95% of Gen AI Initiatives Fail…Yours don’t need to!

3 ways CFOs can avoid the common Gen-AI Pitfalls

MIT NANDA’s latest report on the “State of AI in Business” is quite a topic of discussion these days in tech and business circles. In case you missed it, MIT’s NANDA project carried out a study of 300 public implementations of Gen AI (over six months from January to June 2025). The results are no less than astonishing, and equally disturbing. 95% of Gen AI initiatives bring zero return!

While we can debate or question the methodologies of the report, one thing is clear. Despite all the hype, the enterprise productivity from Gen AI remains largely elusive. In a way, the report is not very surprising. In the past, we have seen the digital initiatives’ failure rate being reported to be somewhere between 70%-80% by McKinsey and BCG, while the AI initiatives’ failure rate has been around 80% (Think Tank RAND report, 2024).

Now, one important differentiation we need to make here is that this does not mean Gen AI as a technology is failing. It is simply a matter of businesses not being able to find the practical use of a powerful technology, which promised trillions of dollars of productivity. In fact, this presents a huge opportunity for incumbent firms to step back and rethink how Gen AI can unlock such productivity. And as I argue in my book “Reimagine Finance” as well, CFOs are uniquely placed to lead this change.

Here are three important considerations or actions you should take as a CFO.

Copilot is not AI Strategy

Both the MIT report above and several pre-workshop surveys I do reveal one peculiar pattern of enterprise AI use. Almost all the firms I have worked with have one form of an enterprise version of ChatGPT or Microsoft Copilot. Now, somehow, leadership is expecting these tools to create workplace productivity. However, the usage is limited to creating emails, writing memos, drafting town hall speeches, etc. There is no significant enterprise productivity unlock.

The biggest problem with these tools is that they are general-purpose large language models. We cannot reasonably rely on the output as it is probabilistic and non-deterministic in nature. For example, you cannot (and should not!) just use an analysis alone, done by Copilot, to decide which sales promotion to run or which machinery to invest in. The reason is that, firstly, the output depends largely on what prompts were made to the model. Every person’s prompting will use different words, and hence the output will be different. Secondly, how context-aware a model is. If it cannot, for example, bring in your company’s financial and sales data in context before answering what kind of promotion you should run, the answer is generic advice as opposed to a specific recommendation based on your situation.

Therefore, it is essential to understand the limitations of these enterprise Copilot or GPT models. Unless we are making them context aware, performing rigorous testing, and training employees to use them in the right way, they will always  be restricted to sporadic individual productivity use.

Choose the correct AI for the right business problem

Similar to how a general-purpose Copilot is not suitable for building a promotional or pricing strategy, an AI agent might not be the right solution for workflow automation. Your workflow might need a simpler low/no-code automation as opposed to a full-fledged autonomous AI orchestration.

In 2024 Think Tank RAND issued a report on the major reasons for the 80% AI failure rate. A major finding was that businesses were not solving the correct business problem using AI. Either the problem was not an ‘AI problem’ in the first place, or there was an overconfidence in AI’s capabilities to address it.

It is, therefore, essential that we create an organisational literacy to evaluate the comprehensive spectrum of AI capabilities. Which problems are suitable for traditional AI, like Process Mining and Predictive Analytics, to solve, and which ones are appropriate for advanced probabilistic models?

Once we build an appreciation of different AI applications and how they are well-suited to solve specific business problems, we will eventually notice that it is not one AI for one problem but a convergence of different AI techniques to solve one or more business challenges. For example, you might use process mining to understand the current procure-to-pay process (P2P), but then use traditional machine learning to automate invoice processing. On top of that, you might want to train your Gen AI models on the entire P2P data to be able to interact with data and generate reports, variance analysis, etc. Our role becomes more of a conductor of an orchestra as opposed to using individual tools and hoping to create a symphony.

Experiment…Thoughtfully

Technology is evolving fast, and so are the business problems we are facing. Standing still is not an option. Nor is it possible to open up our wallets and start spending on every other tool we can find in the market. What we need to do is carefully evaluate the key business problems we are trying to solve, define clear and measurable goals for what we are trying to achieve if we solve that business issue, and then prioritize based on business needs. Only then do we ‘earn’ a right to experiment with technology. Again, the anchoring here is the business problem we are trying to address, and not technology itself. This is what we can call a thoughtful experimentation.

One word of caution, though. While experimentation is crucial, building an AI and Digital strategy in parallel is equally important. The balance between operational execution and strategic thinking is the key here.

Today, the world is filled with a mix of hope, fear, and enthusiasm when it comes to technology. On one hand, it is promising trillions of dollars of economic value, while on the other, there are a lot more questions than answers on how it will be able to do that.

As I always say, there is no single hack or framework that is going to cut it for you since every situation will be different and would require a peculiar orchestration of AI, human-centricity, and change management. However, the above principles of acknowledging limitations of Generative AI, choosing the right solution for the right problem, thoughtfully experimenting, and most importantly, anchoring technology decisions in clear business needs, are how we can create a high likelihood of ending up in the 5% side of the MIT statistic.

About the Author – Tariq Munir

Tariq is the Author of “Reimagine Finance” and advises businesses on unlocking the potential of AI, Data, and Digital. He is also an international keynote speaker, trainer, and monthly columnist at CFO Magazine A/NZ.

He can be reached at [email protected] or www.tariqmunir.me