We often hear or read that artificial intelligence will change the world as we know it. This misses the mark; AI has changed the world that we live in forever. It isn’t some futuristic technology that will alter our lives someday. It’s a feature in products that we interact with today. Artificial intelligence isn’t a single technique. It’s an umbrella term for a large collection of techniques with a variety of applications. Here are some you may be familiar with:
- Planning a route between two locations for a vehicle
- Speech to text systems
- Recognizing the location of faces in an image
- Placing ads on a web page to maximize click-through
If it’s true that AI has already changed our lives, why is there so much recent interest in the technology? There has been a sea change in artificial intelligence. Continued research in the field has reduced the barrier to entry. It’s no longer the case that you’ll need to produce novel research to use AI to solve your problem. Often, techniques from the literature can be directly applied, once they’ve been identified. This gives you a baseline to tweak as performance and situation demands.
Machine learning is the poster child of this phenomenon. Libraries like scikit-learn and pytorch are essentially house-hold names. What’s more, the largest cloud services offer turn-key solutions for machine learning. This democratization of AI levels the playing field substantially. No longer is AI the sole purview of research houses, large corporations, and nation states. The leveled playing field means that everyone can apply AI to their problems, if we can only learn to recognize the opportunities.
The Shape Of An AI Opportunity
- A repeated task that takes some amount of human attention
- That amount in aggregate is high
- Outcomes have measurable correctness or quality
- Incorrect / suboptimal outcomes are annoying, not catastrophic
If an opportunity has the preceding features, it’s likely that AI can help do it better, faster, or cheaper. Let’s look at each of these pieces in turn.
A repeated task that costs some amount of human attention
There are an infinite number of tasks that take some amount of human attention. Tasks suitable for applications of AI have some additional special properties:
1. They can be rigidly defined
Understanding an image isn’t specific enough. Being able to identify cardinals in a picture with well lit scenes isn’t either, but it’s much closer. Being able to identify male cardinals in a well lit picture with 95% accuracy and taking less than a second per image is very close. What you’re looking for is a narrow definition with acceptance criteria.
2. They require human intervention
There has to be something innate to the task that requires a human mind; it’s not enough that a human is doing it. A person can sort physical objects based on size or opacity, but the person isn’t central to the task. A mechanical device with a light sensor and some physical gates could get the job done too.
3. They’re repeated
There’s a cost associated with developing and deploying an AI solution to a problem. The more often a task is repeated, the easier it is to amortize AI development costs. Using AI to ingest paper registration forms for a company that adds thousands or tens of thousands of new clients a year makes great sense. It’s nonsense if you only serve a handful of high value clients every year.
That cost in aggregate is high
It’s not enough that the task is frequent and demands real human ability. There also has to be some aggregate cost associated with the task. Again, this goes to absorbing the cost of AI development and expected ROI. Trivial, infrequent tasks, such as retrieving my physical mail, are bad candidates for AI. It takes a few seconds, it happens once a day, and building a robot that can navigate my yard or changing the environment with a pneumatic tube are both costly endeavors. On the other hand, digitizing a board of post it notes so that I can have a physical kanban board and a digital copy of it is a good candidate. That work takes many minutes every day, to the point where my digital boards and physical boards are often out of sync for a period of time. Software to digitize the board for me relieves both the cost of doing the input and the cost of lacking synchronized work boards for local and remote collaborators.
Outcomes have measurable correctness or quality
This is really important for all features, but it bares repeating here. The ideal tasks for AI have measurable outcomes. This is important for several reasons, but we’ll highlight two here. First, many businesses don’t have a depth of experience in AI. You’re likely to find yourself doing the first AI feature. That means you’ll be selling the organization on the feature and the application of AI. Thus, a direct way to show impact is desirable. Second, measuring how the AI performs lets you know when the AI performs well enough to stop development and release. Like any feature, AI development is somewhat costly. Measuring performance lets us know when to move on to more urgent or rewarding tasks.
If the AI is Wrong, it’s annoying, not catastrophic
I’m not sure if this is universal, but I find it difficult to relinquish control of things. Moving agency from people to machines can be a troubling prospect because of that feeling. It’s easier to swallow is if bad outcomes in automation aren’t catastrophic to the end goal. This may mean we accept that the machine will be incorrect from time to time. It might mean that the solution we return isn’t the best.
That isn’t to say AI is inappropriate in high risk / reward scenarios. AI flies planes and identifies cancer in medical imaging. The trick is finding the right tradeoff. If a process is critical, it demands high accuracy, and that is costly. It isn’t so much that critical tasks should be avoided, but rather that it’s harder to make an ROI argument in these situations due to increased development costs.
Some Common Examples of AI
I wrote another article on five common uses of AI that we interact with every day which you can see here. Here are some additional examples not mentioned in that article:
- Scheduling materials and parts through a production facility
- Digitizing hand filled in forms using machine vision and OCR
- Identifying duplicate entities in databases
- Summarizing Customer feedback from social media
- Predictive maintenance of appliances from telemetry
Each of these has many of the properties we outlined above as desirable. Consider the first, scheduling materials and parts through a production facility. If the shop is doing business, new orders stream in and must be scheduled. For shops producing multiple items, planning production to meet shipping deadlines is time consuming. There are a number of ways of measuring the quality of a schedule, including things like idle time and makespan. Finally, a less than optimal plan isn’t catastrophic. The jobs are still completed and shipped, but perhaps not as quickly as we’d like. As long as performance is comparable to or better than human scheduling, saving the scheduling labor hours to produce more things is likely a good tradeoff.
Summary and Conclusion
AI is an important tool to transfer responsibilities from humans to machines in processes and products. This transfer of responsibilities reduces friction and increase throughput. In turn, this increases the value of what we produce. When incorporating AI into your work, the first step is recognizing the opportunity. In general, an AI opportunity has the following shape:
- A repeated task that takes some amount of human attention
- That amount in aggregate is high
- Outcomes have measurable correctness or quality
- Incorrect / suboptimal outcomes are annoying, not catastrophic
Once you’ve identified an opportunity, the next steps are:
- Validating user demand for that opportunity
- Estimating ROI for developing AI feature
In future articles, we’ll talk about how to de-risk AI opportunities by validating user demand, and how to estimate the time and resources needed to implement an AI enabled feature as part of your products.
