Most enterprise artificial intelligence projects fail. They suffer from a fundamental flaw that prevents them from achieving their stated business goals. The flaw is so deep and so embedded that it can’t be engineered away or risk managed out. The flaw is not about improper learning data or poor platform selection. It’s not even about project management. This AI flaw is more devastating than those could ever be. Enterprise AI projects fail because they never started with large enough business value to support their cost of implementation.
Artificial intelligence is about making better decisions at scale, which can be implemented through computers and not humans. AI is fundamentally designed to replace one of the most time-consuming processes humans have – decision making. In order to economically justify an AI program, therefore, we must start with an understanding of the business value that results when we make better decisions. Not just any decision, but those that result in measurable actions. Measurable actions that result in better outcomes. It all starts with understanding value-based outcome.
AI business value is not the only consideration we need to make when when justifying a project. We need to also take a look at its cost, the economic impact of the effort we put into realize the capability. If AI to achieve its business end game, we need to ensure that the implementation cost is much less than the business benefits it achieve. This is common sense, but often overlooked. A question we struggle with in this area is, “How much more value does an AI project need to generate over its cost before we justify the start of the project?” I am glad you asked.
Best practices in the industry show that at the start of an AI project the baseline value to cost ratio should be at least 10 to 1. This mean for every 10x of arbitrary business value created, the cost of realizing the program should not exceed 1x. This results in the 10:1 model. This kind of return on value is a model that anybody would agree to. Who wouldn’t line up at a bank teller if they were giving ten dollars out for every one dollar given. But there’s a problem with this rule of thumb.
The problem is that humans overestimate value and underestimate costs all the time. Business benefit of AI projects are often overestimated by as little as 2X. That original 10x in business value only generates 5x in real results. At the same time, these projects woefully underestimate the effort it takes to build them. Instead of that 1x in cost, we see real costs are at least twice that. At the end of an actual project, the business is achieving more of a 5 to 2 return on value (5:2). This is still a great return. Again, who wouldn’t want to get $5 for every $2 given?
But estimating modern AI programs doesn’t stop with the value-based economic model. We also need to economically manage risk across all stages of its implementation. Implementations that run from proof of value (POV), to pilots, and into enterprise deployments. Each of these stages should explicitly generate economic value on the effort it took to build them. Again, there are some new rules of thumb that increase the likelihood of economic success for AI projects.
An AI project starts with a proof of value phase. This phase is not a proof of concept (POC) or a proof of technology (POT). POV explicitly demonstrates end-user economic value that can be scaled through pilot and enterprise phases in order to achieve the targeted business results. Our economic value target on the POV phase is just 1% of the cost it takes to build. This gets the “RPM gauge” off the lower peg. It shows the engine is running. It is a minimal demonstration of real business value. So for every 1x of cost to implement a POV project, we are looking to achieve a 0.01x of value in return.
Next is the pilot phase. This stage is all about scaling the AI implementation demonstrated in the POV phase. It’s not about implementing more AI features or functions. It’s about demonstrating that the value from deploying this minimal AI capability across a larger user base (a region, a class of product, etc) can generate more revenue than the cost of doing so. In many cases, a pilot implementation cost around 0.5x to deploy with a targeted 1x of economic return. This provides for a breakeven result under similar assumption from above, should the implementation cost are higher and benefits are lower.
Finally, the enterprise stage is all about the mass rollout of the piloted AI capability across all targeted user groups (all regions, products, etc.). For this phase, the rule of thumb is that for the additional 0.1x in enterprise deployment costs, there should be another 2x in economic value generation. This extreme high return ratio is conditioned on the assumption that there is no additional development costs. This is about deployment for value generation only.
Following this approach of proof of value, pilot, enterprise deployments driven by a value return, we see that we get an overall program return of about 2 to 1(1.9:1). This is a reasonable net return for any global AI program while managing risk using evaluate each stage. The highest economic risk is limited to the POV phase, where only 6% of the project cost is ensured before value is proven.
Artificial intelligence is all about value. The same value generated by their human counter parts. AI project fail because they do not explicitly start out by both defining that economic value and ensuring the value to cost ratio is high enough to achieve a targeted risk weighted returns. In addition, to effectively manage AI development risk, each phase of the project needs to have it own phased value to cost targets. By managing to a value-based model, AI projects will sustain 10:1, 5:2, or at worst 2:1 returns, while exposing only 6% of the project cost before customer value is proven. Who wouldn’t want that.
The post Using Economic Value to Reduce Artificial Intelligence Development Risks appeared first on Data Scientist Insights.
