Typically, AI innovation teams want to use AI to solve important problems that people or the business really need solved, which could not be solved by other traditional approaches. In practice, we tend to misjudge what the customer needs, under-estimate what alternative approaches can accomplish, and overestimate our AI abilities.  This is a standard recipe for disaster. 

Not too hard, not too easy, but just right: the “AI Goldilocks Zone” is the sweet spot where a customer’s needs meet the ability of the technology; and where the problem or project is not too easy nor too hard to be solved effectively. These considerations may sound basic, but I’m often surprised when I work with an AI company that is months or years down the road, but hasn’t taken them into account. 

Question #1: Customer Needs

The first question to ask is whether there is an overlap between what the customer needs and what the technology can do. Often referred to as product/market fit, locating this “opportunity sweet spot” (shown below) challenges the very best of organizations: it is a difficult, but crucial step. And many projects have failed for a lack of product/market fit, or one that is found, but then the market drifts and the company doesn’t realize the change. 

Once you’ve determined what the customer needs, you should then work backwards from your understanding of the features and requirements of the application or service as a whole to the characteristics of the AI “brain” inside. These may include its necessary performance (speed), latency, and accuracy. (Note that decision intelligence provides a structured methodology to help with this exercise).

Question #2: Non-AI solutions

Ensuring a fit between your capabilities and the customer’s needs is necessary, but not sufficient.

You should also ask about the alternatives to an AI solution, as shown below. What are the other solutions the customer has at its disposal? If a conventional, non-AI, system is available, then you should have good reason to think that yours could be considerably better.  You’ll have a lot of existing habits and inertia to overcome, so there needs to be a clear benefit. Don’t fall for the trap of thinking that AI is worth doing for its own sake: you want to have good reason to believe AI can do better.   

Ideally, you should measure the accuracy, performance, costs, and other characteristics of competing conventional solutions (including just plain humans doing the task – often your greatest competitor). This helps define the required AI bar, as well as to help to determine if an AI model could actually be the best method to solve this particular problem.

The good news regarding this question is that many AI solutions are surprisingly good: there is an explosion of new use cases for which an AI solution can provide a substantial bump to existing ones; we have not yet even begun to exhaust these possibilities in many domains, especially as you start to look at human-in-the-loop, non-fully-autonomous uses cases, as I’ll be discussing in a later post.

If you’re simply not sure about the ability of AI to solve your problem, then start with a proof-of-concept (POC): a project with the limited goal of answering the question: “Do we have reason to believe that an AI system’s performance along the metrics we care about can exceed conventional solutions?”.  During a POC, you can start with a bare-bones approach: out-of-the-box data (without worrying about feature engineering) and a straightforward algorithm without a lot of tuning, which is trained and measured manually (as opposed to embedded in a fully automated system for training, measurement, and retraining).

Question #3: Is it too hard?

Potential product/market fit and adequate evidence for improvement over conventional solutions are not enough, however, as shown below. What if the problem is so hard to solve with AI that we’ll never produce a viable solution?  Wise counsel as to the feasibility of an AI approach can be very important here, and I recommend working by analogy to existing successful solutions. 

For instance, if you’re looking to use AI to label cancer cells, then read up on other AI-based systems that do the same on a different kind of cancer, or some related pathology. Another example: you’re looking to predict future financial performance of a class of companies: look for AI systems that predict financials in a related domain. 

Also, beware of “false AI prophets”: don’t be fooled by an academic paper claiming success in the lab: there’s a big challenge and risk to take lab-based results into a commercial setting.  To minimize that risk, you’re ideally looking for a successful implementation of a justifiably comparable system that has been deployed at scale multiple times. 

Question #4: Is it too easy?

A too-easy problem can be an issue, too, because it might be easy for others as well. Assuming you’ve already ruled out non-AI solutions, is there an oversaturated market and/or an AI solution that’s already been proven to work cheaply and reliably, at scale? If so, you might be wise to use a simpler AI solution, or buy the core AI solution instead of building it, and then to focus your efforts on the remaining challenges of forging your path to market, including decisions like pricing, features, channels, and marketing messaging.  These are hard enough and still require the art of AI applications to be successful. 

You will usually not be able to find definite answers to all of these questions, perhaps not even any of them. Your goal, instead, is to develop your best initial hypotheses regarding the answers, and then to track them as your project progresses over time (perhaps with a formal risk management program).  

Asking these questions as a first step is what’s key in starting on your road to applied AI success.