In the business world, every conversation about artificial intelligence seems to fall into one of three categories:
1. AI is a world-changing technology poised to transform everything.
2. AI is the biggest threat to jobs since the industrial revolution.
3. AI as we have it today isn’t really AI.
To be honest, they are all fair points. The problem is that the first camp’s most vocal groups often include many self-interested vendors whose businesses benefit from the AI-adoption frenzy.
Don’t get me wrong. The technology can certainly be transformative for organizations that deploy it to automate routine tasks, drive efficiencies and find insights in data that were previously unanalyzable. Selling based on that premise can be valid. Where vendors tend to go overboard is selling AI as if it is some sort of magic trick that companies can implement with the flip of a switch. This approach is often successful for the vendor but ultimately leaves the customer disappointed.
As the CEO of an AI company myself, I believe we’d all be better off getting rid of the rose-colored glasses and instead showing prospective customers the full truth: Getting to the pot of gold at the end of the AI rainbow isn’t the easy trip that some would want you to believe. It can take time and resources, and the more companies that understand this, the better the industry will be able to truly adopt AI.
While humans need to walk before we can run, machine learning (ML) -- which is generally what we’re referring to when we say AI -- needs to learn how to learn. That learning is powered by algorithms, and those algorithms need to be trained before they’ll do what you need them to do.
To train an algorithm, you need to feed it with data -- usually a substantial amount. The reason Google’s language translation and Facebook’s facial recognition are arguably best in class is because they’ve had access to billions and billions of publicly available consumer data points. With every data point, the machine gets smarter and more accurate.
Enterprises, by comparison, have a much larger challenge with the training process than do data-gobbling behemoths like Google, Facebook and Amazon. But this fact isn’t always made clear to companies during the sales cycle. For example, a business hoping to automate invoice processing may only have a few hundred invoices with which to train the ML, and that may not be enough to get the algorithms to deliver on key automation objectives immediately.
The customer may be completely unaware that this is a problem with big consequences. I’ve counseled a number of companies who went into their AI deployment with aspirations of 50% efficiency savings or reduced headcount. Those goals are achievable, but when you factor in the training process and adoption ramp-up, they may take longer than you realize. Without the right expectations in place, the organizations may run out of patience -- or money -- if they don't budget the increased costs and initial investment period before ROI.
With the right expectations and education about the more SMB-friendly ways to deploy AI, it is possible for businesses to have positive experiences with their AI transformation. Here are a few things to consider during your vendor selection process.
Remember: You Don't Have To Do All The Training
During vendor selection, ask the question: Is the algorithm pretrained, or will you need to do it yourself?
Not all AI solutions depend on their customers’ local training to achieve results. Many come pretrained out of the box, which is the result of the fact that these algorithms have already run on data sets supplied by relevant companies.
Pretrained MLs can enable organizations without Google-sized data repositories to achieve meaningful results with less of their own data contributions. If the task is one that’s fairly common, they may not need to contribute to the training at all.
If you’re amenable to contributing your data to their algorithm’s continued training, you may even be able to negotiate better rates.
Choose Between Going Big And Going Niche
ML today is being applied to a wide variety of use cases. There are vendors whose technology can be used by many of those use cases, and there are vendors whose technology caters specifically to a particular sector.
For businesses that operate in a particular vertical or are looking to automate a particular kind of business process, I believe it’s wise to first examine if one of these sector-specific vendors is an option for you. If you find a solution with a pedigree in your area, it’s likely to be well-trained for your needs. AI for larger or more generalized use cases is typically okay for research purposes or for experiments, where business benefits are not necessarily immediately present. In addition, general AIs can be used for various consumer purposes where specificity to a domain might not be necessary (i.e., a simple command to "figure out if this a picture of a cat" as opposed to "identify the particular phylum that this species of cat belongs to").
Require Proof Of Efficacy
Companies generally measure the accuracy of their ML algorithms by using a classic metric in computer science: the F-score. A “1” represents certainty on the F-score scale. It makes sense -- except that, in my experience, F-scores rarely return a result lower than 0.55, which would mean nothing is uncertain. That doesn’t make sense in the real world.
In my opinion, F-scores simply don’t reflect the full reality, nor are they representative of the way either people or businesses think about certainty. Companies may not want to accept them as the only valid measurement of AI performance. Instead, they can push vendors to use an “efficacy score” that would determine whether the pattern is either right or wrong -- a more binary measure of performance based on real-world outcomes.
Remember: There’s No Magic Bullet
Whether you’re training your own algorithms or using pretrained technology, make sure you budget for the learning curve. Think of ML-based automation like a fine wine -- it gets better over time. It’s great if you get results on day one, but the real impact will increase as time goes on.
