How To Achieve the Promise of Generative AI

Using a generative AI tool is like owning a Ferrari.
You look great in it.
It’s a fabulous piece of machinery.
You perform better on the road.
Your friends admire you.
Here’s the problem.
Your commute has not gotten any shorter.
You’re still stuck at the red light.
So, no matter how fast you can accelerate from zero to 60, you’re still waiting with everyone else.

That metaphor aptly describes the challenge large enterprises face in adopting generative AI. I am shocked at the conviction that this tool is powerful. Yet our ability to get significant value from it continues to be stymied. No killer apps have emerged.

Enterprises have done countless pilots. Yet very few get into production. Our research says somewhere between 70 and 90% fail. That’s a horrendous failure rate.

These pilots probably are a headwind in their revenues without impacting productivity or transforming their businesses. In the end, they have NOT unleashed additional value in a significant way, given how much money they spent.

Why? The pilots are stuck at some form of red light, like our Ferrari.

Why the Generative AI Pilots Fail

Generative AI is indeed a useful tool. The pilots fail at an astronomical rate because enterprises are not putting the tools to work properly. They don’t assign tasks with a meaningful and significant ROI.

For example, we’ve heard stories that generative AI can create a 30-60% improvement in code development. Yet I interviewed a hyperscaler developer who pinpointed the issue. He said it is NOT speeding up his ability to develop great code.

Why? First, he had to clearly understand what he wanted to do, which takes time. Understanding the problem is what is so time-consuming. Then, he had to have generative AI understand exactly what he wanted to do.

He said he could just do the coding himself in the same amount of time.

Generative AI did help him debug.

Overall, he said generative AI did NOT make a material difference in his ability to develop great code. Like the Ferrari, the car is really fun to drive but doesn’t make the commute meaningfully shorter.

Stop Searching for the Killer App

There is no big diamond in the ring. When there is one, it doesn’t seem justifiable given the enormous effort required to redo the digital core.

Building the app isn’t the issue. It’s providing the environment in which the app can run. This work makes killer apps uninvestable because they don’t lower the cost to serve.

Where Generative AI HAS Made a Difference

The pilots that did go into production actually had compelling results. Studying these pilots demonstrates what it takes to get a disruptive and significant return from generative AI, the shorter commute.

For example, generative AI has been successful in the audit function. Radically improving the audit process did not involve a killer app. Instead, the auditors reframed the problem. You get an investable process when you reframe the problem into: How do I infuse the tech stack with AI?

Success was a journey of hundreds and hundreds of adjustments and small investments. We call this an infusion process.

The auditors wanted to:

• Reduce substantially the man hours required in an audit
• Cut the error rate
• Improve fraud detection

No single app does that. But reengineering the entire tech stack and the audit process allows you to get to that goal.

You have to frame these journeys in terms of holistic solutions against a business process or function.

The successful generative AI pilots received significant investment across three distinct towers.

• Technology. These companies purchased access to a CMM. They had to rearchitect their cloud systems so they produced more reliable data. This included moving their legacy mainframe systems into the cloud. They had to retire technical debt and close the white spaces between their different technology solutions so that the rest could work seamlessly.
• Data. Their tech stack needed to get better and more reliable. This required data cleansing. They made data management a priority. The thesis to date has been, let’s move the data to the AI and build a database from that. But that’s insufficient. We also have to move AI to the data. Enterprises have small pools of data generated by the applications. We can use the data where it sits by infusing AI into those applications. Data engineering needs to happen to uplift the technical stack so that it becomes deeply infused with AI.
• Business process engineering. This is often the most overlooked element. The companies with successful pilots had to change their operations team. They had to alter how they interacted with the workflow. AI is different from our existing technologies and requires more changes than others to how we do work if we are to really leverage it. Employees have to learn how to work with generative AI.

Tools are emerging which can:

• Monitor how the tech stack is being used
• Highlight where the bottlenecks or red lights are
• Show where people need training or where the process needs to be adjusted or could be adjusted

Every successful generative AI project we studied required significant investment across all three dimensions.

You have to copilot as you go along. The three stacks required constant iteration since they are intricately interrelated. They change each other. As you change the business operations, you uncover a need to invest in the tech, which exposes requisite changes in the data. A myriad of small adjustments need to happen.

Note: Although infusing AI is the primary thing, you’re doing a lot of the work that is not actually AI.

It’s a constant journey as these stacks infuse capability into each other. The business operations improve. This changes the quality and accessibility of the data in a profound way.

To go back to the Ferrari example, we have to eliminate the red lights and add lanes to the road. Maybe build a bullet train for the data. This is the formula for using generative AI to unleash business productivity.

Finally, we can now build real, measurable metrics that allow us to monitor our journey and identify where we’re failing to meet our objectives. This is far different from building metrics for a single application.

Handsome Rewards When Done Right

Deloitte and Ernst & Young used generative AI to completely transform their audit function. For example, fraud detection substantially improved. They need fewer people to achieve far better results.

We’re looking for step changes in productivity for a business function. We’re looking for new areas of value and improvement, such as fraud detection and audit or fraud detection in claims adjustments

We’re looking for improvements in the customer experience. These are measurable, not at the app level, but at the functional and the stack levels.

We believe you will enjoy generative AI success once you reframe your mindset into one of infusion rather than searching for that killer app. The question is NOT “Where is the killer app that will change my life?” but “How do I uplift this HR function? Sales function? Finance function?”

The Generative AI Journey Is Neither Short Nor Cheap

These journeys are neither short nor cheap.
Currently, there’s a standard one-to-one ratio between implementing an app and the effort for tech services to implement it. These AI implementations are one-to-12:50 to 20. Much more investment in the organization is required.

DeepSeek quickly got close to what the leading generative AI companies were able to do. They are selling AI at a dramatically lower price. DeepSeek will dramatically add momentum to the infusion process.

There are a lot of places where good enough is good enough at a low price. But there are places where the best is required. You have to be selective in the lowering the cost of AI, which is still high.

Conclusion

In my opinion, you should not be looking for the 100 places to add AI across all your business functions. Instead, select one or two areas where you think generative AI can make a substantial difference…and be willing to make the sizable investment across all three towers of work to actually get the benefits ChapGTP originally promised. And stop searching for that killer app!