saving the money? Because if you have to verify everything, thing? Well, maybe we are not.

Case Study: Starbucks Inventory Counting

Starbucks, they implemented AI. It was a laser tablet technology that was supposed to be counting bottles, milk, syrup, and everything.

And they promised 99 % accuracy. And it's going to work great.

And it double counted. Or it didn't count properly.

So then humans who are doing baristas who had to count all their inventory, before that, they had to recount what AI allegedly counted, and then they scrapped the system in May this year, because it was extra work.

But it was a great idea that was supposed to work, but it only lasted nine months. But there's a, I mean, what is the problem?

And I think this is one of the core problems, Michael, I think he left the room, this was, I mean, I thought he was going to see it.

The Hidden Blocker: Low AI Literacy and Resistance to Change

We had an interview, I think it was last week, but in most organizations, 80 % plus employees don't have general AI knowledge across the board yet.

That means you are doing the transformation without the basic knowledge you actually need to have.

So, what we have here is a phenomenal problem, is that employees don't really quite understand exactly what it is, they have a real resistance to use this thing, and the management doesn't quite know exactly how AI works too.

How could you run a transformation when you don't fully understand the thing and there's a real resistance? Like it doesn't work and then there's no owner and there's a lot of other things that are happening.

We should probably just do that. And that's the problem. That is a real problem, Michael.

I'll show you the quote later. But it could be done right.

What Doing It Right Looks Like

Case Study: Ford Rebuilds the Workflow Instead of Scrapping AI

Ford did it really well. at least they recovered very well. Ford, they had their quality control department, and then they outsourced it to AI, they got all the design requirements into AI, the AI took it in, and it wasn't working.

And they fired, I think they fired a couple hundred of their QA engineers, quality control engineers. And there were defects, and Ford did it really right.

They They rehired the guys, didn't scrap AI, rebuilt the workflow, went number one in quality in 2026, first time since 2010. That's what it's done right.

They got these guys who are smart, and they made them better. They used AI for the stuff that makes experts better. And that's really how it needs to be thought about, not we are going to automate and we don't need people.

Because with that approach, well, there's a lot of companies that tried and will try and not going to succeed with that. Because you need expertise.

Practical Example: A Conference-Ranking AI System

Well, this is, I'm going to show you a quick demo. This is a prerecorded one because it will take a bit of time. I've been working on an app. This is, we've been working with one company to create a system that shows which conferences are more likely to give sales results.

Which conferences in B2B are most likely to be effective in bringing sales. sales. I have put it together.

Let me just see where this thing is. There we go. This is just an animation.

By and large, it works around like that. You have a list of companies, a list of conferences that you put in. You could also do a discovery.

You run the analysis. does a lot of takes a lot of pages in which conferences are most likely to bring in the sales for a b2b company so this is uh this particular company is in biotech very niche

b2b space and they are looking to attend hyper niche conferences in life sciences sciences, biosciences, and all these kinds of stuff that nobody knows what it is. And I barely know what it is, but it's in five languages.

A lot of it is, these are niche European conferences for serious hardcore scientists. So what they're looking for in the marketing, which ones are we attending and which ones are we should not attend?

Because if you talk to the sales team, they'll say, of course, go for this one. But that's not quite objective. And if you're a leader in the marketing team, doing all this research yourself is hard. So how are we going to do that?

Well, this system is going to take all the conferences and it's going to rank them based on the real evidence that we could find. And you're going to get something like that.

You're going to get a few recommendations. There's an inside report, but here you get like three conferences that they think is a good idea. and then you can get some information about what it is, why it's a good idea, like buyers, competitor.

So you don't have to spend the time figuring it out all yourself. And this is only based on the publicly available data. That's how it's designed at the moment.

So really you're taking, and then you can look at the buyers. There's also conferences that have specific buyers that you have for your ICP.

Then you have competitor that's getting there. and then there's like if you have a deal and they had a particular deal uh that they closed then you know the system will uh will take that as well and so you're basically taking like 2 700 or like

close to 3 000 pages this is for 56 conferences to i mean really what you're looking for is this you're looking for really like six points like a marketing lead or ceo or executive leader they don't really care about the numbers they care about like which conferences do i have to go So tell me quickly, which ones do we care about?

So what we try to do is we try to take a lot of data, and we try to say, these are the ones that we care about. Consider those.

And so there's a bunch of technical stuff that happens in the background. And we'll get to that.

Let me just see where is it. No, hold on. This is not it. Yeah, there we go.

And we can come back to it.

How the System Works (Crawling, Verification, and Modeling)

So we crawl the pages. We get the information we need.

We're going to verify it, and we're going to judge it. and then we're going to make sure we verify it again to make sure what we are saying is true.

And then we're going to get the dashboard. And so a little bit more of a compressed technical piece

is we're using the Python, then we're using essentially three models, and we're trying to make sense of the type of information that we can get online to make it easier

to judge what we really care about. And that is based back to the verification verification and back to the stuff that matters. Now, this is only using public data.

Where It Fits in the Business Process

Of course, if you are building it into a certain flow, you want to add the data from the sales team. You want to add the data from people who are on the floor from B2B. Maybe you connect it to Slackbot that sends message. And then there's a feedback loop that improves the system and it gets better.

But it only matters. This stuff only matters if B2B conferences are a big part of your budget and there's a real velocity that you have. You're attending a lot of these conferences and you need to figure it out.

I know, for example, quite a lot of B2B companies, they spend one, two million, three million dollars in conferences, they go to six. They don't need this.

Those who have real velocity for really niche conferences potentially may find value in stuff like that.

So AI is supposed to collapse the expensive part. That's the whole point. It's supposed to take the stuff that takes a long time to make it faster so you verify it quickly. quickly.

And then there's real value. If you have to do all the checking yourself, it doesn't really make a lot of sense.

And companies are not tool builders. And what happens is they take AI and they put it to the same flow that they had before, the same way they worked before, the same how things worked.

And it just doesn't work that way because the workflow needs to be changed. People need to change.

That transformation takes a lot of time. This This is not an overnight process.

And if you try to remove people, that's a fast way of going fast and breaking things. It doesn't work. It doesn't work very well.

May work for some startups, but not really. Not very often.

A Practical Checklist to Avoid Failure

So we want to start from a problem with a real number when we are looking at how do we, to make sure this AI doesn't really fail and it works well. Simple. Problem with a real number.

What are we trying to improve? What are we trying to change?

We need to get it to somebody who is going to own it, who understands this stuff and is going to do it better and is going to work through it over time because systems change super fast, architecture changes fast, all of these things are in very, very quick motion.

We need to look at the budget. Is it going to be 50 million tokens?

That system that I've been working on, if you don't use API with search to connect to Brave or Serpa, it's going to consume like 80 million tokens. It's going to be like $12 per conference.

And there's not a lot of value because it's going to take, scrape one website, which is official site of the conference, and two other ones that are just directories, just like taking the numbers and move it on the screen for a lot of money.

It doesn't make any sense. And then

we want to make sure it's not full auto. Better. It's better if it's not full auto.

Like there's some checks that human has to do. And I think this one is, Michael, this one is yours.

Don't rush. I totally agree. I think the bigger the company, the slower approach you want to take.

Unless you are willing to rip it apart and then take it off, but then it doesn't work super well.

Conclusion

That's pretty much it. Thank you, guys.