85pctAIfailure

85% of AI projects fail in the business world, or how to be in the 15%

There is a reason it is your business, you have put in some of yourself into the business and it has worked for you and your customers. You have heard many things about Analytics, data, AI. But how can you make your business better, without losing what makes it special?

Yes, hiring a bunch of programmers can help. Especially nowadays, learning programming is not as difficult as before, programming courses and schools abound. Hiring people with programming skills is not difficult.

For example, 5 years ago I was in Myanmar trying to create and run operational efficiency and analytics for an insurer. This was Myanmar, after only 5 years after first ‘free’ general elections. However, I could find local talent in Machine Learning.

Businesses are running, finding coding talent is not an issue, so why is it that 85% of AI projects fail (1)?

You may realise that the points put forward are not new, I have been griping about them for a while in the context of Analytics and Machine Learning. But how has AI/GenAI changed things?

Preamble: Some Intuitive Definitions

Analytics is about solving business issues, and I usually refer to Analytics/Data Science as having a predictive element, as opposed to Business Intelligence that is backward looking. The aim of Analytics is prediction (or prescription) in a business context.

There are many tools for Analytics, and Artificial Intelligence (AI) is one of them.

In simple terms, AI is about using machines to mimic humans, and this can be applied to solve problems. This could be to solve how to get a pair of legs to walk across obstacles.

Within AI, Machine Learning (ML) is a specialization that uses algorithms to automatically extract information from data and help detect patterns and make decisions. Crucially the ability of ML improves as the models are exposed to more relevant data. Many techniques of Analytics involve ML. ML may or may not involve Neural Networks, but even then, the networks tend to be simple, having an input where data is read and an output where the result is obtained.

Deep Learning is a specialization within Machine Learning whereby the Neural Networks that mimic a human brain get more complex, and have more than two layers, layers that allow for more complex interactions, adding depth to the model.

GenAI is an application of some deep neural networks where the focus is on the ‘creation’/’generation’ of “new content” based on the data that the networks have learnt (existing data).

Hence, to me, these are different tools that someone can pick up and use depending on the problem they are trying to solve.

For example, if you are faced with a leaky faucet, it doesn’t mean that the latest electric screwdriver that allows you to screw at an angle will be the solution.

85% of AI Projects fail

The main reasons are straightforward:

There is a gap between business and coders that is hard to bridge, or as they say in the region, it’s like a chicken trying to talk to a duck.
The data that is required is simply not available
The focus is on the shiniest toy in the toy store, rather than the toy that works.
The organisation does not know how to use AI/Data
The problem is not one AI could or should solve

1 Gap between business and coders

Many people are initially surprised when I claim that the coding part is the easiest piece of the implementation of an AI/Analytics model. The important part is actually conjuring a problem that should be solved, and agreeing how to define success.

In my leaky faucet example, it may be quite clear what the business problem is, or is it? There are many ways it can be solved, for example placing a bucket to contain the dripping water, shutting off the mains, plugging the faucet, creating new plumbing to place a new faucet next to the leaky one… All these deal with the leaky faucet, true, but are they an acceptable solution for the problem?

2 Availability of Required Data

There is no escaping that data is the lifeblood of analytics/ML/AI. The advent of Chat-GPT/Gemini/DeepSeek… may make you believe that all answers can be found by AI. But it is important to know that everybody gets similar answers. If everyone runs their business the same, what do you think will happen? The basic fact remains that the AI does not know your business.

It is crucial to unravel the complexities and address the basic questions. Do you, today, have sufficient and the right data to embark on a journey that would use data to improve your business performance? It’s not just having tons of data, but the data needs to be useful, relevant to the problems. The old adage “garbage in garbage out” still holds. If you don’t or you are not sure, how can you start?

Masayoshi Son of Softbank (2) once said “those who rule data will rule the entire world”. Would you like to rule your data?

In my leaky faucet example, imagine there are dozens of faucets, and water supply has been cut-off. You have no idea which faucet to deal with. Or, you look at your tools and realise that instead of carrying your work (plumbing) tools, you have only your hobby (oil painting) tools.

3 Shiniest toy

Today AI is ubiquitous. A geek would not be a geek if he/she was not kept awake by the new developments. Does it matter to you are a business or isn’t it better to use the tool that will give you the results you can use, and allow you to grow?

Is it better to get a fish, or to learn how to fish, especially depending on where you are fishing?

In my leaky faucet example, maybe I insist on using the amazing new screwdriver. It will work if the issue was with the cartridge within the tap itself, not say if the leak is due to a broken washer, or deteriorated o-ring…

4 Organisational Readiness

This is something I learnt the hard way at the beginning of my career, just because I could build a model that shows which customer is likely to get a specific credit card doesn’t mean anything until the bank I was working for put that model to use.

(Obviously, by extension, it does not matter what toy I used to build that model)

Many people forget implementation is crucial to any piece of data or technology, and if the business/organisation is not ready to use the output of a data exercise, then the exercise has been a waste of resources.

In the leaky faucet example, may be nobody brings the plumber to the faucet to be fixed, say due to unclear authorisations.

5 AI/Analytics is not the answer to all problems

To me, this is related to quite a few of the issues highlighted above, but basically not all problems are worth solving. You run a business, and resources are not unlimited. In most organisations it is important to use resources in a way consistent with achieving your business goals.

This usually means that the questions that the AI/ML/Data products are meant to solve should be directly related to your KPIs, therefore related to the levers of what can make your organisation sell more, manage costs better, be more efficient in operations, minimise risks for the same level or returns, give your customers a better experience.

One of my ex-bosses, a Maths PhD from Berkeley used to argue whether we were moving the needle. To me, if the needle is not being moved, then there is little point in the exercise.

In my leaky faucet example, maybe the person trying to solve the problem is a data scientist rather than a plumber.

Conclusion

It is not rocket science to be in the 15% whose AI/ML/Data endeavours are successful. It takes level headed and systematic thinking and planning. The data journey does not have to be complicated, it takes the business, data and technology to truly work together to achieve common goals.

And for those interested in fixing leaky faucets, a nice explanation can be found (3)

By Kailash

Report abuse